Outpatient accompanying service and quality evaluation method and system based on multi-agent cooperation

Abstract

The present application belongs to the technical field of intelligent accompanying diagnosis service, and particularly relates to a method and system for outpatient accompanying diagnosis service and quality evaluation based on multi-agent cooperation, aiming to solve the problem that the quality of accompanying diagnosis service is difficult to be objectively evaluated, and the evaluation result is difficult to be used for task adjustment and subsequent task allocation. The accompanying diagnosis service order is split into task nodes and distributed, audio, location trajectory, business timestamp and text data are collected to calculate service quality evaluation indexes, intervention is carried out according to the evaluation value and task state, and the task is continued to be executed based on the task context when the intelligent agent is switched; the evaluation result is formed after the service, the unqualified indexes are determined, the subsequent task matching parameters are formed according to the unqualified indexes and fed back to the central dispatch intelligent agent, so as to adjust the matching priority of the subsequent order. The present application can improve the objectivity of service quality evaluation and the pertinence of task scheduling and personnel management.

Description

Technical Field

[0001] This invention relates to the field of intelligent escort service technology, and in particular to an outpatient escort service and quality assessment method and system based on multi-agent collaboration. Background Technology

[0002] With the accelerating aging of the population, the increasing demand for cross-regional medical treatment, and the increasingly complex departmental distribution and treatment processes within large general hospitals, the traditional method of family members accompanying patients to medical appointments faces certain difficulties in terms of time management, physical burden, and familiarity with hospital procedures. As a result, outpatient escort services have gradually developed. Escort services not only provide elderly patients with assistance such as registration, waiting, examinations, payment, medication dispensing, and hospital guidance, but their target audience has also gradually expanded to include pregnant women, children, and other groups with special medical needs. Existing regulations such as the "Professional Skills Standards for Escorts" and "Escort Service Standards" have set forth corresponding requirements for escort services in terms of medical knowledge, service awareness, professional ethics, and service content. However, how to implement these standards in every specific escort service and conduct a full-process quality evaluation of the services actually provided by escorts remains a problem that needs to be solved in the management of escort services.

[0003] In the prior art, for example, Chinese invention patent CN112242195A discloses an intelligent escort system and its escort method, which mainly includes an information collection platform, an information identification platform, a cloud storage platform, and an information processing platform. The information processing platform is equipped with functions such as triage, appointment registration, medical orders, route planning, and query. It can also locate patients and provide in-hospital navigation via mobile terminals, and assist in calculating and pushing fees through payment services. The above technologies can, to a certain extent, assist patients in managing their medical process, navigating within the hospital, and handling related business.

[0004] However, existing escort systems typically operate according to pre-set workflows, with communication primarily revolving around information exchange. When unexpected situations arise during the escort process, such as temporary closure of examination departments, changes in available appointment slots, or malfunctions of self-service equipment, the existing system struggles to autonomously adjust the allocation of affected tasks based on the current task execution status, dependencies between tasks, and changes in service resources. It may still require continuing with the original workflow or relying on manual intervention, impacting the continuity and flexibility of escort services in abnormal situations.

[0005] Meanwhile, the quality of medical accompanying services does not only depend on whether services such as appointment registration, payment, and medicine collection are completed, but also involves multiple aspects including whether the accompanying staff are familiar with the hospital environment and medical consultation procedures, whether their service attitude is friendly, enthusiastic and patient, whether they can respond to patients' consultations in a timely manner, whether their problem explanations are clear and accurate, and whether the communication process is smooth. Existing intelligent medical accompanying systems mainly focus on process scheduling and physical space positioning, and can obtain information such as whether a patient or accompanying staff has arrived at the corresponding location and whether related services are completed. However, for service process details such as whether the accompanying staff take obvious detours while leading the way, whether they show negative emotions when communicating with patients, and the quality of their answers to patients' questions, there is usually a lack of means for continuous quantitative evaluation based on service process data. The quality evaluation of existing medical accompanying services still largely relies on post-service patient feedback or manual evaluation, the evaluation results are susceptible to the influence of subjective feelings and post-event memory, and it is also difficult to accurately trace the specific service link that causes service quality problems.

[0006] In addition, the task execution process and service quality management in existing medical accompanying services are usually relatively independent, and the service quality evaluation results are mainly used for statistics after service completion or personnel management, making it difficult to feed back to the task scheduling link in a timely manner during the medical accompanying service. When the service quality of the accompanying staff declines in the process of communication and interaction, route guidance or abnormal problem handling, existing systems are difficult to adjust task allocation or service strategies in a timely manner according to the corresponding quality status; at the same time, existing evaluation methods usually focus on forming the overall evaluation result of a single service, and it is difficult to further determine the specific service capability dimensions where the accompanying staff fail to meet the standards from different service quality evaluation indicators, so that there is a lack of effective connection between the evaluation results and targeted training for the accompanying staff as well as subsequent task matching, making it difficult to form a continuous feedback process from service execution and quality evaluation to personnel capability improvement and subsequent service scheduling.

[0007] Therefore, how to use the multi-source service data formed during the medical accompanying service process to objectively quantify different service quality dimensions, enable the service quality evaluation results to participate in the execution of medical accompanying tasks, and determine the weak service capability links of the accompanying staff according to the evaluation results of different service quality evaluation indicators, so as to further act on personnel training and subsequent task matching, is a problem that needs to be further solved in the current outpatient medical accompanying service quality management. Summary of the Invention

[0008] This invention aims to address the problems of existing outpatient escort systems, which mainly focus on task process management and struggle to objectively quantify multiple service quality evaluation indicators based on audio data, text data, location trajectories, and business timestamps generated during the escort service. Furthermore, the service quality evaluation results are disconnected from task scheduling, making it difficult to intervene in task execution based on real-time evaluation status. At the same time, existing evaluation methods are unable to further identify specific service quality dimensions that escorts have not met, resulting in a lack of effective connection between evaluation results and escort training and subsequent task matching.

[0009] In a first aspect, this invention proposes a method for outpatient escort service and quality assessment based on multi-agent collaboration, comprising: The order for accompanying medical services is broken down into task nodes with time-dependent relationships. The central scheduling agent allocates the task nodes according to the capability declarations of candidate agents and the current load, and records the task dependencies and task status through a shared task graph. Collect data on the process of accompanying medical visits, including audio data, location trajectory, business timestamps, and text data transcribed from the audio data; The evaluation values ​​of multiple preset service quality evaluation indicators are calculated using the data from the accompanying medical service process. Each of the aforementioned evaluation values ​​is fed back to the central scheduling agent, which intervenes in the task execution based on each of the aforementioned evaluation values ​​and the current task status. When the intervention causes the execution agent of the task node to switch, the takeover agent continues to execute the task node based on the corresponding task context. After the service is completed, each evaluation value is weighted according to the preset evaluation weights corresponding to each service quality evaluation indicator to obtain the service quality evaluation result. Each evaluation value is compared with the corresponding evaluation threshold to determine the service quality evaluation indicators that do not meet the standards for the corresponding escort. The subsequent task matching parameters for the corresponding escort are formed based on the substandard service quality evaluation indicators and fed back to the central scheduling agent to adjust the matching priority of the corresponding escort's subsequent orders.

[0010] Furthermore, the allocation of the task nodes includes: The central scheduling agent issues task tenders to candidate agents for the task nodes to be assigned. The candidate agents generate bidding data based on their own capability declarations and current load; The central scheduling agent determines the execution agent based on the bidding data and assigns the task nodes to be allocated to the execution agent.

[0011] Furthermore, the shared task graph also records the binding relationship between task nodes and service resources; The emergency response agent identifies abnormal events and determines abnormal service resources; The central scheduling agent determines the affected task nodes based on the abnormal service resources, the binding relationships, and the task dependencies. It retains the execution results of completed task nodes, sets the executing task nodes bound to the abnormal service resources as paused and reassigns them, and moves the pending task nodes that depend on the affected task nodes into the reordering queue. The task nodes in the rearranged queue are reallocated according to the task dependencies.

[0012] Furthermore, regarding the service attitude evaluation indicator among the multiple service quality evaluation indicators: The negative sentiment values ​​for the voice and text are determined based on the audio data and the text data, respectively. The speech fusion weights and text fusion weights are adjusted based on the audio signal-to-noise ratio corresponding to the audio data and the text transcription confidence level corresponding to the text data. Based on the adjusted voice fusion weights and text fusion weights, the negative emotion values ​​of the voice and the text are weighted to obtain a negative emotion fusion value, and the evaluation value of the service attitude evaluation index is determined based on the negative emotion fusion value.

[0013] Furthermore, the intervention in task execution includes: When the evaluation value of the service quality evaluation index based on the model is lower than the corresponding evaluation threshold, and the corresponding judgment confidence is lower than the preset confidence threshold, or when a negative feature appears in a single round of dialogue, a service prompt is sent to the companion's terminal. When the negative emotion fusion value of multiple consecutive dialogue rounds reaches the preset takeover threshold, and the number of dialogue rounds reaches the preset round threshold, when the current communication interaction task node is being executed, the central scheduling agent will schedule the communication interaction agent to take over the communication interaction task node. When the number of times the communication and interaction task nodes in the same escort service order are taken over reaches a preset threshold, or when the preset key service quality evaluation indicators meet the preset task reconstruction conditions, the central scheduling agent will reconstruct the task.

[0014] Furthermore, the multiple service quality evaluation indicators also include process familiarity evaluation indicators and response timeliness evaluation indicators; The trajectory turnaround rate is calculated based on the planned path generated by the triage planning intelligent agent and the location trajectory, and the evaluation value of the process familiarity evaluation index is determined based on the trajectory turnaround rate. The audio data is subjected to voice endpoint detection, the average time difference between the end of the patient's question and the beginning of the companion's response is calculated, and the evaluation value of the response timeliness evaluation index is determined based on the comparison result of the average time difference and the preset response threshold.

