A medical accompanying diagnosis decision system based on multi-agent cooperation and CoT mechanism combination

Abstract

The application discloses a medical accompanying diagnosis decision system based on multi-agent cooperation and CoT mechanism combination, relates to the technical field of intelligent medical treatment, and comprises a multimodal unified semantic module, a multi-agent cooperative scheduling module, a chain thinking reasoning module and an end-cloud cooperative optimization module, wherein the multimodal unified semantic module outputs unified semantic representation; the multi-agent cooperative scheduling module distributes a medical treatment task to a target role agent according to a complexity threshold based on a complexity score of a calculation task; the chain thinking reasoning module generates a multi-path reasoning chain, outputs accompanying diagnosis decision data through multiple verifications, and simultaneously generates a reasoning cost signal; and the end-cloud cooperative optimization module performs end-side deployment and incremental updating on each module model, the multi-agent cooperative scheduling module normalizes the reasoning cost signal into a dimensionless cost relative to a historical distribution, and the dimensionless cost is used as an online strategy optimization reward reverse calibration weighting coefficient and a complexity threshold to form a closed loop, so that reasoning difficulty becomes a distribution prior, and distribution and reasoning mismatch is eliminated.

Description

Technical Field

[0001] This invention relates to the fields of smart healthcare and artificial intelligence technology, specifically to a medical escort decision-making system based on the combination of multi-agent collaboration and CoT mechanism. Background Technology

[0002] With the increasing aging of the population and the advancement of digital transformation in healthcare processes, there is a significant demand for in-hospital assistance and companionship decision support for the elderly, people with mobility impairments, and patients with chronic diseases. Upon arrival at the hospital, patients often need to make complex decisions regarding registration, triage, department selection, examination scheduling, waiting lines, and patient flow arrangement. This process involves multi-source heterogeneous information, including voice complaints, medical images, electronic medical records, and physiological timing signals from wearable devices. Combining the reasoning capabilities of large language models, the role division of multiple agents, and the fusion of multimodal information to construct a companionship decision-making system capable of outputting navigation parameters such as department routing, waiting time prediction, and patient flow arrangement has become an important technological direction in the field of smart healthcare.

[0003] In the field of medical decision-making combining multi-agent collaboration and enhanced reasoning, existing technologies have explored various approaches. Chinese patent application CN120544940A discloses a collaborative reasoning method and system that integrates medical knowledge graphs and large models. This system dynamically updates the medical knowledge graph in the cloud using a medical information monitoring agent and creates a clinical decision-making collaborative reasoning model comprising a multimodal fusion layer, a collaborative reasoning layer, and an interpretation generation layer. The multimodal fusion layer achieves modal alignment based on medical ontology to obtain unified semantics. The collaborative reasoning layer uses cross-attention and temporal convolution to perform interactive reasoning on the large model and knowledge graph, outputting candidate reasoning hypotheses with confidence assessments. The interpretation generation layer generates traceable causal chain reasoning reports. Finally, a lightweight model is pushed to the medical terminal via a federated gateway and iteratively optimized using federated learning and differential privacy. However, the system adopts a single-model architecture that is connected layer by layer. It does not set up a scheduling mechanism to allocate tasks to intelligent agents with different roles based on the complexity of the medical treatment task, nor does it bring the reasoning difficulty exposed in the reasoning stage back to the task allocation stage. Therefore, once the complexity assessment of the task is determined, it cannot be corrected according to the actual reasoning process. The mismatch between the task complexity assessment and the actual reasoning difficulty cannot be detected by the system itself.

[0004] US Patent Application Publication No. US20170124269A1 discloses a system for identifying new knowledge in clinical decision support. This system uses a supervising agent to manage multiple collaborating agents, decomposing complex problems into manageable sub-tasks. Based on available patient information, it selects the most effective agent from among the multiple agents and feeds back the accuracy of diagnostic results from each agent to the supervising agent, thereby learning which agent to use under what patient information. However, the quantity used for feedback is the diagnostic accuracy obtained after the fact, which is a sparse signal dependent on result labeling. It can only be obtained after the task is completed and a correct conclusion is obtained, and cannot reflect the difficulty of the reasoning process itself. Therefore, it also cannot correct the mismatch between complexity assessment and reasoning difficulty in the task allocation stage before reasoning.

[0005] In summary, existing multi-agent medical decision-making systems assess the complexity of medical tasks using static feedforward quantities calculated once before reasoning begins. This results in a systematic mismatch between the static feedforward quantity and the actual reasoning difficulty, which only gradually emerges during the reasoning process. Tasks that appear to have many modalities and high terminology density are not necessarily difficult to reason with, while truly difficult tasks often involve repeated stumbling and error correction during reasoning—difficulties that cannot be pre-characterized by static feedforward quantities. More critically, this mismatch lacks a feedback path from the reasoning process back to the allocation decision, making it impossible for the feedforward quantity itself to detect and correct. Ultimately, this leads to tasks with easier reasoning being assigned to agents with heavier reasoning burdens, while difficult tasks are assigned to weaker agents, hindering the accuracy and real-time performance of medical accompaniment decisions. How to make the reasoning difficulty, which is only visible after reasoning, a priori for task allocation, thereby eliminating this mismatch, is a pressing technical problem to be solved in this field. Summary of the Invention

[0006] To address the core technical bottleneck in existing medical accompaniment decision-making systems—namely, the mismatch between static feedforward task complexity assessment and actual reasoning difficulty, which cannot be detected by the system itself—this invention provides a medical accompaniment decision-making system based on multi-agent collaboration and a CoT mechanism. This system normalizes the reasoning cost signal representing retry and error correction intensity generated during the chain-thinking reasoning stage relative to the historical reasoning cost distribution of the role agents, and uses this as a reward for online policy optimization. This process reverse-calibrates the weighting coefficient and complexity threshold of the task complexity score. Without relying on post-diagnosis result labeling, it achieves self-detection and elimination of the mismatch between allocation and reasoning based on the closed-loop feedback control principle from reasoning difficulty to allocation decision.

