Online outpatient service intelligent reservation method based on deep learning

Abstract

The invention belongs to the field of outpatient service intelligent reservation, and particularly relates to an online outpatient service intelligent reservation method based on deep learning, and the method comprises the steps: collecting user data, and carrying out the standardization processing; predicting the doctor seeing time of the user; judging the doctor seeing probability of the user and calculating a doctor seeing congestion risk; constructing link-level constraint modeling and initial arrangement; setting rolling rearrangement of the doctor-seeing trigger event; carrying out prediction model joint training and iterative optimization on the premise of protecting the privacy of the user; according to the method, on the basis of the prediction of the uncertainty of the ready time of the inspection report and the joint evaluation of the diagnosis arrival probability and the congestion risk, a rolling rearrangement mechanism driven by the link-level constraint can be used, so that the same-day completion rate of a patient is remarkably improved on the premise that resources are not expanded, the invalid re-diagnosis rate is reduced, and the effect of relieving the peak congestion on site is achieved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent outpatient appointment, specifically involving an online intelligent outpatient appointment method based on deep learning. Background Technology

[0002] Chinese patent application number CN202211480178.1 discloses an intelligent online appointment management system. This system includes a login module, an outpatient information display module, an online appointment module, a data analysis module, and a customer service module. The online appointment module includes an intelligent triage unit, an appointment allocation unit, and an appointment priority unit. The intelligent triage unit recommends corresponding departments to patients, and the appointment allocation unit allocates appointment slots to patients. The system features intelligent triage recommending corresponding departments, and the appointment allocation unit dynamically allocating slots based on department capacity, the number of appointments already booked, and the importance of the patient's condition to ensure rational resource allocation. The appointment priority unit determines the order of visits based on factors such as the importance of the patient's condition and appointment fulfillment rate, prioritizing urgent and important patients. Furthermore, by rating patients, it encourages timely appointment fulfillment, thus improving the overall online appointment experience for patients.

[0003] While this invention, in the field of appointment management technology, considers the patient's condition when making online appointments, allowing patients with urgent conditions to be prioritized for consultation without queuing for examinations, and also addresses the issue of low appointment rates due to the increasing number of online appointments, which delays doctors' time and affects other patients' consultations, it fails to consider the fixed cycle and outcome-dependent nature of follow-up visits for some chronic disease patients. For example, patients may need to undergo examinations first, followed by follow-up visits and prescriptions. It also fails to address the inconsistency between appointment slots and testing times between higher-level specialist clinics and community clinics, the issue of scheduling follow-up visits before patients' test reports are ready, and the possibility of patients being unable to attend due to holidays or weather conditions without online correction. This leads to both ineffective consultations and congestion at clinics, resulting in repeated rescheduling and increased communication costs. Therefore, a deep learning-based intelligent online outpatient appointment method is urgently needed to solve these problems. Summary of the Invention

[0004] This invention provides a deep learning-based intelligent online outpatient appointment method. The aim is to address issues such as: failure to consider the fixed cycles and outcome-dependent nature of follow-up visits for some chronic disease patients; failure to resolve inconsistencies in appointment availability and testing times between higher-level specialists and community clinics; failure to schedule follow-up visits before patients' test reports are ready; and failure to account for patients' inability to visit due to holidays and weather conditions, coupled with a lack of online correction. These issues lead to both ineffective patient visits and overcrowding, resulting in repeated rescheduling and increased communication costs.

[0005] The technical solution adopted by this invention to solve the above problems is: an online outpatient intelligent appointment method based on deep learning, comprising: accessing a basic data set; and establishing cross-system field mapping relationships based on the basic data set;

[0006] By performing a time-series model of the patient's testing process, a vector that can be used to calculate the duration of each testing stage is obtained; the patient's testing process is modeled in a time sequence.

[0007] By using quantiles to predict the output of the test report, the upper and lower bound windows and median estimate of the report completion time are obtained.

[0008] By constructing a comprehensive risk scoring formula to calculate the window expiration probability and the abnormal delay probability, risk labels that can be directly consumed by the orchestration layer are obtained.

[0009] A monotonic constraint scoring formula is constructed based on the outputs of the patient's no-show probability, late consultation probability, and consultation congestion probability.