[0015] Furthermore, the text data is encoded into a text feature matrix, and text semantic features are obtained through multi-head attention calculation. The text semantic features are then input into a text sentiment classifier to obtain the negative sentiment value of the text. The accompanying person's voice segments are separated from the audio data, and the Mel frequency cepstral coefficients, fundamental frequency, energy profile, and zero-crossing rate are extracted to form an acoustic feature matrix. The acoustic feature matrix is ​​convolved and pooled to obtain local frequency domain features. The local frequency domain features are input into a bidirectional long short-term memory network, and the obtained temporal coding results are attention-weighted to obtain the negative emotion value of the voice.

[0016] Furthermore, the evaluation values ​​of each service quality evaluation indicator in each accompanying medical visit are summarized according to the accompanying medical visit identifier, forming a historical trend profile of each service quality evaluation indicator for the corresponding accompanying medical visitor. Based on the historical trend profile, the personalized training focus and retraining course suggestions for the corresponding escorts are determined, and the subsequent task matching parameters are adjusted according to the historical trend profile.

[0017] Furthermore, the takeover agent continuing to execute the task node based on the corresponding task context includes: The central scheduling agent maintains a task context object for the task node and configures a task context identifier for the task context object. The task context object includes a patient needs summary, a dialogue context pointer, a business processing progress snapshot, and the execution result of the preceding task. Before switching agents, freeze the task state of the task node and the task context object, and release the execution lock held by the original agent. The takeover agent reads the task context object based on the task context identifier, acquires the execution lock, and then continues to execute the task node.

[0018] In a second aspect, the present invention proposes an outpatient escort service and quality assessment system based on multi-agent collaboration, including a task orchestration module, a data acquisition module, an indicator evaluation module, a feedback scheduling module, and an evaluation management module; The task orchestration module is used to split the escort service order into task nodes with time-series dependencies. The central scheduling agent allocates the task nodes according to the capability declaration of the candidate agents and the current load, and records the task dependencies and task status through a shared task graph. The data acquisition module is used to collect data on the escort service process, which includes audio data, location trajectory, business timestamps, and text data transcribed from the audio data. The indicator evaluation module is used to calculate the evaluation values ​​of multiple preset service quality evaluation indicators using the data from the escort service process. The feedback scheduling module is used to feed back each of the evaluation values ​​to the central scheduling agent, intervene in the task execution according to each of the evaluation values ​​and the current task status, and when the intervention causes the execution agent of the task node to switch, the takeover agent continues to execute the task node based on the corresponding task context. The evaluation management module is used to weight each evaluation value according to the preset evaluation weight corresponding to each service quality evaluation indicator after the service is completed, to obtain the service quality evaluation result, compare each evaluation value with the corresponding evaluation threshold, determine the service quality evaluation indicators that do not meet the standards for the corresponding escort, form the subsequent task matching parameters for the corresponding escort based on the substandard service quality evaluation indicators, and feed the subsequent task matching parameters back to the central scheduling agent to adjust the matching priority of the corresponding escort's subsequent orders.

[0019] The beneficial effects of this invention are: This invention breaks down escort service orders into task nodes with temporal dependencies. A central scheduling agent allocates tasks based on the capability declarations of candidate agents and the current load. Simultaneously, a shared task graph records task dependencies and task status, enabling escort service tasks to be organized and adjusted according to the actual execution status. When a department is closed, appointment slots change, or service equipment malfunctions, the affected task nodes can be identified based on the relationship between abnormal service resources and task nodes. While retaining the execution results of completed tasks, the affected tasks are reallocated, thereby improving the continuity of escort service task execution under abnormal circumstances.

[0020] This invention utilizes audio data, text data, location trajectories, and business timestamps generated during the escort service to generate evaluation values ​​for multiple service quality indicators. This allows service quality dimensions such as service attitude, process familiarity, and response timeliness to be quantitatively analyzed through service process data, and enables the evaluation results to be mapped to specific service links, thereby improving the objectivity and traceability of the escort service quality evaluation.

[0021] For service attitude evaluation, this invention determines negative sentiment values ​​for both speech and text, and adjusts the speech fusion weights and text fusion weights based on the audio signal-to-noise ratio and text transcription confidence. When environmental noise or the reliability of text transcription changes, the contribution of the speech modality and text modality to the negative sentiment fusion result can be adjusted accordingly, thereby reducing the impact of changes in the reliability of a single modality on the service attitude evaluation result.

[0022] This invention feeds back the evaluation values ​​of each service quality evaluation indicator to the central scheduling agent and intervenes in task execution in conjunction with the current task status. This allows service quality evaluation to participate in the ongoing escort service, rather than just being used for statistical evaluation after the service ends. When the intervention causes a switch in the execution agent of a task node, the takeover agent can continue to execute the task node based on the corresponding task context, thereby reducing the loss of patient needs, dialogue content, and business processing progress caused by the switch in the execution agent and realizing feedback linkage between service quality evaluation and the task execution process.

[0023] This invention weights each evaluation value according to a preset evaluation weight corresponding to each service quality evaluation indicator after the service is completed, forming a service quality evaluation result. Each evaluation value is then compared with its corresponding evaluation threshold to determine the corresponding unmet service quality evaluation indicators for the accompanying staff. This ensures that the service quality evaluation not only reflects the overall quality of a single accompanying service but also identifies specific service quality dimensions that the accompanying staff needs to improve. Furthermore, based on the unmet service quality evaluation indicators, the invention determines the training focus for the corresponding accompanying staff and, combined with the accompanying staff's previous evaluation results, creates a historical trend profile for personalized training, retraining course suggestions, and subsequent task matching. This allows the service quality evaluation results to be linked to accompanying staff capability improvement and subsequent service scheduling, improving the relevance of training and task matching. Attached Figure Description

[0024] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the outpatient escort service and quality assessment method in the first embodiment of the present invention; Figure 2 This is a flowchart illustrating the task node allocation and abnormal reordering process in the first embodiment of the present invention; Figure 3 This is a flowchart illustrating the calculation process of service quality evaluation indicators in the first embodiment of the present invention; Figure 4 This is a flowchart illustrating the task execution intervention process in the first embodiment of the present invention; Figure 5 This is a schematic diagram of the architecture of the outpatient escort service and quality assessment system in the second embodiment of the present invention. Detailed Implementation

[0025] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] The first embodiment of this invention proposes a method for outpatient escort service and quality assessment based on multi-agent collaboration, including the following steps: S10: The escort service order is split into task nodes with time-series dependencies. The central scheduling agent allocates the task nodes according to the capability declaration of the candidate agents and the current load, and records the task dependencies and task status through a shared task graph. S20: Collect data on the escort service process, including audio data, location trajectory, business timestamp, and text data transcribed from the audio data; S30: Calculate the evaluation values ​​of multiple preset service quality evaluation indicators using the data from the accompanying medical service process; S40: Feed back each of the evaluation values ​​to the central scheduling agent, intervene in the task execution according to each of the evaluation values ​​and the current task status, and when the intervention causes the execution agent of the task node to switch, enable the takeover agent to continue executing the task node based on the corresponding task context. S50: After the service is completed, each evaluation value is weighted according to the preset evaluation weight corresponding to each service quality evaluation indicator to obtain the service quality evaluation result. Each evaluation value is compared with the corresponding evaluation threshold to determine the unqualified service quality evaluation indicators of the corresponding escort. The unqualified service quality evaluation indicators are used to form the subsequent task matching parameters of the corresponding escort. The subsequent task matching parameters are fed back to the central scheduling agent to adjust the matching priority of the corresponding escort's subsequent orders.

[0028] To more clearly illustrate the outpatient escort service and quality assessment method based on multi-agent collaboration of the present invention, the steps in this embodiment are further explained below.

[0029] S10: The escort service order is split into task nodes with time-series dependencies. The central scheduling agent allocates the task nodes according to the capability declaration of the candidate agents and the current load, and records the task dependencies and task status through a shared task graph.

[0030] Specifically, after a patient makes an appointment for a medical escort service through a mobile terminal, a medical escort service order is formed. The central scheduling intelligent agent receives the medical escort service order and, based on the medical items and hospital procedures corresponding to the medical escort service order, breaks down the medical escort service order into task nodes with time-dependent relationships.

[0031] In this embodiment, the task nodes include both information processing or business handling tasks directly executed by a specialized intelligent agent, and service tasks organized and maintained by a specialized intelligent agent and completed on-site by the accompanying staff. The task nodes may include appointment registration, arrival at the hospital, waiting in the department, accompanying for examinations, payment, medication pickup, medical order delivery, and service completion. Different task nodes form temporal dependencies based on the outpatient business process. For example, the payment task node depends on the payment items generated during the corresponding diagnosis or examination process, the medication pickup task node depends on the payment results generated by the payment task node, and the medical order delivery task node depends on the medical order information generated after the doctor completes the diagnosis. Task nodes that do not have direct temporal dependencies and do not have service resource conflicts can be executed in parallel.

[0032] In this embodiment, the allocation of task nodes includes: Step S11: The central scheduling agent issues task tenders to candidate agents for the task nodes to be assigned. Step S12: The candidate intelligent agent forms bidding data based on its own capability declaration and current load; Step S13: The central scheduling agent determines the execution agent based on the bidding data and assigns the task nodes to be allocated to the execution agent.

[0033] Specifically, the central scheduling agent uses the Contract Network protocol for task bidding. The central scheduling agent issues task bidding to candidate agents for the task nodes to be assigned. Each candidate agent forms bidding data based on its own capability declaration and current load and returns it to the central scheduling agent. The central scheduling agent determines the execution agent to execute the corresponding task node based on the bidding data and assigns the task node to the execution agent.

[0034] The capability declaration indicates the business functions that the candidate intelligent agent can directly execute or manage. The triage and planning intelligent agent executes patient flow arrangement, in-hospital route planning and navigation guidance, and queries department locations, queue lengths, and appointment status in real time; the communication and interaction intelligent agent, based on a large language model and medical knowledge graph, performs patient consultation responses, interprets and relays medical orders, and provides reminders about precautions for patient visits; the business processing intelligent agent executes business nodes such as registration, payment, medication pickup, and appointment scheduling, and generates business status receipts; the emergency response intelligent agent identifies abnormal events in real time, formulates alternative solutions based on the abnormal events, and requests the central dispatch intelligent agent to adjust tasks.

[0035] The current load is used to reflect the current task status of the candidate agents, so that the central scheduling agent can further consider the current task carrying status of different candidate agents based on the fact that the candidate agents have the ability to complete the corresponding task nodes.

