A clinical auxiliary decision-making hierarchical interaction method and system based on a cognitive phase state machine

Abstract

The application discloses a kind of clinical auxiliary decision layering interaction method and system based on cognitive phase state machine, belong to medical informatics field.System by state machine will be forced to be divided into the interaction process and be received in hospital capture state, analysis development state, decision plan state and reflection check state, respectively inject the first to fifth level domain in question template in different states, to constrain large language model respectively execute contradiction point focus, multi-hypothesis horizontal identification and multidisciplinary vertical review, with inflection point early warning dynamic plan construction, and exhaustive self-refutation check verification.The application will the spiral rising cognitive path of high seniority clinical experts be engineered into executable state machine interaction logic, overcome the defects that existing flat question and answer type system is easy to produce anchor bias, miss rare disease and high-risk, significantly improve the depth and reliability of auxiliary decision information.

Description

Technical Field

[0001] This invention relates to medical informatics and decision support systems, and more particularly to a hierarchical, progressive interaction method and system based on state machines that simulates the cognitive transition path of senior clinical experts. Background Technology

[0002] Existing clinical decision support systems typically employ a linear "question-answer" query model. Users input symptoms, and the system provides a list of diagnoses. This interaction is flat and lacks the cognitive layering simulation required for clinical reasoning, including contradiction identification, lateral differentiation, multidisciplinary team (MDT) considerations, and reflective verification. Especially when using large-scale language models, if the cue word hierarchy is singular, anchoring bias is highly likely to occur, leading to the omission of rare diseases or complex comorbidities, resulting in low-quality decision support. Summary of the Invention

[0003] The technical problem this invention aims to solve is to provide an interactive method and system capable of simulating the complete cognitive leap process of senior clinical experts, from "grasping the main contradiction" to "conducting identification and multidisciplinary review," then to "generating contingency plans with escape routes," and finally "self-refutation and verification." The technical solution of this invention is as described in the claims. Attached Figure Description

[0004] Figure 1 This is a system module architecture diagram of the present invention.

[0005] Figure 2 It is the state transition diagram of the cognitive stage state machine.

[0006] Figure 3 This is a schematic diagram showing the information partitioning of the interactive interface at different stages. Detailed Implementation

[0008] Example 1: System Architecture and Core Processes

[0009] The underlying foundation of this system is a large-scale language model (LLM) based on the Transformer architecture. The parameter size can be selected as needed; this embodiment uses a general pre-trained model with 70B parameters. The core contribution of this invention lies in the external state machine control module and the hierarchical prompt template engine. Instead of fine-tuning the base model, it achieves the engineering of higher-order cognitive paths through stateful prompt engineering.

[0010] State machine definition

[0011] Figure 2The state machine transition relationships of this embodiment are illustrated. State transitions are triggered by the system capturing specific operation commands from the user interface or by automatic conditions. Confirmation commands are explicit clicks, and automatic conditions are triggered immediately after the system completes a specific output action.

[0012] Round 1: Reception and Capture Status

[0013] The user enters their chief complaint. The system then sets the current status to "patient capture state".

[0014] The prompt template injection engine retrieves a pre-defined first-level domain question template from the template library. The text structure of this template is as follows:

[0015] "Ignoring all possible diagnoses, you are now an emergency triage specialist. Strictly follow these steps and do not output any content outside of these steps: 1. Extract the most dangerous combination of contradictions from the following chief complaint; 2. Determine if there are any 'three no' conditions (cannot rule out myocardial infarction, cannot rule out aortic dissection, cannot rule out pulmonary embolism); 3. Output only the 'Critical Signal Summary' in the format '[Contradictory Warning]...[Critical Gap]...'. Chief Complaint Content: [User Input]."

[0016] The intent recognition and focus locking module assembles the prompt and user input, sends it to the LLM, retrieves the response, and forcibly blocks subsequent token generation for reaction diagnosis. The final user interface only displays contradictory warnings and missing key signs, such as: "[Contradictory Warning] Acute coronary syndrome may vs. aortic dissection rupture to be ruled out; [Key Gap] No blood pressure in both arms, no description of tearing pain like aortic dissection, no ECG." At this time, the interface does not provide any disease name.

[0017] Round 2: Analyzing the expanded state

[0018] The user adds vital signs and clicks "Confirm In-Depth Analysis". The state machine transitions to the "Analysis Deployment State".

[0019] The wide-area identification and subject decomposition module has started working, injecting commands concurrently in two streams:

[0020] Second level (differential diagnosis transverse chain generation unit):

[0021] "Stop diagnostic convergence. You need to generate at least three independent hypothesis profiles that cannot be merged from all available information. Each profile must include: a pathophysiological category, a list of supporting evidence, and a list of excluding evidence. It is strictly forbidden to merge any two profiles into the same disease framework. Dialogue history: [All dialogues]."