[0010] Construct link-level constraint modeling and initial orchestration;

[0011] Set up a rolling reordering of medical visit trigger events; conduct joint training and iterative optimization of prediction models while protecting user privacy.

[0012] As a preferred implementation, the specific steps for establishing cross-system field mapping relationships based on the basic data set are as follows: by irreversibly hashing the patient's personal identifier and de-sensitizing the patient's appointment number, a controlled data availability boundary is obtained by retaining only the association keys available during off-peak hours; by uniformly dividing and recording the doctor's registration source, the time period of the testing equipment, and the pharmacy window into fixed-granularity time slices, a time resource view that can be computed across platforms is obtained; by setting ordered relationships, off-peak interval thresholds, mutual exclusion constraints, and marking the positions of the patient's consultation process, a set of rules for determining the patient's consultation link is obtained.

[0013] As a preferred embodiment, the specific steps for predicting the output of the test item report using quantiles are as follows:

[0014] By using monotonic calibration of quantile coverage and integrating abnormal time periods in the form of small-scale multi-model, a ready time window consistent with the confidence level and actual coverage is obtained; by setting listening events for changes in the stage status of the inspection process and updating the remaining time according to the changes in the stage of the inspection process, a ready time window that narrows as the inspection process progresses is obtained.

[0015] As a preferred implementation, the specific steps for calculating the window expiration probability and the abnormal delay probability by constructing a comprehensive risk scoring formula are as follows:

[0016] By constructing a comprehensive risk scoring formula to calculate the window expiration probability and the abnormal delay probability, risk labels that can be directly consumed by the orchestration layer are obtained. The comprehensive risk score calculation formula is as follows:

[0017] ,

[0018] A represents the overall risk score. The weighting factor for reporting unprepared risks is Readay, which represents the risk of reporting unpreparedness. Here, represents the weighting coefficient for the probability of abnormal delay, and Delay represents the probability of abnormal delay. B represents the weighting coefficient for time-period congestion risk, and B represents the time-period congestion risk. is the weighting factor for buffer adequacy, where Buffer is the buffer adequacy.

[0019] As a preferred embodiment, the specific steps for constructing a monotonic constraint scoring formula based on the outputs of the patient's no-show probability, late arrival probability, and congestion probability are as follows:

[0020] Based on the outputs of patient no-show probability, late consultation probability, and consultation congestion probability, a monotonic constraint scoring formula is constructed. While maintaining the original medical resources, priority is given to allocating low-risk, feasible time slots. The monotonic constraint scoring formula is as follows:

[0021] ,

[0022] C represents the appointment priority of the candidate time slot T, and T is the time parameter of the candidate time slot. For normalization function, For biased items, As a weight for the reliability of patient arrival, Let T be the probability function of no patient arrival during the candidate time period T. Weighting for on-time arrival, Let T be the probability function of being late for the candidate time period T. Weighting for low congestion preference, Let T be the congestion risk function for the candidate time period T. For buffer sufficiency weights, Let T be the buffer sufficiency function for the candidate time period T. Weighting for inadequate preparation penalties The report-in-readiness risk function is the candidate time period T.

[0023] As a preferred implementation, the specific steps for constructing link-level constraint modeling and initial orchestration are as follows:

[0024] By abstracting the patient's testing process into time nodes and setting hard and soft constraints, a computable feasible domain for the link is obtained. By setting hierarchical objectives, the coverage of feasible links is first increased, and then the total consultation time and peak congestion rate are reduced, thereby reducing the rearrangement radius and obtaining an interpretable optimization order.

[0025] As a preferred embodiment, the specific steps for setting the rolling reorder of the medical visit trigger event are as follows:

[0026] By subscribing to report-ready events, comprehensive risk scores, and monotonic constraint scores, and identifying affected link subgraphs through the event engine, the target patient and their adjacent time slice set are obtained. Lock-free incremental solving is performed on the affected subgraph and the candidate time slice set, and the event engine is scheduled to generate alternative paths to obtain a new schedule, thereby reducing rescheduling and minimizing secondary queuing. By setting upper limits on the rearrangement radius and notification window, the execution module controls the granularity of message delivery, resulting in a controllable online correction process.