[0036] Each agent exchanges messages via an event bus using publish and subscribe methods, and reads and writes to a shared blackboard database, in which a shared task graph is stored. The shared task graph records the task dependencies between task nodes and the task status of each task node.

[0037] For example, the triage planning agent generates the estimated waiting time based on the hospital's operational status and publishes the corresponding status information. After the communication and interaction agent subscribes to the status information, it proactively explains the waiting situation and corresponding precautions to the patient, thereby realizing status sharing and collaborative execution among different agents.

[0038] In this embodiment, the shared task graph also records the binding relationship between task nodes and service resources.

[0039] Step S14: The emergency response agent identifies abnormal events and determines abnormal service resources; Step S15: The central scheduling agent determines the affected task nodes based on the abnormal service resources, the binding relationship, and the task dependency relationship; retains the execution results of completed task nodes; sets the executing task nodes bound to the abnormal service resources as paused and reassigns them; and moves the pending task nodes that depend on the affected task nodes into the reorder queue. Step S16: Reallocate task nodes in the rearranged queue according to the task dependency relationship.

[0040] Specifically, the central scheduling agent maintains five task states for task nodes: Pending, Running, Done, Suspended, and Cancelled. It also records the task dependencies between task nodes and the binding relationships between task nodes and service resources in the shared task graph.

[0041] The service resources include departments, doctor appointment slots, self-service equipment, manual service windows, and examination equipment.

[0042] The emergency response intelligent agent identifies abnormal events based on the status of the hospital's business systems and information generated during the escort service. These abnormal events include equipment malfunctions, departmental closures, and changes in available appointment slots.

[0043] When identifying abnormal events based on text data, the emergency response agent uses natural language processing to identify abnormal trigger information such as "the machine is broken", "there are no more appointments", and "the department is closed".

[0044] When an abnormal event is a device malfunction, the central scheduling agent identifies the task nodes that directly use the malfunctioning device as affected task nodes based on the binding relationship between task nodes and abnormal service resources. For example, when a self-service payment device malfunctions, the payment task node bound to the self-service payment device is identified as an affected task node.

[0045] When an abnormal event is a department's suspension of outpatient services or a change in appointment slots, the central dispatching agent starts from the task node directly related to the abnormal service resource, and determines the scope of impact by moving backward along the task dependencies between task nodes. The downstream task nodes whose location constraints correspond to the abnormal department or whose input depends on the execution results of the affected task nodes are identified as the affected task nodes.

[0046] For affected task nodes that are already in the completed state, retain their execution results; for task nodes that are bound to abnormal service resources and are in the execution state, set the task nodes to paused and reassign tasks; for task nodes that depend on affected task nodes and are in the pending execution state, add them to the reordering queue; task nodes that are not identified as affected task nodes maintain their original task state.

[0047] The central scheduling agent reallocates task nodes in the rearranged queue based on the task dependencies between task nodes.

[0048] The emergency response intelligent agent also generates alternative plans based on abnormal events. For example, when a self-service payment machine malfunctions, the emergency response intelligent agent generates an alternative plan to transfer the payment to a human service window and sends the alternative plan to the central dispatch intelligent agent. The central dispatch intelligent agent then adjusts the corresponding task nodes, and the communication and interaction intelligent agent simultaneously explains the reason for the task change to the patient.

[0049] In other implementations, while keeping the task nodes, the task dependencies between task nodes, the task states, and the data relationships of the shared task graph unchanged, task allocation can also adopt an auction mechanism, centralized arbitration, or an allocation method based on a reinforcement learning policy network.

[0050] S20: Collect data on the process of accompanying medical visits, including audio data, location trajectory, business timestamps, and text data transcribed from the audio data.

[0051] In this embodiment, step S20 includes: Step S21: Collect the audio data, the location trajectory, and the service timestamp; Step S22: Perform noise suppression, voice endpoint detection, and voiceprint separation on the audio data to obtain the voice segment of the patient companion; Step S23: Perform automatic speech recognition on the voice segment of the companion to obtain the text data and the corresponding text transcription confidence level.

[0052] In step S21, after obtaining authorization from the patient and the accompanying person, data is collected on the accompanying service process through the terminal sensing device.

[0053] The terminal sensing device can be a smartphone carried by the accompanying person or a smart work badge integrating a microphone and indoor positioning module. It can also be a smartwatch, smart headphones, or other wearable devices with audio acquisition and location sensing capabilities.

[0054] Patients use their mobile devices to submit orders for escort services and to enter their final service evaluation after the service is completed.

[0055] The hospital is equipped with Bluetooth positioning beacons or Wi-Fi positioning devices, which work in conjunction with the terminal sensing devices carried by the accompanying person to form an indoor location trajectory.

[0056] The audio data is environmental audio information during the escort service; the location trajectory is formed by GPS positioning coordinates or Bluetooth / Wi-Fi indoor positioning coordinates; the business timestamp is used to record the actual occurrence time of registration, waiting, examination, payment, medication collection and other business nodes.

[0057] The intelligent patient guidance system simultaneously provides in-hospital route planning, navigation guidance, and waiting information based on department location, queue length, and availability of appointments.

[0058] In step S22, noise suppression is first performed on the acquired audio data to reduce background noise in the hospital lobby, waiting area and other public environments.

[0059] Then, voice endpoint detection is performed to identify the start and end times of the voice signal.

[0060] Further blind source voiceprint separation is performed to distinguish different sound sources from the mixed audio data that simultaneously contains the patient's voice and the companion's voice, and obtain the companion's voice fragment and the patient's voice fragment.

[0061] In step S23, the voice segment of the companion is input into the medical vertical field automatic speech recognition engine, the voice segment of the companion is transcribed into text data, and the text transcription confidence score corresponding to the text data is formed.

[0062] The audio signal-to-noise ratio is used to reflect the relative relationship between the effective speech signal and the ambient background noise, in order to characterize the reliability of the current audio mode.

[0063] In addition to the data on the aforementioned escort service process, when calculating other service quality evaluation indicators, we further obtain text or voice evaluations input by patients via mobile terminals after the service is completed, images collected by user terminals, original doctor's medical order texts, the hospital's historical business time baseline, and knowledge information from the medical knowledge graph.

[0064] Various data are linked based on the escort service orders and corresponding task nodes, so that the subsequent evaluation results can correspond to specific service orders, specific business nodes and corresponding service periods.

[0065] In one implementation, the data platform establishes data connections with the hospital information system, electronic medical record system, registration and payment system, pharmacy system, examination appointment system, and in-hospital self-service terminals through authorized data interfaces. The in-hospital self-service terminals include department check-in machines, self-service payment machines, and other terminal devices capable of tracking patient transaction status.

[0066] When a patient completes check-in at the department check-in machine, pays at the self-service payment machine, or completes the corresponding business operation at other in-hospital self-service terminals, the corresponding business system or in-hospital self-service terminal generates a business event including the business event type, the time of the business event, the terminal identifier, and the business status. The patient's corresponding department or business area is determined based on the deployment location corresponding to the terminal identifier, and the patient's current business process is determined based on the business event.

[0067] The data platform converts data from different hospital business systems and in-hospital self-service terminals into standard business events through a unified data adaptation interface, and receives and distributes them asynchronously through an event bus. These standard business events are associated with corresponding task nodes based on the escort service order identifier and task node identifier. The central scheduling agent and various specialized agents obtain patient check-in, waiting, payment, examination, medication pickup, and other business statuses by subscribing to the corresponding standard business events.

[0068] For business events generated by hospital business systems or in-hospital self-service terminals, a business timestamp is generated or calibrated based on the occurrence time of the corresponding business event. Different business events are transmitted asynchronously through an event bus. Each intelligent agent subscribes to the required business events according to the corresponding escort service order and task node, eliminating the need for a central scheduling intelligent agent to continuously poll different hospital business systems or in-hospital self-service terminals.

[0069] S30: Calculate the evaluation values ​​of multiple preset service quality evaluation indicators using the data from the accompanying medical service process.

[0070] Specifically, this embodiment uses an outpatient escort service quality evaluation index system to evaluate the quality of the escort service process.

[0071] The outpatient escort service quality evaluation index system adopts a three-level index structure. The first-level index includes structural indicators, process indicators, and outcome indicators. The combined weight of the structural indicators is 0.1326, the combined weight of the process indicators is 0.4337, and the combined weight of the outcome indicators is 0.4337. Nine second-level indicators and 30 third-level indicators are set.

[0072] Among the structural indicators, the combined weight of systems and regulations is 0.0274, under which the combined weight of companion service regulations is 0.0022 and the combined weight of privacy protection system is 0.0042; the combined weight of organizational management structure is 0.0546, under which the combined weight of clear organizational management structure is 0.0216 and the combined weight of completeness of business information platform is 0.0157.

[0073] The combined weight of pre-hospital accompaniment services in the process indicators is 0.1289, under which the combined weight of informed consent signing completion rate is 0.0341 and the combined weight of patient assessment and planning improvement is 0.0394; the combined weight of in-hospital accompaniment services is 0.1372, under which the combined weight of communication and care behavior standardization rate is 0.0219 and the combined weight of safe transfer qualification rate is 0.0276; the combined weight of post-hospital accompaniment services is 0.1676, under which the combined weight of health follow-up execution rate is 0.1676.

[0074] The combined weight of complaints and handling in the outcome indicators is 0.1354, under which the combined weight of complaint handling response rate is 0.0438 and the combined weight of complaint handling satisfaction rate is 0.0521; the combined weight of satisfaction is 0.1706, under which the combined weight of patient or family satisfaction is 0.1706.

[0075] Within the aforementioned evaluation index system, twelve core service quality evaluation indicators that can be quantitatively calculated based on multimodal data were further selected.

[0076] The first item is the dress code evaluation index. The quality supervision intelligent agent uses computer vision object detection models to identify uniform and name tag features based on images collected from user terminals, or extracts entities and evaluation tendencies related to dress codes from patient evaluation texts through natural language processing.