[0022] LLM will output something like: "Contour A: Blood vessel wall rupture (e.g., aortic dissection), Supporting: sudden chest pain, history of hypertension..., Excluding: no mediastinal widening imaging...; Contour B: Blood vessel occlusion (e.g., ACS), Supporting: sweating, 60-year-old male..., Excluding: no ST elevation on ECG...; Contour C: Non-vascular visceral trauma (e.g., esophageal rupture), Supporting: ..., Excluding: ...."

[0023] Tier 3 (Multidisciplinary Vertical Review Unit):

[0024] This unit contains pre-stored templates for internal medicine, surgery, radiology, and laboratory medicine. The unit extracts the names of each independent hypothesis profile from the LLM output and triggers the following multi-role prompts:

[0025] "Now you must play the roles of a radiology expert and a cardiology expert respectively. Regarding the aforementioned profiles A and B, the radiology expert asks: In cases where a CTA cannot be performed, what indirect signs must be looked for on a bedside chest X-ray? The cardiology expert asks: If the onset of illness is only 2 hours old and high-sensitivity troponin is negative, how should this be interpreted dynamically? Each of you should provide a one-sentence question."

[0026] The interactive interface is divided into sections: the left side displays a "Multi-Hypothesis Logic Tree," and the right side displays "MDT Multi-Window Questioning." This mandatory multidisciplinary perspective questioning information is presented simultaneously with the independent hypotheses, constituting a key distinguishing feature of this invention.

[0027] Third round: Decision-making contingency plan

[0028] After the user reviews the partition information, they click "Generate Plan". The state machine then switches to the "Decision Plan State".

[0029] The dynamic contingency plan construction module loads the fourth-level template:

[0030] "Based on the current information, construct a dynamic emergency response plan with escape doors. The plan should include:"

[0031] 1. Main path: Named 'Myocardial Infarction Path', the execution steps are...

[0032] 2. Alternative path: Named 'Interstitial Fuse Failure Path', with the trigger condition being 'Intima-media flap or pericardial effusion seen on bedside ultrasound'.

[0033] 3. Inflection Point Signal Monitoring: If blood pressure drops below 90 / 60 mmHg within 15 minutes of the main pathway being initiated, immediately switch to the 'hypotension branch' and mandate bedside echocardiography. Please output in timeline format.

[0034] The inflection point signal injection interface in the module will output a node graph labeled with time windows. Each node is marked with an observation target (such as "T+5 minutes: blood pressure & heart rate") and a preset threshold. When subsequent monitoring data is received, path switching can be automatically triggered (in this embodiment, this is presented as a prominent reminder on the interface).

[0035] Fourth round: Reflection and verification status

[0036] After the contingency plan is output, the state machine does not wait for user instructions and immediately and unconditionally activates the "reflection and verification state".

[0037] The self-questioning and closed-loop verification module injects the fifth-level template:

[0038] "Perform a self-reversal without any restrictions. Forced to complete the following tasks:"

[0039] 1. Exhaustive list of rare diseases: Among all current hypotheses, is spontaneous esophageal rupture (Boerhaave syndrome), a rare but fatal option, overlooked? What are the latent characteristics of this disease?

[0040] 2. Cognitive bias identification: Did the patient's chief complaint of 'chest pain' lead to the failure to consider pulmonary embolism, with lower extremity swelling as the only early symptom? Please list the evidence that has ruled out pulmonary embolism. If there is no evidence, issue a warning.

[0041] 3. Absolute Contraindication: If a heavy dose of antiplatelet and heparin is administered according to the main pathway, and an undetected aortic dissection is present, catastrophic consequences will occur. This warning must be placed at the highest level of alert.

[0042] The counterfactual inference engine scans for uncovered rare ICD-10 codes in the aforementioned dialogue. If a gap is found, it forces the generation of a corrective question, such as: "Has the patient been explicitly asked whether they have a history of severe vomiting recently?"

[0043] All output is condensed into a black warning bar, attached below the contingency plan. The system then generates a closed-loop confirmation control stating "All risks are known; confirm execution / modify contingency plan." Prior to this, the text interaction box was restricted. The user must make an explicit risk awareness selection to end this round of interaction.

[0044] Through this phased state machine control, the present invention transforms AI-assisted decision-making, which is prone to empirical bias, into a "cognitive exoskeleton" with reverse verification and multidisciplinary checks and balances, thereby achieving the technical effect of enhancing decision robustness.