[0027] As a preferred embodiment, the specific steps for setting the rolling reordering of medical visit trigger events further include:

[0028] By writing conflict proofs to the replacement time slices after rearrangement and the reason for the change, a traceable rearrangement loop is obtained, which can improve the verifiability of authorization. By providing a backup for the candidate strategy triggered by high-risk tags, the strategy unit allocates spare time slices to the triggered candidate strategy, resulting in a fault-tolerant chain from failure to rollback to recovery.

[0029] As a preferred implementation, the specific steps for jointly training and iteratively optimizing the prediction model while protecting user privacy are as follows:

[0030] By training sub-models on local data for both community and specialist sides, extracting intermediate representations from the trained sub-models, and then using a federated coordinator to securely aggregate the gradients of the extracted representations and sub-model parameters, a global model is obtained without leaking the original data. By employing a secure aggregation protocol on the submitted update packages and adding configurable differential privacy noise after aggregation, a privacy-preserving global update vector is obtained. By performing knowledge distillation on the aggregated global model and generating lightweight inference sub-models that are periodically distributed to hospitals and community institutions, an efficient predictor can be deployed in environments with weak networks and low computing power.

[0031] As a preferred implementation, the specific steps for jointly training and iteratively optimizing the prediction model while protecting user privacy further include:

[0032] After each round of model distribution, coverage and quantile error are calculated in each institution using a reserved validation set, and aggregated metrics are fed back to obtain consistent performance evaluation results across institutions. By periodically feeding back the actual arrival time, report issuance time, and rescheduling results of patients' online procedures as supervision labels and conducting A / B trials in an isolated environment to verify the strategy, an empirically validated parameter adjustment scheme is obtained.

[0033] The beneficial effects of this invention are:

[0034] 1. By focusing on the uncertainty prediction of test report readiness time and combining the prediction results with the probability of arrival and congestion risk to drive link-level orchestration, the completion rate of same-day patient consultations can be improved and the rate of invalid follow-up visits and average waiting time can be significantly reduced, thereby reducing resource waste and improving the overall accessibility of medical services.

[0035] 2. By solving the link-level constraints of the time slice graph and combining the lock-free incremental local replacement of the affected subgraph to achieve rolling rearrangement, the peak congestion and queuing peaks are significantly alleviated and the resource utilization is significantly improved. It can smooth the treatment load and improve the system throughput without expanding the capacity.

[0036] 3. By employing secure aggregation federated training and knowledge distillation based on the aggregation teacher model to distribute lightweight inference sub-models and feed back online execution data for threshold adaptation, the platform can maintain high-precision prediction and deployable edge inference capabilities without exchanging raw data, while ensuring privacy compliance. This achieves a balance between prediction accuracy, cross-institutional generalization, and compliance auditability, thereby supporting large-scale promotion and robust implementation. Attached Figure Description

[0037] Figure 1 This is a flowchart of an online outpatient appointment method based on deep learning.

[0038] Figure 2This is a technical rendering of an online outpatient appointment system based on deep learning. Detailed Implementation

[0039] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0040] Example 1 Figure 1 The flowchart of an online outpatient intelligent appointment method based on deep learning includes: collecting user data and performing standardized processing; predicting users' appointment times; judging users' appointment probabilities and calculating the risk of congestion; constructing link-level constraint modeling and initial orchestration; step three, setting up rolling reordering of appointment trigger events; and conducting joint training and iterative optimization of the prediction model while protecting user privacy. The specific implementation steps are as follows:

[0041] Step 1: Collect user data and perform standardized processing; predict users' consultation time.