[0077] The second item is the professionalism evaluation index. The communication and interaction intelligent agent obtains the dialogue text of the accompanying person generated by automatic speech recognition, and the quality supervision intelligent agent calculates the similarity between the vocabulary or semantic content used by the accompanying person and the medical knowledge base and standardized triage scripts, and forms the evaluation value of the professionalism evaluation index based on the similarity.

[0078] The third item is the service attitude evaluation index. Based on the voice, physical acoustic characteristics, and text semantic information of the escort, negative emotions such as anger, impatience, and indifference are identified to evaluate the escort's friendliness, enthusiasm, and patience during the service process.

[0079] The fourth item is the process familiarity evaluation index. Based on the location trajectory and business timestamp, and using the planned path generated by the triage planning agent as the baseline, the smoothness of the trajectory and the number of backtracking incidents of the accompanying person during the movement between different departments are analyzed, and the business timestamp is used to determine whether the person arrives at the target department at the corresponding time.

[0080] The fifth item is the response timeliness evaluation index. Based on the time information generated by the voice endpoint detection, the average time difference between the end of the patient's question and the start of the companion's response is calculated, and the evaluation value is determined according to the preset response threshold.

[0081] The sixth item is the evaluation index for clarity, accuracy, and professionalism of the answers. Syntactic dependency analysis is performed on the transcribed text to determine the logical coherence of the companion's responses, and the occurrence of meaningless pauses and interjections such as "uh" and "that" is statistically analyzed based on the audio data.

[0082] The seventh item is the respect and care evaluation indicator. Based on the patient's voice or text feedback after the service is completed, the emotional polarity in the feedback is analyzed, and multi-head attention mechanism is used to identify semantic content such as "thank you" and "thoughtful" that can represent that the patient feels respected and cared for.

[0083] The eighth indicator is the communication fluency evaluation index. This is based on the overlap between the companion's and the patient's voices, as well as the number of conversation rounds. When the companion frequently interrupts the patient, increasing the overlap between their voices, the evaluation value of this communication fluency index is lowered accordingly.

[0084] The ninth item is the demand grasp evaluation index. The central dispatching intelligent agent compares the initial order demand recorded in the reservation stage with the actual business node execution results, invoices, and location trajectories, and forms the evaluation value of the demand grasp evaluation index based on the degree of consistency between the two.

[0085] The tenth item is the evaluation indicator for effective information delivery. During the interpretation of medical orders, the communication and interaction intelligent agent obtains the original text of the doctor's medical orders and the text relayed to the patient by the accompanying person. Through automatic summarization and core medical entity coverage calculation, the corresponding evaluation value is determined.

[0086] The eleventh item is the task speed evaluation indicator. The business processing intelligent agent records the actual processing time for payment, medication collection, and other business nodes. The quality supervision intelligent agent compares the actual business processing time with the hospital's historical average processing time baseline and calculates the efficiency quantile.

[0087] The twelfth item is the evaluation indicator for problem and difficulty resolution. The emergency response intelligent agent monitors abnormal trigger information in real time, and the quality supervision intelligent agent forms an evaluation result based on the time between the occurrence of the abnormal event and the companion's proposal of an alternative solution, as well as the patient's acceptance of the alternative solution.

[0088] In this embodiment, for the service attitude evaluation indicator among the multiple service quality evaluation indicators, step S30 includes: Step S31: Determine the negative sentiment value of the voice and the negative sentiment value of the text based on the audio data and the text data, respectively; Step S32: Adjust the speech fusion weight and text fusion weight according to the audio signal-to-noise ratio corresponding to the audio data and the text transcription confidence level corresponding to the text data; Step S33: Based on the adjusted speech fusion weights and text fusion weights, the negative emotion values ​​of the speech and the text are weighted to obtain a negative emotion fusion value, and the evaluation value of the service attitude evaluation index is determined based on the negative emotion fusion value.

[0089] In step S31, the negative sentiment value of the text and the negative sentiment value of the voice are determined in the following ways.

[0090] The text data is encoded into a text feature matrix, and the text semantic features are calculated through multi-head attention. The text semantic features are then input into a text sentiment classifier to obtain the negative sentiment value of the text.

[0091] Specifically, the text data is encoded to obtain a text feature matrix. Query matrix Key matrix Sum matrix They are obtained according to the following relationships: ; ; ; in, Represents the text feature matrix; Indicates querying the projection matrix; Represents the key projection matrix; Represents the projection matrix of values; Represents the query matrix; Represents the key matrix; Represents a value matrix.

[0092] The output of a single attention head is: ; in, This represents the output of a single attention head; Represents the probability normalization function; superscript Indicates matrix transpose operation, not technical parameters; This represents the feature dimension of the key vector; This represents the relative position offset matrix.

[0093] Relative position offset matrix Used to represent the relative temporal relationship between different lexical units. For example, "the patient was still confused after the escort explained" and "the patient was confused and then the escort explained" contain some of the same textual information, but their different order corresponds to different service quality semantics.

[0094] Multiple attention heads perform the aforementioned attention calculations, and the outputs of each attention head are concatenated along the feature dimension to obtain the text semantic features. By using multiple attention heads, semantic and emotional information in communication texts and end-user feedback texts from different semantic relationships can be captured.

[0095] The semantic features of the text are input into a text sentiment classifier to obtain the probabilities corresponding to different sentiment categories, and the probabilities corresponding to negative sentiment categories are used as the negative sentiment value of the text.

[0096] In one implementation, the bag-of-words features formed based on the sentiment lexicon are combined with the text semantic features obtained by multi-head attention computation to help form the negative sentiment value of the text.

[0097] The accompanying person's voice segments are separated from the audio data, and the Mel frequency cepstral coefficients, fundamental frequency, energy profile, and zero-crossing rate are extracted to form an acoustic feature matrix. The acoustic feature matrix is ​​convolved and pooled to obtain local frequency domain features. The local frequency domain features are input into a bidirectional long short-term memory network, and the obtained temporal coding results are attention-weighted to obtain the negative emotion value of the voice.

[0098] Specifically, Mel frequency cepstral coefficients, fundamental frequency, energy profile, and zero-crossing rate are extracted from the voice segments of the accompanying person, and arranged in chronological order to form an acoustic feature matrix.

[0099] In one implementation, the acoustic feature matrix input includes a speech emotion recognition network consisting of four convolutional layers, four pooling layers, and a bidirectional long short-term memory network.

[0100] The first convolutional layer has 64 output channels, the second convolutional layer has 128 output channels, and the subsequent two convolutional layers each have 256 output channels. The kernel size is 5, the pooling window size and stride are both set to 2, and the output unit size of the bidirectional long short-term memory network is 128.

[0101] Convolutional layers are used to extract local frequency domain texture changes, and bidirectional long short-term memory networks extract temporal dependency information of speech emotion changes along the forward and backward time directions.

[0102] set up This indicates that the bidirectional long short-term memory network is in the first... The encoded vector output at time step i is used to perform a dot product with the sentiment evaluation parameter matrix to obtain the i-th time step i. The sentiment evaluation results at each moment: ; in, Indicates the first The encoding vector corresponding to each time step; Indicates the time sequence number; This represents the sentiment evaluation parameter matrix obtained by training with sentiment-labeled training data; Indicates the first The emotional evaluation result corresponding to each moment.

[0103] Calculate the first based on the emotional evaluation results at each time point. Attention weights corresponding to each time step: ; in, Indicates the first Attention weights corresponding to each moment; Represents an exponential function; This represents the summation index during the attention weight normalization process; This indicates the total number of time points included in the timing coding result; Indicates the first The encoding vector corresponding to each time step; Indicates the first The emotional evaluation result corresponding to each moment.

[0104] The speech representation vector is obtained by weighting each encoded vector according to the attention weight corresponding to each time step: ; in, This represents a speech vector.

[0105] Therefore, short speech segments with clear emotional polarity can receive greater attention weight, thus avoiding excessive dilution of short-term emotions such as impatience by longer neutral speech segments.

[0106] The speech representation vector The input is processed by an emotion classifier to obtain the probabilities corresponding to different emotion categories. The emotion categories include friendly, neutral, irritable, angry, and indifferent. The negative emotion value of the speech is formed based on the probability of belonging to a negative emotion category.

[0107] The sentiment evaluation parameter matrix can be trained using data containing sentiment labels. The training data includes the interactive sentiment binary motion capture corpus IEMOCAP and the Berlin sentiment speech database EMO-DB.

[0108] In steps S32 and S33, the speech fusion weight and text fusion weight are adjusted according to the audio signal-to-noise ratio corresponding to the audio data and the text transcription confidence corresponding to the text data, and the negative emotion fusion value is obtained according to the adjusted speech fusion weight and text fusion weight.

[0109] Specifically, let the negative emotion value of the voice be... The negative sentiment value of the text is The speech fusion weight is The text fusion weight is Negative sentiment fusion value ,but: ; in, Indicates the negative sentiment value of the voice; Indicates the negative sentiment value of the text; Indicates speech fusion weights; Indicates text fusion weights; This indicates the negative emotion fusion value.

[0110] In one implementation, the corresponding fusion weights are determined based on preset intervals in which the audio signal-to-noise ratio (SNR) and text transcription confidence levels fall. When the audio SNR is high and the text transcription confidence level is high, both the speech fusion weight and the text fusion weight can be set to 0.5. When the audio SNR is low and the text transcription confidence level is high, the speech fusion weight can be reduced to 0.3 and the text fusion weight increased to 0.7. When the text transcription confidence level is low and the audio SNR is high, the speech fusion weight can be increased to 0.7 and the text fusion weight decreased to 0.3. The audio SNR interval, the text transcription confidence level interval, and the corresponding fusion weights can be preset or adjusted based on the actual operating environment and historical evaluation data.

[0111] When a patient's or accompanying person's regional accent causes numerous errors in the text generated by automatic speech recognition, reducing the confidence level of the text transcription, the speech fusion weight should be increased. This allows for a greater reliance on speech acoustic modalities to determine negative emotion fusion values.

[0112] Therefore, instead of using the static average of the speech and text modalities, the fusion weights of the two modalities are dynamically adjusted according to the current operating environment.

[0113] The quality supervision intelligent agent determines the negative emotional state corresponding to the current communication segment based on the negative emotional fusion value, and determines the evaluation value of the service attitude evaluation index based on the negative emotional states corresponding to each communication segment during the service process. When the level of negative emotion increases or the number of negative emotional events increases, the evaluation value of the service attitude evaluation index is correspondingly reduced.