[0007] The technical solution of this invention is: a medical escort decision-making system based on multi-agent collaboration and the CoT mechanism, comprising a multimodal unified semantic module, a multi-agent collaborative scheduling module, a chain-based reasoning module, and an edge-cloud collaborative optimization module. The multimodal unified semantic module is used to perform cross-modal alignment and fusion of the patient's multimodal medical input, outputting a unified semantic representation. The multi-agent collaborative scheduling module is configured with multiple agents with different roles, used to calculate the task complexity score of the medical task based on the unified semantic representation, and to allocate the medical task to the target agent based on the task complexity score and a complexity threshold. The chain-based reasoning module is used for the target agent to generate a multi-path reasoning chain based on a reasoning template, and to perform multiple verifications on the multi-path reasoning chain, outputting escort decision data, and generating representations of retry and error correction intensity during the reasoning and verification process. The inference cost signal; the edge-cloud collaborative optimization module is used to perform edge-side lightweight deployment and incremental updates of the models of the multimodal unified semantic module, the multi-agent collaborative scheduling module and the chain-thinking inference module; the multi-agent collaborative scheduling module is also used to normalize the inference cost signal relative to the historical inference cost distribution of the target role agent into a dimensionless cost, and use the dimensionless cost as the reward for online policy optimization, and reverse-calibrate the weighting coefficient of the task complexity score and the complexity threshold, so that the inference cost signal output by the chain-thinking inference module and the allocation decision of the multi-agent collaborative scheduling module form a closed loop.

[0008] Furthermore, the inference cost signal is a weighted combination of the number of times the target agent triggers re-inference in the multiple verifications, the pruning depth of the multi-path inference chain, and the number of error corrections between the multi-path inference chain and the medical knowledge graph. The multiple verifications include logical consistency verification, knowledge graph matching verification, and clinical guideline comparison verification. The normalization is to map the inference cost signal to its quantile in the historical inference cost distribution of the target agent, and to construct a profile of the agent's reasoning ability based on the historical inference cost distribution statistics of each agent. The online policy optimization constructs a dense reward with the dimensionless cost and weights it with the sparse reward. The update of the weighting coefficient and the complexity threshold is superimposed with a stability regularization term based on Fisher information, and the Fisher information and the incremental update of the edge-cloud collaborative optimization module share the same estimate.

[0009] The beneficial effects of this invention are as follows: Firstly, this invention uses the inference cost signal to calibrate the task complexity score and form a closed loop, which can eliminate the mismatch between the static feedforward complexity assessment and the actual inference difficulty. The mechanism is that the inference cost signal is the most sensitive ex-post measure of complexity assessment error during the inference stage. Bringing it back to the allocation decision constructs a negative feedback path on the original single feedforward allocation process, so that mismatches that the feedforward quantity itself cannot detect can be continuously exposed and corrected in real time by the inference process.

[0010] Secondly, this invention normalizes the inference cost signal relative to the historical inference cost distribution of the role agent into a dimensionless cost before feeding it back. The mechanism is that the inference costs of different role agents and different inference templates are not consistent in magnitude. If they are directly aggregated, the reward will be dominated by a single role agent with a large inference cost, resulting in calibration distortion. After quantile normalization, the inference costs of each role agent are mapped to the same dimensionless interval, making them comparable and additive. Furthermore, the inference ability profile of the role agent produced by the normalization process further feeds back into the professional dimension of the task complexity score, forming a mutual promotion between means and effect.

[0011] Third, this invention superimposes a stability regularization term based on Fisher information onto the updates of weighting coefficients and complexity thresholds. The mechanism is that closed-loop feedback continuously pulls the complexity threshold; without constraints, the allocation strategy is prone to oscillations and forgetting previously effective allocation experiences. Fisher information imposes constraints on key parameters of the previous allocation strategy, ensuring that the update converges stably while maintaining calibration capabilities. Furthermore, this Fisher information shares the same estimate with the incremental updates of the edge-cloud collaborative optimization module, ensuring that the stability of the allocation strategy and the anti-forgetting capabilities of the edge model are uniformly supported by the same mechanism, achieving cross-module collaboration.

[0012] Fourth, the aforementioned features are coupled to form a positive feedback loop, generating a system-level synergistic effect that exceeds the sum of the individual effects of each feature. The mechanism is that the improvement in task allocation accuracy reduces the mean of the inference cost signal, thereby making the estimation of the dimensionless cost more accurate, which in turn makes the allocation more accurate. This cycle causes the complexity threshold to converge to a stable fixed point. At this fixed point, the system's response delay and hallucination rate for accompanying patients are both lower than the corresponding values ​​when the closed loop is broken, thus achieving improvements in both real-time performance and reliability. This improvement cannot be achieved by any single feature, reflecting the nonlinear gain of multi-module collaboration. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a schematic diagram of the structure of the multimodal unified semantic module of the present invention; Figure 3 This is a schematic diagram of the chain-like reasoning module of the present invention; Figure 4 This is a schematic diagram of the data flow of the multi-agent collaborative scheduling module and closed-loop calibration of the present invention; Figure 5 This is a schematic diagram of the structure of the edge-cloud collaborative optimization module of the present invention. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to illustrate the technical solutions of the present invention and do not limit the scope of protection of the present invention.