[0042] Specifically, the steps for collecting and standardizing user data are as follows: First, by accessing a basic data set, basic data covering testing, reporting, follow-up visits, and medication dispensing is obtained. Second, by implementing mandatory field validation on key fields of the basic data set, default values ​​are assigned to missing data segments, and abnormal time-slice data is identified, quantifiable data is obtained. Third, the basic data set is composed of data from community clinic schedules, higher-level specialist clinic schedules, testing progress status, pharmacy window opening hours, historical medical records, holidays, and weather calendars. Fourth, a cross-system field mapping relationship is established based on the basic data set, resulting in a unified and stable field naming standard, thereby eliminating discrepancies in terminology between community clinics and specialist hospitals. Data ambiguity caused by label differences; by irreversibly hashing patient personal identifiers and de-identifying patient appointment numbers, retaining only association keys available during off-peak hours, a controlled data usability boundary is obtained, thereby reducing the risk of patient privacy leaks and meeting compliance audit requirements; by uniformly segmenting and recording doctor registration sources, laboratory equipment time slots, and pharmacy windows into fixed-granularity time slices, a cross-platform computeable time resource view is obtained, providing a unified basic unit for subsequent constraint solving and rolling rearrangement; by setting ordered relationships, off-peak interval thresholds, mutual exclusion constraints, and marking positions for patient consultation processes, a patient consultation link determination rule set is obtained, enabling the handling of irreversible data before orchestration generation. Combining filtering is performed to reduce subsequent conflict rates; the underestimated interval threshold is the time required for the patient's test report after examination plus the minimum time required for the patient to move to the follow-up appointment point; the mutual exclusion constraint is that within the same time slice, patients cannot be assigned to resource time periods with overlapping time ranges; duplicate patient consultation records are eliminated by using a unified time zone and timestamp alignment to eliminate duplicate data, resulting in a time-series dataset without duplication and conflict, thus eliminating duplicate patient consultation information; by pushing patient consultation information to the doctor's workstation or mobile terminal responsible for receiving patients, and maintaining a time slice cache and hash index of the patient's last 7 days in memory, low-overhead fast reading is achieved, thus enabling online orchestration and scrolling. The reordering provides millisecond-level data access; by recording the patient's actual arrival time, report readiness time, and waiting time, and binding them with the time slice and rule snapshot at that time, labeled samples that can be directly used for joint model training and threshold optimization are obtained; the joint model, without sharing the original data, classifies the patient's report readiness time window and the patient's arrival risk according to the comprehensive risk score, and sorts them according to the priority of the appointment from high to low, constrains and solves the time slice allocation of the patient's cross-platform test-report readiness-follow-up visit-medication pickup link, and performs local rolling reordering after the patient's appointment schedule and events such as report readiness, congestion, and arrival are triggered, thereby improving the hospital's same-day completion rate and reducing patients' ineffective follow-up visits.

[0043] The specific steps for predicting a user's appointment time are as follows: The patient's testing process is modeled sequentially, and the start time, end time, and queue length of each stage are recorded to obtain a vector that can calculate the duration of each testing stage. This allows the report of each stage of the testing process to be converted into a predictable time quantity. The sequential modeling of the testing process involves dividing it into a sequence of stages: sample reception → triage → machine processing → review → report generation. By aggregating observable signal data from online appointment and testing platforms, the current processing capacity of the hospital and the congestion of testing equipment on that day can be obtained. This allows for the prediction of hospital congestion without relying on the hospital's testing equipment data, providing a stable input for predicting patient appointment times. The observable signal data includes statistics on patient appointment creation / cancellation / rescheduling, patient past appointment fulfillment rates and arrival / late arrival records, pharmacy dispensing window status, holiday / weather / traffic indicators, and institutional resource capacity announcements. By constructing a task profile for the patient, the patient's testing tasks can be determined. Relevant individual characteristics are used to distinguish the delay differences of different tasks in the same environment; the task profile of patients is constructed; the output of test reports is predicted by quantiles to obtain the upper and lower bound windows and median estimate of the report completion time, thus providing input for subsequent orchestration; the coverage of quantiles is monotonically calibrated and abnormal time periods are integrated in the form of small-scale multi-models to obtain a ready time window consistent with the confidence and actual coverage, thereby reducing the occurrence of patients modifying their consultation dates due to window bias; by setting listening events for changes in the stage status of the test process and updating the remaining time according to the stage changes of the test process, a ready time window that narrows as the test process progresses is obtained, thus providing a real-time available window for the orchestration layer when the patient's test report is early or late; by constructing a comprehensive risk scoring formula to calculate the probability of window expiration and the probability of abnormal delay, a risk label that can be directly consumed by the orchestration layer is obtained, and low-risk time periods are preferentially allocated in the link solution. The comprehensive risk score calculation formula is as follows:

[0044] ,

[0045] A represents the overall risk score. The weighting factor for reporting unprepared risks is Readay, which represents the risk of reporting unpreparedness. Here, represents the weighting coefficient for the probability of abnormal delay, and Delay represents the probability of abnormal delay. B represents the weighting coefficient for time-period congestion risk, and B represents the time-period congestion risk. is the weighting factor for buffer adequacy, where Buffer is the buffer adequacy.