[0114] In this embodiment, the multiple service quality evaluation indicators also include process familiarity evaluation indicators and response timeliness evaluation indicators.

[0115] Step S34: Calculate the trajectory turnaround rate based on the planned path generated by the triage planning intelligent agent and the location trajectory, and determine the evaluation value of the process familiarity evaluation index based on the trajectory turnaround rate; Step S35: Perform voice endpoint detection on the audio data, calculate the average time difference between the end time of the patient's question and the beginning time of the companion's response, and determine the evaluation value of the response timeliness evaluation index based on the comparison result of the average time difference and the preset response threshold.

[0116] In step S34, the triage planning agent generates a planned path based on the target department involved in the current task node and the internal access relationships within the hospital.

[0117] In one implementation, the trajectory reversal rate is determined based on the proportion of the trajectory length in the location trajectory that repeatedly traverses already traversed paths or involves reverse reversals to the total trajectory length. A higher trajectory reversal rate results in a lower process familiarity evaluation value. For example, a trajectory reversal rate of no more than 10% can be set to a higher level of process familiarity evaluation; a trajectory reversal rate of 10% to 30% can be set to a medium level; and a trajectory reversal rate greater than 30% can be set to a lower level. Specific thresholds can be adjusted based on the hospital layout and historical patient accompaniment trajectory data.

[0118] An increase in the trajectory reversal rate indicates a decrease in the companion's familiarity with the hospital environment or the corresponding outpatient business process, and consequently a decrease in the evaluation value of the process familiarity evaluation index.

[0119] For example, if the location trajectory shows that the accompanying person went back and forth three times in the corridor on the second floor of the outpatient department, causing an additional 20 minutes of time, the corresponding trajectory information will be used to reduce the evaluation value of the process familiarity evaluation index.

[0120] In step S35, the end time of the patient's question is determined based on the voice endpoint corresponding to the patient's voice segment, and the start time of the companion's response is determined based on the voice endpoint corresponding to the companion's voice segment. The time difference between the two in multiple effective dialogue rounds is calculated, and the average time difference is formed.

[0121] The evaluation value of the response timeliness evaluation index is determined based on the comparison between the average time difference and the preset response threshold.

[0122] S40: Feed back each of the evaluation values ​​to the central scheduling agent, intervene in the task execution according to each of the evaluation values ​​and the current task status, and when the intervention causes the execution agent of the task node to switch, enable the takeover agent to continue executing the task node based on the corresponding task context.

[0123] Specifically, the real-time evaluation results generated by the quality supervision intelligent agent are not only used for comprehensive evaluation after the service is completed, but are also fed back to the central dispatch intelligent agent in real time through the event bus, enabling the central dispatch intelligent agent to implement dynamic intervention during the service process based on each evaluation value and the current task status recorded in the shared task graph.

[0124] In this embodiment, the intervention in task execution includes: Step S41: When the evaluation value of the service quality evaluation index based on the model is lower than the corresponding evaluation threshold, and the corresponding judgment confidence is lower than the preset confidence threshold, or when a negative feature appears in a single round of dialogue, a service prompt is sent to the companion terminal. Step S42: When the negative emotion fusion value of multiple consecutive dialogue rounds reaches the preset takeover threshold, and the number of dialogue rounds reaches the preset round threshold, when the current communication interaction task node is being executed, the central scheduling agent schedules the communication interaction agent to take over the communication interaction task node. Step S43: When the number of times the communication and interaction task node in the same escort service order is taken over reaches a preset threshold, or when the preset key service quality evaluation indicators meet the preset task reconstruction conditions, the central scheduling agent performs task reconstruction.

[0125] In step S41, the service prompt is used when the service quality is slightly abnormal or the current quality assessment is insufficient to change the task execution subject. The service quality evaluation index based on model judgment refers to the service quality evaluation index that obtains the evaluation result through a classification model or recognition model and simultaneously forms the corresponding judgment confidence level.

[0126] In one implementation, when the evaluation value of a service quality evaluation indicator based on the model is lower than the corresponding evaluation threshold, but the corresponding judgment confidence is lower than 90%, or only a slight negative feature is identified in a single round of dialogue, the central scheduling agent sends a vibration prompt or service prompt to the companion terminal, such as outputting "Please pay attention to the tone of communication", without changing the task allocation of the current task node.

[0127] In step S42, when negative emotional events related to service attitude continue for multiple dialogue rounds, and the current task node is a communication and interaction task such as patient consultation response and medical order relay, the central scheduling agent schedules the communication and interaction agent to take over the current communication task.

[0128] In one specific implementation, the takeover operation is performed when the negative emotion fusion value reaches a preset takeover threshold, the confidence level of the corresponding negative emotion event reaches 90% or more and continues for two consecutive dialogue rounds, and a communication interaction task node is currently being executed.

[0129] During the takeover process, the original escorts continued to perform non-communication tasks such as queuing, accompanying patients, and guiding routes.

[0130] In step S43, when a communication and interaction task node in the same escort service order is taken over multiple times, the central scheduling agent performs task reconfiguration.

[0131] In one specific implementation, when the same escort service order experiences two or more takeover-level events, the central scheduling agent re-determines the executing entity of the unfinished task node, initiates an escort replacement request to the escort agency scheduling pool, writes the corresponding escort's capability profile into the retraining queue, and lowers the matching priority of the escort's subsequent orders.

[0132] When key service quality evaluation indicators such as process familiarity evaluation indicators or problem and difficulty resolution evaluation indicators seriously fail to meet the corresponding evaluation requirements, it is also considered as another situation that meets the pre-set task reconstruction conditions.

[0133] Wherein, when the intervention causes a switch in the executing agent of the task node, enabling the takeover agent to continue executing the task node based on the corresponding task context includes: Step S44: The central scheduling agent maintains a task context object for the task node and configures a task context identifier for the task context object. The task context object includes a patient needs summary, a dialogue context pointer, a business processing progress snapshot, and the execution result of the preceding task. Step S45: Before switching the execution agent, freeze the task state of the task node and the task context object, and release the execution lock held by the original execution agent. In step S46, the takeover agent reads the task context object according to the task context identifier, acquires the execution lock, and continues to execute the task node.

[0134] Specifically, the blackboard database maintains a TaskContext object for different task nodes and configures a corresponding TaskContext identifier ContextID for each TaskContext object.

[0135] In addition to patient needs summary, dialogue context pointer, business processing progress snapshot and previous task execution results, the task context object also includes escort service order identifier and escort's current location.

[0136] The patient needs summary is used to store the patient's confirmed needs for the current service task; the dialogue context pointer is used to locate the dialogue content formed before the task takeover; the business processing progress snapshot is used to store the processing status of current registration, payment, examination or medication collection, etc.; the previous task execution result is used to store the previous business results that the current task node can read; the companion's current position is used as the starting position when the triage planning agent re-executes path planning.

[0137] When a switch of execution agents is required, the central scheduling agent freezes the task state and task context object of the current task node and writes them to the blackboard database. The original execution agent stops holding the corresponding task state and releases the execution lock.

[0138] The takeover agent reads the task context object from the blackboard database based on the task context identifier ContextID, and continues to execute the task node after acquiring the execution lock.

[0139] For example, when the takeover agent is a communication and interaction agent, the communication and interaction agent reads the patient's needs summary, the current dialogue context, the business processing progress and the results of the previous task execution, and continues to execute the patient consultation response or medical order relay task.

[0140] Through the aforementioned task context inheritance, when the executing agent of a task node changes, patient needs, existing dialogue information, business processing progress, location, and previous business results can be continuously transmitted.

[0141] After the communication and interaction agent takes over, the quality supervision agent continues to perform quality evaluation on the new communication content.

[0142] In one implementation, when the takeover task is completed and no negative emotional events are detected again for 5 consecutive minutes, the central dispatching agent releases the takeover state, restores the original communication responsibilities of the companion, and writes the takeover process, the release process, and the task status change into the corresponding service log.

[0143] As a further explanation of the first embodiment of the present invention, when an abnormal event causes a task node to be reassigned, the central scheduling agent determines the task node directly related to the abnormal service resource based on the binding relationship between the abnormal service resource and the task node, and determines the affected downstream task node along the task dependency relationship; retains the execution results of the completed task nodes, sets the executing task node bound to the abnormal service resource as paused, and moves the pending task nodes that depend on the affected task node into the reordering queue; for the reassigned executing task node, freezes the task state and task context object of the task node and releases the execution lock held by the original executing agent, so that the takeover agent can read the task context object and acquire the execution lock to continue executing the task node.

[0144] In this embodiment, the shared task graph simultaneously records the task dependencies between task nodes and the binding relationships between task nodes and service resources. When the emergency response intelligent agent recognition device malfunctions, a department suspends services, or appointment slots change, the task nodes directly associated with the abnormal service resources are first determined based on the abnormal service resources. Then, the downstream task nodes affected by the directly associated task nodes are determined based on the task dependencies.

[0145] For example, when a self-service payment device malfunctions, the payment task node that is bound to the self-service payment device and is currently executing is identified as the directly affected task node. If a subsequent medication pickup task node requires the execution result of the aforementioned payment task node as input, then the medication pickup task node is identified as the affected downstream task node. The central scheduling agent retains the execution results of completed preceding task nodes, sets the payment task node currently using the self-service payment device as paused, moves medication pickup task nodes that depend on the payment task node but have not yet started execution into the reordering queue, and maintains the original task status for other task nodes not affected by the abnormal service resources.

[0146] The emergency response agent formulates alternative solutions based on the abnormal events. For example, the payment task originally performed using self-service payment equipment is changed to be completed through a manual service window, and the central dispatch agent re-determines the executing entity for the corresponding task node.

[0147] When the reallocation involves switching the execution agent of an ongoing task node, the central scheduling agent freezes the task state and task context object of the task node and writes them to the blackboard database, causing the original execution agent to release the corresponding execution lock. The task context object includes a patient needs summary, a dialogue context pointer, a snapshot of the business processing progress, and the execution results of previous tasks. It may also include the escort service order identifier and the current location of the escort.