[0015] like Figure 1 As shown, the medical escort decision-making system based on multi-agent collaboration and CoT mechanism provided in this embodiment is deployed in the hospital information environment. It includes a multimodal unified semantic module 1, a multi-agent collaborative scheduling module 2, a chain-like reasoning module 3, and an edge-cloud collaborative optimization module 4 from bottom to top. The four form a deeply coupled closed-loop collaborative system.

[0016] The system receives multimodal patient input from patients, including voice complaints, medical images, electronic medical record text, and physiological time-series signals such as heart rate, blood pressure, and blood oxygen saturation collected via mobile applications, wearable devices, or in-hospital smart terminals. The multimodal unified semantic module 1 performs cross-modal alignment and fusion of the above multimodal patient input, outputting a unified semantic representation, which is then sent to the multi-agent collaborative scheduling module 2. The multi-agent collaborative scheduling module 2 is configured with multiple intelligent agents with different roles. These agents calculate the task complexity score of the patient task based on the unified semantic representation, and then assign the patient task to the target intelligent agent based on the task complexity score and complexity threshold. The target intelligent agent generates a multi-path reasoning chain based on a reasoning template in the chained reasoning module 3, performs multiple verifications on the multi-path reasoning chain, and outputs accompanying decision data including department routing, waiting time prediction, and patient flow arrangement. Simultaneously, the chained reasoning module 3 generates a reasoning cost signal representing the intensity of retry and error correction during the reasoning and verification process. The edge-cloud collaborative optimization module 4 is responsible for deploying and incrementally updating the models of the above modules on the edge side in a lightweight manner, so as to maintain the real-time performance and privacy protection of the system on edge devices with limited computing power.

[0017] In terms of deployment, the reasoning components of the multimodal unified semantic module 1, the multi-agent collaborative scheduling module 2, and the chain-like reasoning module 3 are deployed on in-hospital edge devices to ensure that sensitive patient information is processed locally and meets real-time requirements; the maintenance and updating of the medical knowledge graph and clinical guideline base, as well as federated aggregation, are deployed in the cloud. The edge-cloud collaborative optimization module 4 acts as a bridge between the edge and the cloud. On the one hand, it quantifies and distributes updated knowledge and models from the cloud to the edge; on the other hand, it uploads locally updated information from the edge to the cloud for aggregation after privacy protection, thereby maintaining the system's reasoning ability and knowledge timeliness on the edge with limited computing power. The four modules are linked by a unified semantic representation, task complexity scoring, accompanying medical decision data, and inference cost signals: the unified semantic representation flows from the multimodal unified semantic module 1 to the multi-agent collaborative scheduling module 2; the task complexity scoring drives allocation decisions within the multi-agent collaborative scheduling module 2; the accompanying medical decision data is output from the chain-like reasoning module 3; and the inference cost signals flow back from the chain-like reasoning module 3 to the multi-agent collaborative scheduling module 2, forming a feedback loop from the inference stage back to the allocation stage. This data link ensures both the integrity of the forward processing chain and the closure of the feedback loop, eliminating any suspended intermediate data.

[0018] Unlike conventional single-feedback architectures, in this embodiment, the inference cost signal generated by the chain-thinking inference module 3 is not only used for the final decision on whether to include or exclude data in the accompanying medical examination, but is also fed back to the multi-agent collaborative scheduling module 2. The multi-agent collaborative scheduling module 2 normalizes the inference cost signal relative to the historical inference cost distribution of the target agent into a dimensionless cost, and uses this dimensionless cost as a reward for online policy optimization, inversely calibrating the weighting coefficient and complexity threshold of the task complexity score. Thus, a closed loop is formed between the inference cost signal output by the chain-thinking inference module 3 and the allocation decision of the multi-agent collaborative scheduling module 2, making the inference difficulty, which is only visible after inference, a priori for the next task allocation. This closed loop is the core of the deep coupling and synergistic gain among the modules of this system, and its specific implementation will be detailed in Embodiment Six.

[0019] like Figure 2 As shown, the multimodal unified semantic module 1 includes a visual-text alignment unit, a cross-modal attention fusion unit, and a temporal physiological signal encoding unit, which are used to map the multimodal medical input of patients into a unified semantic space and fuse it into a unified semantic representation.

[0020] The visual-text alignment unit learns to align medical images with their corresponding diagnostic report text through comparison. It extracts image and text features using an image encoder and a text encoder respectively, then maps these features to the same dimension via a projection layer. The unit is trained using a joint loss, expressed as: , in: is the joint training loss of the visual-text alignment unit, which is a scalar, dimensionless quantity, calculated by this formula, and is used to drive the semantic alignment of image features and text features. For image-text contrast loss, is a scalar, dimensionless, and adopts the information noise contrast estimation form to bring matching image-text pairs closer and push away mismatched pairs. The image-text matching loss is scalar and dimensionless, and binary cross-entropy is used to determine whether the image-text pair comes from the same patient sample. The loss for language modeling is scalar and dimensionless, and the text generation cross-entropy is used with image as a condition. and Let be the loss weight, and be a scalar with a range of values. Take respectively , The value is determined by optimization on the validation set and is used to balance the contributions of the matching task and the generation task to the alignment training. The image-text contrast loss is further expressed as: , in: The number of samples in a training batch is a positive integer. , used to define the range of negative samples in contrastive learning; , Here is the sample sequence number, which is a positive integer and its value range is [value range missing]. to ; For the first The image embedding vector of each sample is... The dimensional vector is extracted by the image encoder and mapped through a projection layer; , For text embedding vectors, The dimensional vector is extracted by the text encoder and mapped through a projection layer; The cosine similarity function has a range of values. This is used to measure the semantic proximity of image-text embeddings; Let be the temperature coefficient, and be a scalar. The unitless parameter is used to control the sharpness of the negative sample distribution. Through the above joint training, the image-text alignment error rate of the visual-text alignment unit was reduced to 1.8%.