[0046] Step two: Assess the probability of users seeking medical treatment and calculate the risk of congestion; construct link-level constraint modeling and initial orchestration.

[0047] Specifically, the steps for judging the probability of a user's visit and calculating the risk of congestion are as follows: By mapping the time slices and patient task profiles described in step one to event flow features, a device-independent input tensor is obtained, enabling generalizable risk modeling across institutions; a monotonic constraint scoring formula is constructed based on the outputs of the patient's no-visit probability, late-visit probability, and congestion probability, thereby prioritizing the allocation of low-risk executable time slots while keeping the original medical resources unchanged, thus reducing ineffective follow-up visits and peak congestion. The monotonic constraint scoring formula is as follows:

[0048] ,

[0049] C represents the appointment priority of the candidate time slot T, and T is the time parameter of the candidate time slot. For normalization function, For biased items, As a weight for the reliability of patient arrival, Let T be the probability function of no patient arrival during the candidate time period T. Weighting for on-time arrival, Let T be the probability function of being late for the candidate time period T. Weighting for low congestion preference, Let T be the congestion risk function for the candidate time period T. For buffer sufficiency weights, Let T be the buffer sufficiency function for the candidate time period T. Weighting for inadequate preparation penalties The report is not ready risk function for candidate time period T. The Buffer and Readay functions here have the same meaning as the formula in the comprehensive risk score.

[0050] After constructing the monotonic divisor scoring formula, uncertainty can be transformed into a directly sortable quantitative signal. By monitoring online events such as patient testing processes, report status changes, and patient check-ins, the report readiness time window and patient consultation risk can be updated online in real time, thereby correcting the time period with reduced disturbance. Through federated training, joint training can be performed without exchanging raw data, thereby obtaining a consistent risk profile across institutions. This balances data privacy and regulatory compliance with the generalization ability of the joint model under cross-platform deployment.

[0051] The specific steps for constructing link-level constraint modeling and initial orchestration are as follows: By abstracting the patient's testing process into time nodes and setting hard and soft constraints, a computable feasible domain for the link is obtained. This limits the solution to the executable solution space, preventing the situation where patients are scheduled for treatment before their reports are ready. The hard constraints require that the order of testing → report ready → follow-up visit → medication pickup must be followed, and that resources cannot be allocated mutually exclusively within the same resource. The soft constraints use patient walking distance, total treatment time, peak congestion during a given time period, and rescheduling radius as optimizable objectives, and these are prioritized and weighed hierarchically. While satisfying the aforementioned hard constraints, the system minimizes waiting time to reduce patient disturbance. By setting hierarchical objectives, it first increases the coverage of feasible chains, then reduces the total consultation time and peak congestion rate, thereby reducing the rescheduling radius and obtaining an interpretable optimization order. This achieves a verifiable balance between the stability of cross-institutional consultation chain orchestration and the efficiency of time resource utilization. The rescheduling radius is determined by limiting the maximum allowable offset of each rescheduling from the original appointment time to a preset threshold and limiting the number of reschedulings per patient per unit time, thus controlling the impact of rescheduling on other appointments within a minimum time range.

[0052] Step 3: Set up a rolling reordering of medical visit trigger events; conduct joint training and iterative optimization of the prediction model while protecting user privacy.