[0148] The takeover agent reads the task context object from the blackboard database based on the task context identifier, and continues to execute the reassigned task node after acquiring the execution lock. This ensures that the existing patient needs, business processing progress, and previous task execution results do not need to be regenerated, while also preventing the original execution agent and the takeover agent from modifying the task state of the same task node at the same time.

[0149] As a further explanation of the first embodiment of the present invention, when the negative emotion fusion value of multiple consecutive dialogue rounds reaches a preset takeover threshold, and the confidence level of the corresponding negative emotion event reaches a preset confidence threshold, when a communication interaction task node is currently being executed, the central scheduling agent schedules a communication interaction agent to take over the communication interaction task node; during the takeover process, the original escort continues to execute non-communication task nodes in the current escort service order; the communication interaction agent continues to execute the communication interaction task node based on the corresponding task context; when the takeover task is completed and no negative emotion event is detected again for a consecutive preset time, the takeover state is lifted and the escort's communication responsibilities are restored.

[0150] In this embodiment, the quality supervision intelligent agent continuously generates service attitude evaluation results based on audio and text data during the escort service process, and feeds the service attitude evaluation results back to the central dispatch intelligent agent.

[0151] When a negative feature appears only in a single dialogue round, or when the confidence level of the current evaluation result is insufficient to support changing the task execution subject, the central scheduling agent sends a service prompt to the companion terminal without changing the task allocation of the current task node.

[0152] When negative emotional events appear consecutively in multiple dialogue rounds, and the negative emotion fusion value reaches a preset takeover threshold and the confidence level of the corresponding negative emotional event reaches a preset confidence threshold, the central scheduling agent further reads the current task status in the shared task graph.

[0153] If a patient consultation response, medical order relay, or other communication and interaction task node is currently being executed, the central scheduling agent will schedule the communication and interaction agent to take over the corresponding communication and interaction task node; if the currently executed task node is not a communication and interaction task node, the executing entity of the non-communication task node will not be directly changed due to the negative emotional event.

[0154] When taking over a communication and interaction task node, the central scheduling agent freezes the task status and task context object corresponding to the communication and interaction task node, and causes the original executor to release the corresponding execution lock. The communication and interaction agent reads the patient's needs summary, dialogue context, business processing progress and previous task execution results according to the task context identifier, and after obtaining the execution lock, continues to execute the patient consultation response or medical order relay task.

[0155] During the period when the communication and interaction intelligent agent takes over the communication task, the original escort continues to perform queuing, accompaniment, route guidance and other non-communication tasks, so that the switching of the execution subject of the communication task does not affect the continued execution of other task nodes in the current escort service order.

[0156] After the takeover task of the communication and interaction task node is completed, the quality supervision intelligent agent continues to evaluate the service quality of the subsequent communication process; when no negative emotional events are detected again for a preset period of time, the central dispatch intelligent agent releases the takeover state, restores the original communication responsibilities of the escort, and records the takeover, release of takeover and the corresponding task status changes to the service log of the escort service order.

[0157] S50: After the service is completed, each evaluation value is weighted according to the preset evaluation weight corresponding to each service quality evaluation indicator to obtain the service quality evaluation result. Each evaluation value is compared with the corresponding evaluation threshold to determine the unqualified service quality evaluation indicators of the corresponding escort. The unqualified service quality evaluation indicators are used to form the subsequent task matching parameters of the corresponding escort. The subsequent task matching parameters are fed back to the central scheduling agent to adjust the matching priority of the corresponding escort's subsequent orders.

[0158] Specifically, let the number of service quality evaluation indicators participating in the current service quality evaluation be . , No. The evaluation value of each service quality evaluation indicator is , No. The preset evaluation weights for each service quality evaluation indicator are: The service quality evaluation result is ,but: ; in, This indicates the service quality evaluation results; This indicates the number of service quality evaluation indicators participating in the current service quality evaluation; Indicates the serial number of the service quality evaluation indicator; Indicates the first The evaluation value of each service quality evaluation indicator; Indicates the first Each service quality evaluation indicator corresponds to a preset evaluation weight.

[0159] In one implementation, the preset evaluation weights of the twelve core service quality evaluation indicators are set according to the weights of the corresponding evaluation dimensions in the outpatient escort service quality evaluation indicator system, and can be adjusted according to the service quality evaluation needs of different patient groups.

[0160] For example, for routine elderly patients, the process familiarity evaluation index and task speed evaluation index adopt the corresponding standard evaluation weights.

[0161] For pregnant women, given their special needs for prenatal and postpartum recovery and emotional support, we should increase the evaluation weights of the respect and care evaluation indicators as well as the problem and difficulty-solving evaluation indicators, and adjust the retrieval scope of the corresponding medical knowledge graph.

[0162] The weighting of evaluations for different patient groups only changes the relative contribution of different service quality evaluation indicators to the service quality evaluation results, without changing the data source or the calculation method of the evaluation values ​​for the corresponding evaluation indicators.

[0163] After the service is completed, the report generates an intelligent agent that summarizes the algorithmic quantitative scores of twelve core service quality evaluation indicators, and combines them with the hospital's historical big data baseline and the final evaluation text entered by the patient through a mobile terminal to conduct a final comprehensive evaluation, forming a structured service quality report.

[0164] The structured service quality report records the service quality evaluation results, main deduction indicators, time of occurrence of negative emotional events, corresponding business links, and communication overlap of this escort service.

[0165] For example, the structured service quality report recorded that the overall score for this escort was 82 points, with the main deduction being service attitude. It also recorded that negative emotions were detected in the service personnel during the payment process at 10:35, with a confidence level of 91% and a communication overlap rate of over 25%. This allows the service quality evaluation results to be traced back to the specific service time and specific business process.

[0166] While generating the service quality evaluation results, the evaluation values ​​of each service quality evaluation indicator are compared with the corresponding evaluation thresholds, and the non-compliant service quality evaluation indicators of the corresponding escorts are determined based on the comparison results.

[0167] When the service quality evaluation result or the corresponding evaluation indicator is lower than the corresponding service standard or the evaluation requirements set by the escorting institution, the system can trigger the corresponding quality alarm, write the corresponding evaluation result into the escort's evaluation record, and determine the service capability dimensions that the escort needs to focus on improving based on the substandard service quality evaluation indicators, so as to form training priorities.

[0168] In this embodiment, step S50 further includes: Step S51: Summarize the evaluation values ​​of each service quality evaluation indicator in each of the corresponding escort services according to the escort identifier, and form a historical trend profile of each service quality evaluation indicator of the corresponding escort. Step S52: Determine the personalized training focus and retraining course suggestions for the corresponding escort based on the historical trend profile, and adjust the subsequent task matching parameters based on the historical trend profile.

[0169] In step S51, the evaluation values ​​of each service quality evaluation index generated for each escort service are associated with and saved with the corresponding escort ID.

[0170] For each individual escort, the historical evaluation results of twelve core service quality evaluation indicators are summarized based on the escort's identification, forming a historical trend of each service quality evaluation indicator as escort service orders change, thereby creating a historical trend profile of the corresponding escort.

[0171] The historical trend profile is used to represent the historical evaluation of the corresponding escort on different service quality evaluation indicators, so that managers can determine the changes in the service capabilities of the corresponding escort from the dimensions of each evaluation indicator, rather than making a judgment based solely on the comprehensive evaluation results of a single escort service.

[0172] In step S52, based on the service quality evaluation indicators in the historical trend profile that are below the corresponding evaluation baseline, personalized training priorities and retraining course suggestions are generated for the corresponding escorts.

[0173] In one implementation, when the evaluation result of the corresponding escort is lower than the corresponding service standard or the qualification requirements set by the escort institution, training or retraining requirements are generated, and the corresponding evaluation result is written into the escort's professional evaluation record.

[0174] The historical trend profile is fed back to the central scheduling agent, which serves as the basis for task matching when allocating tasks in the future, and is used to adjust the matching priority of the corresponding escort's subsequent orders, thereby realizing the adjustment of the matching parameters of the subsequent tasks.

[0175] The subsequent task matching parameters include matching priority adjustment parameters corresponding to different task types. The central scheduling agent adjusts the matching priority of the corresponding task type based on the service capability dimension corresponding to the service quality evaluation indicators that have not met the standards. For example, when the process familiarity evaluation indicator fails to meet the standards, the matching priority of the corresponding escort in performing cross-departmental guidance tasks is reduced; when the service attitude evaluation indicator fails to meet the standards, the matching priority of the corresponding escort in performing communication and interaction tasks is reduced. The higher the degree of failure, the greater the reduction in the corresponding matching priority.

[0176] The structured service quality report can also be sent to the escort service management backend and the corresponding industry regulatory data terminal.

[0177] The report-generating agent can also perform agency-level quality analysis based on the service quality evaluation results of multiple escort service orders.

[0178] Specifically, the deduction information corresponding to each service quality evaluation indicator is aggregated and analyzed according to the department visited, service link, and patient group to identify the recurring low evaluation indicators and their corresponding service links, so as to identify the systemic weaknesses of the escort service institution.

[0179] For example, identifying frequent service attitude problems during the payment process or abnormally long service times during the medication pickup process.

[0180] For escort services, the statistical results of each service quality evaluation indicator are compared with the regional industry baseline. The quality improvement priority is determined based on the differences with the regional industry baseline, and periodic quality control reports are generated.

[0181] For third-party review agencies, the system saves business timestamps, audio features, location trajectories, event tags, service quality evaluation results, and intervention records according to the escort service order, so as to form a service process evidence chain corresponding to a single escort service order.

[0182] After obtaining the corresponding authority, the third-party review agency will review the service quality evaluation results based on the evidence chain of the service process, and form a list of quality improvement suggestions based on the review results and low evaluation indicators.

[0183] The institutional-level quality analysis results, historical trend profiles of escorts, periodic quality control reports, and quality improvement suggestion lists serve as inputs for subsequent evaluation, training record generation, and service scheduling strategy updates, thereby forming a continuous quality improvement process of evaluation, feedback, improvement, and re-evaluation.

[0184] In one specific application scenario, after a patient books a medical escort service through a mobile terminal, the central scheduling agent breaks down the escort service order into multiple task nodes with time-dependent relationships based on the medical service items, and completes task allocation based on the capability declarations of candidate agents and the current load.