[0021] The cross-modal attention fusion unit integrates patient speech descriptions, medical images, and electronic medical record text into a unified semantic representation. Let the text feature sequence be... Image region features are Speech features are The fusion process is represented as: , in: The query matrix is ​​composed of text feature sequences. Obtained through linear transformation; , Given the key matrix and value matrix, derived from image region features With speech features The result is obtained by linear transformation after splicing; superscript Indicates matrix transpose; Let be the dimension of the key vector, and be a positive integer. , This is a scaling factor used to prevent the gradient from vanishing due to an excessively large dot product; This is a normalized exponential function, whose output is the weights of each value vector, and its range is... The fusion employs multi-head attention, with a number of heads of [number missing]. Through cross-modal attention fusion, the accuracy of multi-source information fusion was improved to 91.7%.

[0022] The temporal physiological signal encoding unit performs time-frequency joint encoding on multi-channel physiological signals such as heart rate, blood pressure, and blood oxygen saturation sampled at 100Hz. The time-domain encoding is obtained by modeling the sliding window signal using a time-series encoding network, while the frequency-domain encoding is obtained by the amplitude spectrum of the signal from DC to 50Hz using the Fast Fourier Transform. The two are concatenated and projected to obtain the encoding features of the temporal physiological signal. The sliding window length is 2s, the step size is 0.5s, and the inter-frame overlap rate is 50%. Through this time-frequency joint encoding, the end-to-end processing latency of the temporal physiological signal is controlled within 15ms. The outputs of the visual-text alignment unit, the cross-modal attention fusion unit, and the temporal physiological signal encoding unit are concatenated after unified projection to obtain a unified semantic representation with a dimension of 512, which is read by the multi-agent collaborative scheduling module 1 to calculate the task complexity score.

[0023] like Figure 4 As shown, the multi-agent collaborative scheduling module 2 is configured with multiple intelligent agents with different roles, including a patient assessment agent, a medical knowledge agent, a process scheduling agent, and a multimodal perception agent. The patient assessment agent is responsible for analyzing the patient's chief complaint, physiological indicators, and medical history, and outputting a patient status assessment; the medical knowledge agent is responsible for retrieving medical knowledge graphs, clinical guidelines, and medication databases, and outputting structured medical knowledge; the process scheduling agent is responsible for arranging hospital processes such as registration guidance, examination scheduling, and waiting time prediction; and the multimodal perception agent is responsible for processing medical images, voice input, and wearable sensor data.

[0024] The four types of intelligent agents mentioned above are independent yet collaborative in their responsibilities: the patient assessment agent focuses on characterizing the individual state of the patient; the medical knowledge agent focuses on the retrieval and structuring of external medical knowledge; the process scheduling agent focuses on the orchestration of hospital resources and processes; and the multimodal perception agent focuses on the perception and preprocessing of raw multimodal signals. For the same medical treatment task, the multimodal perception agent in the multi-agent collaborative scheduling module 2 first completes the perception, and then selects the target agent to undertake the main reasoning based on the task complexity score. The other agents then provide auxiliary information such as state assessment, knowledge retrieval, or process constraints to the target agent in a collaborative manner. This division of labor-based collaboration avoids the capability dilution caused by a single model handling all sub-tasks, and also provides a granular observation unit for subsequently characterizing the reasoning struggles of each agent using reasoning cost signals, and thus constructing a profile of the reasoning capabilities of each agent.

[0025] The multi-agent collaborative scheduling module 2 calculates the task complexity score of the medical visit task based on the unified semantic representation output by the multimodal unified semantic module 1. The task complexity score is represented as: , in: For medical treatment tasks The task complexity score is a scalar with a value range of 1. The value is calculated by this formula. The larger the value, the more complex the task and the more the intelligent agent needs to perform reasoning. For pending medical appointments; The modal dimension score is denoted by , which is a scalar and has a value range of . The value is 0.33 for single-modality, 0.67 for dual-modality, and 1.0 for tri-modality and above. It is determined by the number of modalities involved in the unified semantic representation and is used to reflect the types of information sources involved in the task. The score represents the urgency dimension, and is a scalar with a value range of [value range missing]. It is determined by the degree of abnormality in the patient's vital signs and is used to reflect the time sensitivity of the task. The score represents the professionalism dimension, and is a scalar with a value range of [value range missing]. The density of medical terms used in patient complaints is determined by word frequency-inverse document frequency statistics, which is used to reflect the domain specialization of the task. , , The weighting coefficient for the task complexity score is denoted by , which is a scalar and takes values ​​ranging from . And satisfy the constraints Initially take , , It is dynamically updated by online strategy optimization to balance the contributions of the three dimensions to the task complexity score.