[0053] Specifically, the steps for setting up rolling rescheduling of appointment trigger events are as follows: Subscribing to report-ready events, comprehensive risk scores, and monotonic constraint scores, and identifying affected link subgraphs through the event engine, yields the target patient and their adjacent time slice set, thus precisely limiting the rescheduling scope. The affected link subgraphs represent events such as patients whose examination reports are not yet ready and cannot be scheduled for follow-up appointments, or whose follow-up appointments are delayed due to congestion. Lock-free incremental solving is performed on the affected subgraphs and candidate time slice set, and the event engine is used to generate alternative paths, resulting in a new schedule, thereby reducing rescheduling and minimizing secondary queuing. Upper limits are set on the rescheduling radius and notification window, which are then controlled by the execution module. By controlling the granularity of message delivery, a controllable online correction process can be achieved, which can suppress cascading disturbances. The upper limit is set to limit the maximum allowable offset of each rescheduling to no more than 24 hours before or after the original appointment time, and limit the number of reschedulings for a single patient within 7 days to no more than once, thereby controlling the spread of rescheduling and reducing the burden on the system and patients. By writing conflict proofs of the reason for the change into the time slice replaced after rescheduling, a traceable rescheduling closed loop is obtained, which can improve the effectiveness of authorization verification. By providing a backup strategy triggered by high-risk tags, the strategy unit allocates spare time slices to the triggered backup strategy, thereby obtaining a fault-tolerant chain from failure to rollback to recovery, which can reduce invalid follow-up visits.

[0054] The specific steps for joint training and iterative optimization of predictive models while protecting user privacy are as follows: First, sub-models trained on local data for both community and specialist sides are used. Intermediate representations are extracted from the trained sub-models. Then, the federated coordinator securely aggregates the extracted representations and the gradients of the model parameters to obtain a global model that does not leak the original data, thereby improving the generalization ability of the global model under cross-platform conditions. Second, a privacy-preserving global update vector is obtained by using a secure aggregation protocol on the submitted update packages and adding configurable differential privacy noise after aggregation. This satisfies compliance requirements while reducing model parameter leakage. Risks; By performing knowledge distillation on the aggregated global model and generating lightweight inference sub-models, which are periodically distributed to hospitals and community institutions, efficient predictors can be deployed in weak network and low computing power environments, thereby improving edge prediction capabilities and reducing the load on central inference; the sub-models are local predictors trained on local data, preserving the institution's delay and performance characteristics, and can provide aggregateable intermediate representations without exchanging original data; the global model is aggregated and trained at the central end by combining encrypted gradients submitted from various platforms with configurable differential privacy noise, thereby obtaining a unified prediction model that statistically takes into account the characteristics of each institution and satisfies privacy constraints. This provides a consistent benchmark for downstream distillation and unified strategy deployment. The lightweight inference sub-model generates a parameter-compressed student model using the aggregated global model combined with model pruning and quantization compression techniques. This allows for real-time deployment of an edge inferencer that maintains acceptable error on key prediction metrics even in weak network and low computing power environments. After each round of model deployment, coverage and quantile error are calculated in each institution using a reserved validation set, and the aggregated metrics are returned to obtain consistent performance evaluation results across institutions. This ensures that key metrics are not reduced after model updates and supports rollback decisions. By comparing the actual arrival time of patients during online treatment with the time when reports are issued... The rescheduling results are periodically fed back as monitoring tags, and A / B trials are conducted in an isolated environment to verify the strategy, thereby obtaining empirically validated parameter adjustment schemes. This forms a measurable continuous optimization closed loop and improves the verifiability of implementation. The A / B trial verification strategy involves randomly dividing the patient appointment links into a control group and an experimental group within the same time window. The adjusted strategy is issued to the control group, while the experimental group retains the original strategy. Key performance indicators are continuously collected to obtain experimental data for both the experimental and control groups. The experimental data is used to determine whether the new strategy has improved compared to the original strategy. If it has improved, it is retained; otherwise, it is reverted.

[0055] like Figure 2The figure shows the technical effect of an online outpatient intelligent appointment method based on deep learning. The black bars represent the technical effect of the present invention, while the gray bars represent the technical effect achieved by the prior art. It can be clearly seen that the same-day completion rate and equipment utilization rate of the present invention are significantly higher than those of the prior art. The higher these two values, the better the effect. In particular, the invalid follow-up rate, average waiting time, and percentile waiting time are significantly lower than those of the prior art. The lower these three data are, the better the effect. It can be seen that the effect of the present invention is significantly better than that of the prior art.

[0056] Example 2, based on Example 1, presents a deep learning-based intelligent online outpatient appointment method, specifically as follows:

[0057] Step 1: Perform data synchronization and initialization;

[0058] Specifically, the hospital's unified scheduling platform periodically pulls and merges resource time slices, testing task anchors, and past performance records from the community and specialist departments, and de-identifies and maps key fields to obtain a unified and computable set of time slices and patient link anchors, thereby providing a timely and compliant data foundation for subsequent readiness prediction and constraint solving.