[0185] Once the escort arrives at the hospital and begins the escort service, the escort staff continuously collects audio data and location trajectory during the escort service process using a smart work badge equipped with a microphone and indoor positioning module.

[0186] The audio data undergoes noise suppression, speech endpoint detection, and blind source voiceprint separation to obtain the companion's speech segment, which is then transcribed into text data by an automatic speech recognition engine.

[0187] For example, while queuing to pay, a patient might ask their companion, "Why is it taking so long?" When the companion's response does not show obvious negative semantics in the corresponding text data, but the tone of voice is significantly raised, the speech rate is faster, and the energy fluctuations are obvious, the speech emotion recognition network generates a high negative emotion value in the speech.

[0188] The negative sentiment values ​​of text and voice are dynamically weighted and fused based on the current audio signal-to-noise ratio and text transcription confidence to form a negative sentiment fusion value, and the evaluation value of the service attitude evaluation index is determined based on the negative sentiment fusion value.

[0189] In a specific operating state, when the negative emotion fusion value reaches the preset takeover threshold, and the confidence level of the corresponding negative emotion event reaches 91% and occurs for two consecutive dialogue rounds, the event timestamp, indicator dimension, and confidence level are used to form a quality anomaly state record, which is then sent to the central scheduling agent.

[0190] After the central dispatching agent determines, based on the shared task graph, that the companion is currently performing a communication and interaction task node of patient consultation and response, the dispatching communication and interaction agent takes over the task node.

[0191] The communication and interaction agent reads the current dialogue context, patient needs, and business processing status based on the task context identifier, and continues to answer questions via voice or text through the patient's terminal; the original escort continues to perform non-communication tasks such as queuing, guiding, or accompanying.

[0192] Once the Q&A session is completed and the quality monitoring AI agent does not detect any negative emotional events for 5 consecutive minutes, the central dispatch AI agent will release the takeover status, restore the original communication responsibilities of the escort, and record the intervention process in the service log of the escort service order.

[0193] When a self-service payment device malfunctions during the patient's visit, the emergency response AI identifies abnormal triggering information such as "the machine is broken," generates an alternative solution to pay at a manual service window, and requests the central dispatch AI to readjust the corresponding task nodes. The communication AI simultaneously explains the reason for the service process change to the patient.

[0194] After the service is completed, the report generates an intelligent agent that summarizes the evaluation values ​​of each service quality evaluation indicator, negative emotional events, task status, intervention records, and the patient's final evaluation. It forms a service quality evaluation result according to the preset evaluation weights corresponding to each service quality evaluation indicator, and compares each evaluation value with the corresponding evaluation threshold to determine the service quality evaluation indicators that the corresponding escort did not meet. Based on the service quality evaluation indicators that did not meet the standards, the training focus is determined, and the historical trend profile of the corresponding escort is updated. The historical trend profile is used as the basis for subsequent task matching.

[0195] The second embodiment of the present invention proposes an outpatient escort service and quality assessment system based on multi-agent collaboration, including a task orchestration module, a data acquisition module, an indicator evaluation module, a feedback scheduling module, and an evaluation management module; The task orchestration module is used to split the escort service order into task nodes with time-series dependencies. The central scheduling agent allocates the task nodes according to the capability declaration of the candidate agents and the current load, and records the task dependencies and task status through a shared task graph. The data acquisition module is used to collect data on the escort service process, which includes audio data, location trajectory, business timestamps, and text data transcribed from the audio data. The indicator evaluation module is used to calculate the evaluation values ​​of multiple preset service quality evaluation indicators using the data from the escort service process. The feedback scheduling module is used to feed back each of the evaluation values ​​to the central scheduling agent, intervene in the task execution according to each of the evaluation values ​​and the current task status, and when the intervention causes the execution agent of the task node to switch, the takeover agent continues to execute the task node based on the corresponding task context. The evaluation management module is used to weight each evaluation value according to the preset evaluation weight corresponding to each service quality evaluation indicator after the service is completed, to obtain the service quality evaluation result, compare each evaluation value with the corresponding evaluation threshold, determine the service quality evaluation indicators that do not meet the standards for the corresponding escort, form the subsequent task matching parameters for the corresponding escort based on the substandard service quality evaluation indicators, and feed the subsequent task matching parameters back to the central scheduling agent to adjust the matching priority of the corresponding escort's subsequent orders.

[0196] Specifically, the outpatient escort service and quality assessment system based on multi-agent collaboration in this embodiment adopts a five-layer architecture consisting of a terminal perception layer, a data middleware layer, a multi-agent collaboration layer, a rules and knowledge layer, and an application display layer.

[0197] The terminal sensing layer includes a smart terminal carried by the caregiver, the patient's mobile device, and an in-hospital positioning device. The smart terminal can be a smartphone or a smart name tag integrating a microphone and indoor positioning module, or it can be a smartwatch, smart earphone, or other wearable device with corresponding audio and location acquisition functions.

[0198] The terminal perception layer is used, under the condition of obtaining the appropriate authorization, to collect environmental audio data, GPS or Bluetooth / Wi-Fi indoor positioning coordinate streams, and business node operation timestamps during the escort service process, and to receive the patient's final text or voice evaluation. Different terminals communicate with the server through an encrypted communication channel.

[0199] The data platform layer includes a noise suppression module, a voice endpoint detection module, a blind source voiceprint separation module, an automatic speech recognition engine for the medical vertical field, and a trajectory data cleaning module, and maintains an event bus for each intelligent agent to publish and subscribe to messages.

[0200] The multi-agent collaboration layer includes a central dispatching agent and six specialized agents, namely, a triage and planning agent, a communication and interaction agent, a business processing agent, an emergency response agent, a quality supervision agent, and a report generation agent.

[0201] The central scheduling agent is used to parse and break down the escort service order into task nodes with time-series dependencies, perform task allocation based on the capability declarations of each candidate agent and the current load, and arbitrate service resource conflicts.

[0202] The intelligent patient guidance system is used to orchestrate the patient flow, plan routes within the hospital, and provide navigation guidance, while also obtaining real-time information on department locations, queue lengths, and appointment availability.

[0203] The communication and interaction intelligent agent performs patient consultation responses, interprets and relays medical orders, and proactively reminds patients of precautions for medical visits based on a large language model and medical knowledge graph. When a communication and interaction task node takes over, it continues to execute the task node based on the corresponding task context.

[0204] The intelligent agent for business processing is used to execute business nodes such as registration, payment, medication collection, and appointment for examination, and generates corresponding business status receipts.

[0205] The emergency response intelligent agent is used to monitor abnormal events in real time, formulate alternative plans based on the abnormal events, and request the central dispatch intelligent agent to readjust the tasks.

[0206] The quality supervision intelligent agent operates a text semantic analysis network and a voice emotion recognition network in parallel, and evaluates the escort service process in real time based on twelve core service quality evaluation indicators.

[0207] The report generation agent is used to summarize the evaluation values ​​and event tags of each service quality evaluation indicator after the service is completed, and generate a structured service quality report and quality improvement suggestions.

[0208] Each agent exchanges messages via an event bus using publish and subscribe methods, and jointly reads and writes to the blackboard database.

[0209] The rules and knowledge layer is used to store the quantitative rules for the evaluation indicators of outpatient escort service quality, the preset evaluation weights, evaluation thresholds, preset intervention conditions, medical vertical domain knowledge graphs, and the baseline of historical data of hospital treatment process.

[0210] The application presentation layer is used to provide patients with visual service quality reports, and to provide escort agencies and regulatory agencies with access-controlled data dashboards, quality control reports, and interfaces for scheduling escort skills training.

[0211] The task orchestration module is jointly implemented by a central scheduling agent, an event bus, and a blackboard database.

[0212] The central scheduling agent allocates task nodes based on the capability declarations of candidate agents and the current load, and the shared task graph records the task dependencies between task nodes and the task status of each task node.

[0213] The shared task graph can also record the binding relationships between task nodes and service resources. When the emergency response agent identifies an anomaly in a service resource, the central scheduling agent determines the affected task nodes based on the abnormal service resource, the binding relationship between the task node and the service resource, and the task dependency relationship, and then reallocates the corresponding task nodes.

[0214] The data acquisition module is jointly implemented by the terminal perception layer and the data platform layer. The terminal perception layer acquires audio data, location trajectory, and business timestamps; the data platform layer performs noise suppression, voice endpoint detection, voiceprint separation, and automatic speech recognition on the audio data to form text data and corresponding text transcription confidence scores, and performs data cleaning on the location trajectory.

[0215] The indicator evaluation module is jointly implemented by the text semantic processing network, the speech acoustic processing network, and the quantitative processing logic corresponding to different service quality evaluation indicators in the quality supervision intelligent agent.

[0216] The text semantic processing network uses a multi-head attention mechanism to extract text semantic features and form text negative sentiment values; the speech acoustic processing network uses a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism to form speech negative sentiment values; at the same time, it forms evaluation values ​​for other service quality evaluation indicators based on location trajectory, business timestamps, medical knowledge data, patient feedback, and business node execution results.

[0217] For the service attitude evaluation index, the speech fusion weight and text fusion weight are adjusted according to the audio signal-to-noise ratio corresponding to the audio data and the text transcription confidence corresponding to the text data. The negative emotion value of the speech and the negative emotion value of the text are weighted according to the adjusted speech fusion weight and text fusion weight to obtain the negative emotion fusion value. The evaluation value of the service attitude evaluation index is determined according to the negative emotion fusion value.

[0218] The feedback scheduling module is jointly implemented by the quality supervision intelligent agent, the central scheduling intelligent agent, the communication and interaction intelligent agent, the rule and knowledge layer, and the blackboard database.

[0219] The quality supervision intelligent agent sends each evaluation value to the central dispatch intelligent agent through the event bus; the central dispatch intelligent agent executes service prompts, takes over communication and interaction task nodes, or reconstructs tasks based on each evaluation value and the current task status.

[0220] When the intervention causes a switch of the execution agent of the task node, the central scheduling agent freezes the task state and task context object of the corresponding task node and causes the original execution agent to release the execution lock; the takeover agent reads the corresponding task context object according to the task context identifier, and continues to execute the corresponding task node after acquiring the execution lock.