[0026] The multi-agent collaborative scheduling module 2 assigns medical treatment tasks to target role agents based on task complexity scores and complexity thresholds. The assignment rules are expressed as follows: , in: The target role is the intelligent agent, and the optimal allocation object is selected by this formula; Candidate role intelligent agents; The adaptation score of candidate role agents under a given task complexity score; Let be a high-complexity threshold, and be a scalar with a value range of . The initial value is 0.70; Let be the low-complexity threshold, and be a scalar with a value range of . The initial value is 0.35; and These are collectively referred to as the complexity threshold, which is dynamically updated by online policy optimization. The above formula represents piecewise selection, where both the input and the threshold are dimensionless quantities.

[0027] It should be noted that the weighting coefficients in the task complexity score... , , Complexity threshold in the allocation rules , It is not a fixed constant, but rather a reverse calibration performed by online policy optimization with a dimensionless cost as a reward. The calibration process constitutes the core closed loop of this system, as detailed in Implementation Example 6.

[0028] The various intelligent agents communicate with each other via a medical domain-extended communication protocol. This protocol extends the general remote procedure call protocol, and its message body carries a digest-compressed inference chain, inference confidence, and a list of clinical guideline citations in standard encoding. The digest-compressed inference chain is no longer than 512 tokens, and the inference confidence value ranges from [value missing]. To reduce end-to-end communication overhead, the message body is losslessly compressed before transmission and combined with connection pooling multiplexing. The maximum number of connections in the connection pool is 32, thereby keeping the communication latency between the various roles of the intelligent agents within 50ms.

[0029] like Figure 3 As shown, the chain-like reasoning module 3 generates a multi-path reasoning chain based on the reasoning template by the assigned target role intelligent agent, performs multiple verifications on the multi-path reasoning chain, outputs the accompanying medical decision data, and generates a reasoning cost signal during the reasoning and verification process.

[0030] The chain-based reasoning module 3 has a built-in medical reasoning template library. It automatically extracts reasoning patterns from authoritative clinical guidelines, constructing three types of templates: diagnostic reasoning chain templates, medication reasoning chain templates, and patient visit process reasoning chain templates. Each template consists of prerequisites, several reasoning steps, and a conclusion, with a template coverage rate of 85%. The target agent initializes a multi-path reasoning chain based on the reasoning template. The multi-path reasoning chain consists of multiple candidate reasoning paths, each of which unfolds progressively according to depth. The node value is calculated at each reasoning node, and the node value is represented as: , in: For the first Each reasoning node The node value is a scalar, and its value ranges from 1 to 2. It is used to guide the search and pruning of multi-path reasoning chains; For the first inference chain in a multi-path chain One reasoning node; For reasoning nodes With medical knowledge graph The semantic similarity, with values ​​ranging from 1 to 2. Cosine similarity is used as the metric. The medical assessment model scores the reasonableness of inference nodes, with a value range of [value range missing]. ; The similarity weight for the knowledge graph is a scalar value of 0.6, without units, determined by grid search. A larger value emphasizes knowledge graph verification to suppress model illusions. The system evaluates multiple candidate inference paths based on their average node value and uses bundle search to maintain the active path set and prune low-value paths. The bundle width is 3, and the maximum unfolding depth of the multi-path inference chain is 8.

[0031] The chain-based reasoning module 3 performs multiple verifications on the multi-path reasoning chain, including logical consistency verification, knowledge graph matching verification, and clinical guideline comparison verification. Logical consistency verification calculates the conditional probability between adjacent reasoning steps; if this conditional probability is lower than the logical consistency threshold... The system detects logical breaks and triggers re-reasoning; knowledge graph matching verification calculates the overlap between the set of neighbors of medical entities in the medical knowledge graph and their associated sets in the reasoning context. When the overlap is below 0.5, it is considered a knowledge mismatch and knowledge correction is triggered; clinical guideline comparison verification calculates the Jaccard similarity between the set of keywords in the reasoning conclusion and the set of recommended words in the retrieved clinical guidelines. When the similarity is below 0.45, a warning is issued; when it is below 0.2, it is transferred to manual review. Through the above three verifications, the reasoning consistency error was reduced from 12.3% to 4.1%.

[0032] Unlike conventional systems that only use the triggering results of the above verification for error correction and conclusion selection, the chain-thinking reasoning module 3 in this embodiment further generates a reasoning cost signal representing the intensity of retrying and error correction during the reasoning and verification process. The reasoning cost signal is a weighted combination of the number of times the target agent triggers re-reasoning in multiple verifications, the pruning depth of the multi-path reasoning chain, and the number of error corrections made by the multi-path reasoning chain and the medical knowledge graph, expressed as: , in: Let $\frac{ ... The value is dimensionless and is obtained by weighted summation of the three types of counts in this formula. The larger the value, the more frequently the target intelligent agent retries and corrects errors in this reasoning and verification process, that is, the higher the degree of reasoning struggle. In logical consistency checks, due to conditional probability being lower than... The number of times re-inference is triggered is a non-negative integer with no unit, and is obtained from the statistics of logical consistency verification. The cumulative depth of the inference nodes pruned in the multi-path inference chain during the bundle search is a non-negative integer with no unit, and its value does not exceed the maximum unfolding depth of 8. It is obtained by statistics from the bundle search pruning process. This represents the number of times knowledge correction is triggered during knowledge graph matching verification due to an overlap of less than 0.5. It is a non-negative integer with no unit and is obtained from knowledge graph matching verification statistics. , , Here, is the weighting coefficient of the inference cost signal, and is a scalar with a value range of . And satisfy Initially take , , This is used to balance the contributions of the three sources of struggle.