[0045] The above description is only a preferred embodiment of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A hierarchical interactive system for clinical decision support based on a cognitive stage state machine, characterized in that, include: A state machine control module is configured to mark the cognitive stage state as reception capture state, analysis unfolding state, decision plan state or reflection and verification state according to the interaction process of the current session, and execute state transitions according to predetermined logic; A prompt template injection engine is configured to retrieve the corresponding level of in-domain question template from the template library according to the cognitive stage currently marked by the state machine control module, and inject it into the basic large language model, so as to constrain the basic large language model to generate prompt information limited to the stage sub-goals for the current dialogue record; An intent recognition and focus locking module is configured to load a first-level domain question template in the reception capture state. The first-level domain question template is constructed to force the basic large language model to ignore diagnostic suggestions and only parse the high-risk contradictions in the input chief complaint, generate a summary of key contradictions and force it to be output to the interactive interface, while masking diagnostic output. A wide-area identification and subject expansion module is configured to load a second-level domain question template and a third-level domain question template in the analysis expansion state. The second-level domain question template is constructed to forcibly generate multiple independent hypothesis profiles that cannot be merged with each other, as well as corresponding positive evidence and negative exclusion evidence. The third-level domain question template is constructed to forcibly inject viewpoint templates from at least two different medical disciplines, generate cross-questioning or supporting evidence for the multiple independent hypothesis profiles, and display them in sections in the interactive interface. A dynamic contingency plan construction module is configured to load a fourth-level domain question template in the decision contingency plan state. The template is constructed to build a multi-branch dynamic emergency plan based on the preceding dialogue log, which includes a main path, at least one alternative path, and inflection point signal monitoring conditions for triggering the switching between the main and alternative paths. A self-questioning and closed-loop verification module is configured to load a question template within the fifth-level domain in the reflection and verification state. The template is constructed to perform at least one of the following operations: exhaustive review of rare diseases in the current plan, identification and questioning of cognitive bias, and bottom-line exclusion warning for absolutely prohibited operations, and generate a closed-loop confirmation instruction to lock the interactive output.

2. The system according to claim 1, characterized in that, The predetermined logic of the state machine control module includes: After obtaining the user's complaint for the first time, the status is set to the reception capture state; Upon receiving confirmation from the user regarding the summary of the key contradictions, the system switches to the analysis expansion state. Upon receiving a user's instruction to terminate the review of the content displayed in the partition, switch to the decision contingency plan state; After the multi-branch dynamic emergency plan is output to the interactive interface, the reflection and verification state is automatically activated.

3. The system according to claim 1, characterized in that, The wide-area identification and subject expansion module includes: The differential diagnosis lateral chain generation unit is used to extract symptom entities and their attributes from the dialogue record, calculate the symptom similarity and timeline coupling degree, and generate the multiple independent hypothesis profiles based on this, wherein each independent hypothesis profile does not point to a single specific disease. A multidisciplinary longitudinal review unit is used to forcibly inject at least two of the internal medicine viewpoint template, surgical viewpoint template, radiology viewpoint template, and laboratory viewpoint template, and drive the underlying large language model to output exclusionary questions or supporting evidence for each independent hypothesis profile in the corresponding subject viewpoint.

4. The system according to claim 1, characterized in that, The dynamic contingency plan construction module includes an inflection point signal injection interface, which is used to divide the main path into multiple continuous decision time windows, set at least one observable clinical target in each decision time window, and configure a condition threshold for each clinical target; when the observable clinical target triggers the condition threshold in the corresponding time window, the dynamic contingency plan construction module switches to the alternative path.

5. The system according to claim 1, characterized in that, The self-questioning and closed-loop verification module includes a counterfactual inferencer configured to identify, in the response generated by the underlying large language model, the first rare disease entity with atypical representations that is not covered by the independent hypothesis profile, and to force the underlying large language model to generate a corrective counter-question that points to the key representations of the rare disease entity.

6. A hierarchical interactive method for clinical auxiliary decision-making based on a cognitive stage state machine, using the system described in any one of claims 1 to 5, characterized in that, Includes the following steps: Step A: Receive the initial input complaint, the state machine enters the reception capture state, loads the first-level domain question template, forcibly generates and outputs a summary of key contradictions, and simultaneously suppresses the output of diagnostic conclusions; Step B: After receiving confirmation of the summary of the key contradictions, the state machine transitions to the analysis expansion state, loads the second-level and third-level domain query templates in sequence, independently performs the multi-hypothesis identification analysis operation and the multidisciplinary cross-review operation, and displays the results in partitions. Step C: After receiving the instruction to terminate the review of the content displayed in the partition, the state machine transitions to the decision-making plan state, loads the fourth-level domain question template, and generates a multi-branch dynamic emergency plan containing at least one alternative path and clear inflection point monitoring indicators. Step D: After the multi-branch dynamic emergency plan is generated, the state machine automatically switches to the reflection and verification state, loads the question template in the fifth-level domain, performs exhaustive cross-examination review operation and generates a closed-loop confirmation instruction, and keeps the interaction path open before generating the confirmation statement used to close the interaction.

7. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the method of claim 6.

8. 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 of claim 6.

Images