[0221] The evaluation management module weights each evaluation value according to the preset evaluation weights corresponding to each service quality evaluation indicator to form a service quality evaluation result. It then compares each evaluation value with the corresponding evaluation threshold to determine the service quality evaluation indicators that the corresponding escort does not meet.

[0222] The evaluation management module determines the training focus based on the substandard service quality evaluation indicators and the historical evaluation results of the corresponding escorts, and summarizes the evaluation values ​​of each service quality evaluation indicator in each escort service according to the escort identifier, forming a historical trend profile of each service quality evaluation indicator.

[0223] Based on the historical trend profile, the personalized training focus and retraining course suggestions for the corresponding escort are determined, and the historical trend profile is fed back to the central scheduling agent as the basis for task matching when the central scheduling agent allocates tasks in the future, and is used to adjust the matching priority of the corresponding escort's subsequent orders.

[0224] In this embodiment, each specialized intelligent agent can be implemented using a tool-calling method based on a large language model, a software method based on a rule engine and a finite state machine, or a combination of both methods.

[0225] In addition to using the contract network protocol, task allocation can also employ auction mechanisms, centralized arbitration, or allocation methods based on reinforcement learning strategy networks.

[0226] The data platform layer and the multi-agent collaboration layer can be deployed on public cloud servers, hospital private servers, or edge computing nodes.

[0227] When deployed in a hospital-owned private facility, audio data, text data, and location trajectories can be processed on the hospital's internal servers. When using edge computing, the terminal carried by the patient escort completes some of the audio data preprocessing before sending the processed data to the subsequent processing module.

[0228] In addition to smart work badges, smartwatches, smart headphones, and other wearable devices can also be used to collect audio data and location trajectories.

[0229] Through the aforementioned system, when a department suspends outpatient services, changes appointment slots, or service equipment malfunctions, the central dispatching agent can reallocate tasks to affected task nodes based on a shared task graph. The quality supervision agent can utilize audio data, text data, location trajectories, and business timestamps to generate evaluation values ​​for different service quality indicators. For service attitude evaluation, it can dynamically adjust the fusion weights of the voice and text modalities based on the audio signal-to-noise ratio and text transcription confidence. Furthermore, it can feed back each evaluation value to the central dispatching agent in real time, intervene in task execution based on the evaluation value and the current task status, maintain continuous task execution through task context during agent switching, and further generate service quality evaluation results after the service is completed. This identifies the corresponding escort's substandard service quality evaluation indicators, forms corresponding training priorities, and adjusts the subsequent task matching parameters for the corresponding escort.

[0230] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0231] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0232] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for outpatient escort service and quality assessment based on multi-agent collaboration, characterized in that, include: The order for accompanying medical services is broken down into task nodes with time-dependent relationships. The central scheduling agent allocates the task nodes according to the capability declarations of candidate agents and the current load, and records the task dependencies and task status through a shared task graph. Collect data on the process of accompanying medical visits, including audio data, location trajectory, business timestamps, and text data transcribed from the audio data; The evaluation values ​​of multiple preset service quality evaluation indicators are calculated using the data from the accompanying medical service process. Each of the aforementioned evaluation values ​​is fed back to the central scheduling agent, which intervenes in the task execution based on each of the aforementioned evaluation values ​​and the current task status. When the intervention causes the execution agent of the task node to switch, the takeover agent continues to execute the task node based on the corresponding task context. After the service is completed, each evaluation value is weighted according to the preset evaluation weights corresponding to each service quality evaluation indicator to obtain the service quality evaluation result. Each evaluation value is compared with the corresponding evaluation threshold to determine the service quality evaluation indicators that do not meet the standards for the corresponding escort. The subsequent task matching parameters for the corresponding escort are formed based on the substandard service quality evaluation indicators and fed back to the central scheduling agent to adjust the matching priority of the corresponding escort's subsequent orders.

2. The outpatient escort service and quality assessment method based on multi-agent collaboration according to claim 1, characterized in that, The allocation of the task nodes includes: The central scheduling agent issues task tenders to candidate agents for the task nodes to be assigned. The candidate agents generate bidding data based on their own capability declarations and current load; The central scheduling agent determines the execution agent based on the bidding data and assigns the task nodes to be allocated to the execution agent.

3. The outpatient escort service and quality assessment method based on multi-agent collaboration according to claim 1, characterized in that, The shared task graph also records the binding relationship between task nodes and service resources; The emergency response agent identifies abnormal events and determines abnormal service resources; The central scheduling agent determines the affected task nodes based on the abnormal service resources, the binding relationships, and the task dependencies. It retains the execution results of completed task nodes, sets the executing task nodes bound to the abnormal service resources as paused and reassigns them, and moves the pending task nodes that depend on the affected task nodes into the reordering queue. The task nodes in the rearranged queue are reallocated according to the task dependencies.

4. The outpatient escort service and quality assessment method based on multi-agent collaboration according to claim 1, characterized in that, Regarding the service attitude evaluation indicator among the multiple service quality evaluation indicators: The negative sentiment values ​​for the voice and text are determined based on the audio data and the text data, respectively. The speech fusion weights and text fusion weights are adjusted based on the audio signal-to-noise ratio corresponding to the audio data and the text transcription confidence level corresponding to the text data. Based on the adjusted voice fusion weights and text fusion weights, the negative emotion values ​​of the voice and the text are weighted to obtain a negative emotion fusion value, and the evaluation value of the service attitude evaluation index is determined based on the negative emotion fusion value.

5. The outpatient escort service and quality assessment method based on multi-agent collaboration according to claim 4, characterized in that, The intervention in task execution includes: When the evaluation value of the service quality evaluation index based on the model is lower than the corresponding evaluation threshold, and the corresponding judgment confidence is lower than the preset confidence threshold, or when a negative feature appears in a single round of dialogue, a service prompt is sent to the companion's terminal. When the negative emotion fusion value of multiple consecutive dialogue rounds reaches the preset takeover threshold, and the number of dialogue rounds reaches the preset round threshold, when the current communication interaction task node is being executed, the central scheduling agent will schedule the communication interaction agent to take over the communication interaction task node. When the number of times the communication and interaction task nodes in the same escort service order are taken over reaches a preset threshold, or when the preset key service quality evaluation indicators meet the preset task reconstruction conditions, the central scheduling agent will reconstruct the task.

6. The outpatient escort service and quality assessment method based on multi-agent collaboration according to claim 1, characterized in that, The multiple service quality evaluation indicators also include process familiarity evaluation indicators and response timeliness evaluation indicators; The trajectory turnaround rate is calculated based on the planned path generated by the triage planning intelligent agent and the location trajectory, and the evaluation value of the process familiarity evaluation index is determined based on the trajectory turnaround rate. The audio data is subjected to voice endpoint detection, the average time difference between the end of the patient's question and the beginning of the companion's response is calculated, and the evaluation value of the response timeliness evaluation index is determined based on the comparison result of the average time difference and the preset response threshold.

7. The outpatient escort service and quality assessment method based on multi-agent collaboration according to claim 4, characterized in that: The text data is encoded into a text feature matrix, and the text semantic features are calculated through multi-head attention. The text semantic features are then input into a text sentiment classifier to obtain the negative sentiment value of the text. The accompanying person's voice segments are separated from the audio data, and the Mel frequency cepstral coefficients, fundamental frequency, energy profile, and zero-crossing rate are extracted to form an acoustic feature matrix. The acoustic feature matrix is ​​convolved and pooled to obtain local frequency domain features. The local frequency domain features are input into a bidirectional long short-term memory network, and the obtained temporal coding results are attention-weighted to obtain the negative emotion value of the voice.

8. The outpatient escort service and quality assessment method based on multi-agent collaboration according to claim 1, characterized in that: The evaluation values ​​of each service quality evaluation indicator of the corresponding escort are summarized according to the escort's identification in each escort service to form a historical trend profile of each service quality evaluation indicator of the corresponding escort. Based on the historical trend profile, the personalized training focus and retraining course suggestions for the corresponding escorts are determined, and the subsequent task matching parameters are adjusted according to the historical trend profile.

9. The outpatient escort service and quality assessment method based on multi-agent collaboration according to claim 1, characterized in that, The takeover agent continues to execute the task node based on the corresponding task context, including: The central scheduling agent maintains a task context object for the task node and configures a task context identifier for the task context object. The task context object includes a patient needs summary, a dialogue context pointer, a business processing progress snapshot, and the execution result of the preceding task. Before switching agents, freeze the task state of the task node and the task context object, and release the execution lock held by the original agent. The takeover agent reads the task context object based on the task context identifier, acquires the execution lock, and then continues to execute the task node.

10. A system for outpatient escort service and quality assessment based on multi-agent collaboration, characterized in that, It includes a task orchestration module, a data acquisition module, an indicator evaluation module, a feedback scheduling module, and an evaluation management module; The task orchestration module is used to split the escort service order into task nodes with time-series dependencies. The central scheduling agent allocates the task nodes according to the capability declaration of the candidate agents and the current load, and records the task dependencies and task status through a shared task graph. The data acquisition module is used to collect data on the escort service process, which includes audio data, location trajectory, business timestamps, and text data transcribed from the audio data. The indicator evaluation module is used to calculate the evaluation values ​​of multiple preset service quality evaluation indicators using the data from the escort service process. The feedback scheduling module is used to feed back each of the evaluation values ​​to the central scheduling agent, intervene in the task execution according to each of the evaluation values ​​and the current task status, and when the intervention causes the execution agent of the task node to switch, the takeover agent continues to execute the task node based on the corresponding task context. The evaluation management module is used to weight each evaluation value according to the preset evaluation weight corresponding to each service quality evaluation indicator after the service is completed, to obtain the service quality evaluation result, compare each evaluation value with the corresponding evaluation threshold, determine the service quality evaluation indicators that do not meet the standards for the corresponding escort, form the subsequent task matching parameters for the corresponding escort based on the substandard service quality evaluation indicators, and feed the subsequent task matching parameters back to the central scheduling agent to adjust the matching priority of the corresponding escort's subsequent orders.

Images