[0033] The three types of counts described above characterize three different sources of reasoning struggle: the number of times re-reasoning is triggered reflects the repetition of the reasoning chain at the logical level, the pruning depth reflects the detour of multi-path search in the candidate space, and the number of error corrections reflects the deviation between the reasoning chain and the medical knowledge graph. A weighted combination of these three factors provides a unified scalar characterizing the overall struggle intensity of the target agent in this reasoning process. It is worth noting that the three types of counts used here are originally byproducts of multiple verification, which are discarded after triggering error correction or path pruning in conventional systems; however, this embodiment retains and quantifies them as useful signals of the true reasoning difficulty of the coding task, thus providing an observation that is immediately available during the reasoning process without waiting for the medical results for subsequent closed-loop calibration. The reasoning cost signal is output by the chain-thinking reasoning module 3 and sent back to the multi-agent collaborative scheduling module 2 as input for closed-loop calibration.

[0034] like Figure 5 As shown, the edge-cloud collaborative optimization module 4 includes a model quantization unit, an incremental learning unit, and a federated privacy unit, which are used to deploy the models of the above modules on edge devices with limited computing power and perform incremental updates, while ensuring the privacy and security of sensitive information of patients.

[0035] The model quantization unit performs low-bit quantization on the linear layer weights of the model, employing a layer-by-layer Hessian-aware quantization strategy. The weight matrix is ​​quantized in blocks and corrected using Hessian matrix-aware error compensation, with a quantization bit width of 4 bits. This quantization improves the edge inference speed of the 7B-parameter-scale medical model from the baseline of 7.2 tokens / s to 23.0 tokens / s, reduces edge memory usage from 12GB to 3.8GB, and keeps the accuracy loss on the medical benchmark within 2.1%.

[0036] The incremental learning unit performs incremental fine-tuning on the pre-trained model based on a low-rank adaptation matrix. For knowledge injection into new departments or diseases, only one pair of low-rank adaptation matrices is trained while freezing the pre-training weights. The number of trainable parameters is approximately 0.1% to 0.8% of the total fine-tuning parameters, with a low-rank dimension of 16. To prevent catastrophic forgetting of existing knowledge during incremental fine-tuning, the incremental learning unit adds a regularization term based on Fisher information to the incremental fine-tuning loss. Fisher information imposes constraints on important parameters from the old task. With this incremental learning strategy, the model's forgetting rate in small-sample scenarios decreased from 18.7% to 5.3%, while the accuracy on new tasks remained at 89.4%. It should be noted that the Fisher information used here to suppress catastrophic forgetting is the same estimate used to suppress allocation strategy oscillations in Example 6, and both are maintained by the same mechanism.

[0037] The federated privacy unit performs norm clipping and noise addition on the uploaded model gradients before federated aggregation. Each hospital's client performs local stochastic gradient descent, clips the gradients according to the L2 norm, and adds calibrated Gaussian noise. The server then aggregates the data based on the amount of local data from each client. Through this differential privacy federated learning strategy, the risk of leakage of sensitive patient information is reduced to below 0.15%, and the data localization rate reaches 98%.

[0038] This embodiment details the process by which the multi-agent collaborative scheduling module 2, the chain-like reasoning module 3, and the edge-cloud collaborative optimization module 4 form a closed loop through the reasoning cost signal, thereby generating collaborative gain. The data flow is as follows: Figure 4 As shown.

[0039] The reasoning cost signal output by the chain-thinking reasoning module 3 Due to the influence of the target agent's own reasoning template and knowledge scope, the reasoning costs of different agents are not directly comparable. To make the reasoning cost a unified reward, the multi-agent collaborative scheduling module 2 first normalizes the reasoning cost signal relative to the historical reasoning cost distribution of the target agent into a dimensionless cost, which is expressed as: , in: The cost is dimensionless, and is a scalar with a range of values ​​of . This formula maps the inference cost signal to its quantile in the historical distribution, representing the degree of struggle of the target agent in this inference relative to its own historical level. For target role intelligent agent The historical reasoning cost empirical cumulative distribution function; For target role intelligent agent Historical reasoning cost sample set, Let be the cardinality of the set, and be a positive integer. This is a value of the reasoning cost in the historical reasoning cost sample set; The inference cost signal is defined in the same way as the inference cost signal formula in Example 4. The function is an indicator function, taking the value 1 when the condition within the parentheses is true, and 0 otherwise. The multi-agent collaborative scheduling module 2 also constructs a character agent reasoning ability profile based on the historical reasoning cost distribution statistics of each character agent. The character agent reasoning ability profile feeds back into the professionalism dimension of the task complexity score for role adaptation, so that character agents with lower reasoning costs (i.e., more relaxed reasoning) are adapted accordingly in the professionalism assessment.

[0040] Multi-agent cooperative scheduling module 2 constructs a dense reward with a dimensionless cost and then weights and fuses it with a sparse reward. The fused reward is represented as: , in: For the first The fusion reward for each medical visit is a scalar quantity, with a value range of [value missing]. The value is calculated using this formula, and a larger value indicates a more reasonable allocation. For the first The dimensionless cost of each medical visit is defined in the same way as the aforementioned normalization formula. The negative sign at the beginning indicates that the less struggle there is in reasoning, the higher the reward, forming a dense reward that can be obtained immediately with each reasoning. The reward is sparse and scalar. It is set to 1 when the data for this accompanying medical visit decision is adopted, and 0 otherwise. It is only available after the medical visit result is obtained. Let be the sparse reward fusion coefficient, and let be a scalar with a value range of . We set it to 0.3 to balance dense and sparse rewards.

[0041] The multi-agent cooperative scheduling module 2 uses fusion rewards as a return, optimizes the weighting coefficients and complexity thresholds of the reverse calibration task complexity score through online strategies, and superimposes a stability regularization term based on Fisher information during the update. The update rule is expressed as follows: , in: , The first Next and first The strategy parameter vector for each visit is a five-dimensional vector, and its components are weighted coefficients for the task complexity score. , , With complexity threshold , It is updated iteratively by this formula; Let be the learning rate, and be a scalar with a range of values ​​of . It is 0.01, has no unit, and is used to control the step size of each calibration. The fusion reward is defined in the same way as the aforementioned fusion reward formula; In the strategy parameters Below, given task complexity score Tasks are assigned to role-specific intelligent agents. The probability of the allocation strategy, with a value range of . ; For the first The target role intelligent agent assigned to the next medical visit task; A gradient operator for calculating the gradient of the policy parameter vector; It is the natural logarithm; The stability regularity strength is denoted by , which is a scalar and has a range of values ​​of . The value is 100, with no unit. The larger the value, the stronger the protection against previous allocation strategies. , is the Fisher information diagonal vector, is a five-dimensional vector, and its components are non-negative. It is calculated from historical medical visits and is used to measure the importance of each component of the strategy parameters to the previous allocation strategy. The vector of policy parameters that has converged in the past is a five-dimensional vector; This is the element-wise multiplication operator for vectors. The Fisher information mentioned is the same estimate used in the incremental learning unit of the edge-cloud collaborative optimization module 4 to suppress catastrophic forgetting in the model, thus ensuring that the stability of the allocation strategy and the anti-forgetting capabilities of the edge model are supported by the same mechanism.

[0042] Through the aforementioned normalization, dense reward construction, and online policy optimization with stability regularization, a closed loop is formed between the inference cost signal output by the chain-thinking reasoning module 3 and the allocation decision of the multi-agent collaborative scheduling module 2. This closed loop creates positive feedback between the multi-agent collaborative scheduling module 2, the chain-thinking reasoning module 3, and the edge-cloud collaborative optimization module 4: the improved task allocation accuracy reduces the mean of the inference cost signal, causing the estimated variance of the dimensionless cost to converge with the number of medical tasks, and the complexity threshold tends to a stable fixed point; due to the constraint of the stability regularization term, the complexity threshold does not oscillate during the convergence process, nor does it forget previously effective allocation experience. From an engineering perspective, the stable fixed point converged by this closed loop corresponds to a set of weighting coefficients and complexity thresholds that match the task complexity score with the actual inference difficulty: at this fixed point, tasks judged as high complexity and assigned to the medical knowledge role agent do indeed have a high actual inference cost; tasks judged as low complexity and assigned to the process scheduling role agent do indeed have a low actual inference cost. In other words, the closed-loop calibration gradually pulls the initially set static weighting coefficients and complexity thresholds, which may deviate from the actual reasoning difficulty, back to values ​​consistent with the reasoning difficulty. The stability regularization term ensures that this pullback process does not fluctuate drastically due to individual abnormal patient visits. Compared to schemes with fixed weighting coefficients and fixed complexity thresholds, this closed loop can adapt to the reasoning difficulty distribution of different departments and patient groups without repeated manual parameter tuning. Experimental results show that the system's response delay for accompanying patients at stable fixed points decreased from 2.8s when the closed loop was disconnected to 1.2s, the medical hallucination rate decreased from 8.7% to 2.3%, and the task allocation accuracy reached 92.3%. The simultaneous improvement in response delay and hallucination rate cannot be achieved by any single feature among normalization, dense reward, or stability regularization, but rather by the system-level synergistic gain emerging from the coupling of these three features with the four modules, demonstrating the nonlinear effect of the closed loop described in this invention.

[0043] This embodiment uses a follow-up visit scenario for an elderly patient as an example to illustrate the end-to-end processing of the system. The patient arrives at the hospital with a wearable device and inputs a voice complaint via the hospital's smart terminal, such as "recently high blood pressure, dizziness," while simultaneously uploading recent blood pressure monitoring time-series data sampled at 100Hz. The system reads the patient's electronic medical record text and captures facial expression images. The multimodal unified semantic module 1 encodes the above four multimodal patient inputs respectively: the voice is encoded after recognition and conversion to text, the blood pressure time-series data is encoded using time-frequency joint encoding, the medical record text is encoded using text encoding, and the facial image is encoded using visual encoding; the cross-modal attention fusion unit fuses the four features into a unified semantic representation with a dimension of 512.

[0044] Multi-agent cooperative scheduling module 2 calculates task complexity scores based on unified semantic representation, with modality dimension scores in this scenario. Urgency dimension score Professionalism score The task complexity score, calculated using the current weighted coefficients, is 0.85, exceeding the high complexity threshold of 0.70. Therefore, the medical visit task is assigned to the medical knowledge role agent. In the chain-thinking reasoning module 3, the medical knowledge role agent initializes a multi-path reasoning chain based on the hypertension diagnosis reasoning chain template and expands it through bundle search to select the optimal reasoning path. During reasoning and verification, logical consistency checks do not trigger re-reasoning, the bundle search pruning depth is 2, and knowledge graph matching checks trigger one knowledge error correction, generating a reasoning cost signal accordingly. After multiple verifications pass, the system outputs accompanying decision data, including recommended department routes, waiting time prediction, and visit process arrangement. For example, it recommends cardiology, predicts a waiting time of approximately 35 minutes, and provides a reminder to carry a blood pressure monitor and fasting blood sampling record.

[0045] Subsequently, the multi-agent collaborative scheduling module 2 normalizes the current inference cost signal relative to the historical inference cost distribution of the medical knowledge role agent into a dimensionless cost. It then constructs a weighted coefficient and complexity threshold for the task complexity score, which are optimized and updated via an online policy with stability regularization after integrating rewards. The edge-cloud collaborative optimization module 4 updates the edge model with the current consultation result through the incremental learning unit, and uploads the updated model to the cloud for federated aggregation through the federated privacy unit. Thus, the inference difficulty exposed in this consultation is retained as a priori for the next task allocation. The system continuously corrects the mismatch between task complexity assessment and inference difficulty during continuous operation, thereby continuously improving the accuracy and real-time performance of medical accompaniment decisions.

[0046] Those skilled in the art should understand that the above embodiments are merely preferred embodiments of the present invention, and any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A medical escort decision-making system based on multi-agent collaboration and CoT mechanism, characterized in that, include: The multimodal unified semantic module is used to perform cross-modal alignment and fusion of patients' multimodal medical input and output a unified semantic representation. The multi-agent collaborative scheduling module is configured with multiple role agents with different functions, which are used to calculate the task complexity score of the medical treatment task based on the unified semantic representation, and to allocate the medical treatment task to the target role agent according to the task complexity score and the complexity threshold. The chain-based reasoning module is used to generate a multi-path reasoning chain by the target role intelligent agent based on the reasoning template, perform multiple verifications on the multi-path reasoning chain, output the accompanying medical decision data, and generate a reasoning cost signal representing the intensity of retry and error correction during the reasoning and verification process. The edge-cloud collaborative optimization module is used to perform edge-side lightweight deployment and incremental updates of the models of the multimodal unified semantic module, the multi-agent collaborative scheduling module, and the chain-like reasoning module. The multi-agent collaborative scheduling module is also used to normalize the inference cost signal relative to the historical inference cost distribution of the target role agent into a dimensionless cost, and use the dimensionless cost as the reward for online policy optimization to reverse-calibrate the weighting coefficient of the task complexity score and the complexity threshold, so that the inference cost signal output by the chain thinking inference module and the allocation decision of the multi-agent collaborative scheduling module form a closed loop.

2. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 1, characterized in that, The reasoning cost signal is a weighted combination of the number of times the target agent triggers re-reasoning in the multiple verifications, the pruning depth of the multi-path reasoning chain, and the number of error corrections between the multi-path reasoning chain and the medical knowledge graph; the multiple verifications include logical consistency verification, knowledge graph matching verification, and clinical guideline comparison verification.

3. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 2, characterized in that, The normalization is to map the inference cost signal to its quantile in the historical inference cost distribution of the target role agent to obtain the dimensionless cost; the multi-agent collaborative scheduling module is also used to construct a role agent inference ability profile based on the historical inference cost distribution statistics of each role agent, and the role agent inference ability profile is used for role adaptation in the professional dimension of the task complexity score.

4. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 3, characterized in that, The online strategy optimization constructs a dense reward based on the dimensionless cost. The dense reward is updated immediately after each medical visit, taking into account the weighting coefficients of the modal dimension, urgency dimension, and professionality dimension of the task complexity score, as well as the complexity threshold. The dense reward is then weighted and fused with the sparse reward corresponding to whether the accompanying medical visit decision data is adopted.

5. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 4, characterized in that, The weighting coefficients and the complexity threshold are updated and superimposed with a stability regularization term. The stability regularization term imposes constraints on important parameters of the previous allocation strategy based on Fisher information calculated on historical medical tasks, so as to suppress the allocation strategy oscillation caused by the closed loop. The Fisher information on which the stability regularization term is based shares the same Fisher information estimate with the incremental update of the edge-cloud collaborative optimization module.

6. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 5, characterized in that, The multi-agent collaborative scheduling module, the chain-like reasoning module, and the edge-cloud collaborative optimization module form a positive feedback loop: the improvement in task allocation accuracy reduces the mean of the reasoning cost signal, causing the estimated variance of the dimensionless cost to converge with the number of medical visits, and the complexity threshold tends to a stable fixed point; the system's accompanying decision response delay and hallucination rate at the stable fixed point are both lower than the corresponding values ​​when the closed loop is broken.

7. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 1, characterized in that, The multimodal unified semantic module includes a visual-text alignment unit, a cross-modal attention fusion unit, and a temporal physiological signal encoding unit. The temporal physiological signal encoding unit is used to perform time-frequency joint encoding of multi-channel physiological signals by time-domain encoding and frequency-domain encoding. The cross-modal attention fusion unit is used to fuse text, image, and speech features into the unified semantic representation.

8. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 1, characterized in that, The multi-agent collaborative scheduling module transmits messages between its various roles through a medical domain extended communication protocol. The message body of the medical domain extended communication protocol carries a digest-compressed inference chain, inference confidence, and a list of clinical guideline citations. The message body is also subjected to message compression and connection pool reuse.

9. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 1, characterized in that, The edge-cloud collaborative optimization module includes a model quantization unit, an incremental learning unit, and a federated privacy unit. The model quantization unit is used to perform low-bit quantization on the model weights, the incremental learning unit is used to perform incremental fine-tuning based on the low-rank adaptation matrix, and the federated privacy unit is used to perform norm pruning and noise addition on the uploaded model gradients before federated aggregation.

10. The medical escort decision-making system based on multi-agent collaboration and CoT mechanism according to claim 1, characterized in that, The accompanying medical decision data includes the routing of medical departments, prediction of waiting time, and arrangement of medical procedures. The accompanying medical decision data consists of the technical parameters of the in-hospital medical procedure.

Images