[0059] Step two: Conduct risk assessment and report readiness forecasting;

[0060] Specifically, by inputting staged features and online load vectors into the merged testing task on the central prediction service and performing quantile interval prediction, the probability of patient reports not being ready, the probability of abnormal delays, and the time-period congestion score are calculated and combined into a comprehensive risk with configured weights. This yields a prediction output that includes a ready time window and discretized risk labels, thereby providing a comparable and consumable risk constraint effect for link-level orchestration.

[0061] Step 3: Reorganize and distribute patient appointment schedules;

[0062] Specifically, by using a link-level constraint solver to solve for feasible appointment sequences on the time slice graph and sending the initial schedule to the execution end of each institution, and simultaneously performing lock-free incremental solving and local replacement on the affected subgraph when the progress reporting event is triggered, and issuing rescheduling notifications according to a limited radius, a final time slice instruction set with minimal disturbance adjustment and executable by all participants can be obtained. This can reduce the invalid re-visit rate, reduce the impact of rescheduling, and ensure scheduling consistency without expanding capacity.

[0063] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent online outpatient appointment booking based on deep learning, characterized in that: Access the basic data collection; establish cross-system field mapping relationships based on the basic data collection; By performing a time-series model of the patient's testing process, a vector that can be used to calculate the duration of each testing stage is obtained; the patient's testing process is modeled in a time sequence. By using quantiles to predict the output of the test report, the upper and lower bound windows and median estimate of the report completion time are obtained. By constructing a comprehensive risk scoring formula to calculate the window expiration probability and the abnormal delay probability, risk labels that can be directly consumed by the orchestration layer are obtained. A monotonic constraint scoring formula is constructed based on the outputs of the patient's no-show probability, late consultation probability, and consultation congestion probability. Construct link-level constraint modeling and initial orchestration; Set up a rolling reordering of medical visit trigger events; conduct joint training and iterative optimization of prediction models while protecting user privacy.

2. The online outpatient intelligent appointment method based on deep learning according to claim 1, characterized in that: The specific steps for establishing cross-system field mapping relationships based on the basic data set are as follows: By irreversibly hashing the patient's personal identifier and de-identifying the patient's medical number, a controlled data availability boundary is obtained by retaining only the association keys available during off-peak hours. By uniformly segmenting and recording the doctor's registration source, the time period of the testing equipment, and the pharmacy window into time slices of fixed granularity, a time resource view that can be computed across platforms is obtained; by setting ordered relationships, low-peak interval thresholds, mutual exclusion constraints, and marking the positions of patients' medical treatment processes, a set of rules for determining the patient's medical treatment process is obtained.

3. The online outpatient intelligent appointment method based on deep learning according to claim 1, characterized in that: The specific steps for predicting the output of the test report using quantiles are as follows: By using monotonic calibration of quantile coverage and integrating abnormal time periods in the form of small-scale multi-model, a ready time window consistent with the confidence level and actual coverage is obtained; by setting listening events for changes in the stage status of the inspection process and updating the remaining time according to the changes in the stage of the inspection process, a ready time window that narrows as the inspection process progresses is obtained.

4. The online outpatient intelligent appointment method based on deep learning according to claim 1, characterized in that: The specific steps for calculating the window expiration probability and the abnormal delay probability by constructing a comprehensive risk scoring formula are as follows: By constructing a comprehensive risk scoring formula to calculate the window expiration probability and the abnormal delay probability, risk labels that can be directly consumed by the orchestration layer are obtained. The comprehensive risk score calculation formula is as follows: ， A represents the overall risk score. The weighting factor for reporting unprepared risks is Readay, which represents the risk of reporting unpreparedness. Here, represents the weighting coefficient for the probability of abnormal delay, and Delay represents the probability of abnormal delay. B represents the weighting coefficient for time-period congestion risk, and B represents the time-period congestion risk. is the weighting factor for buffer adequacy, where Buffer is the buffer adequacy.

5. The online outpatient intelligent appointment method based on deep learning according to claim 1, characterized in that: The specific steps for constructing a monotonic constraint scoring formula based on the outputs of the patient's no-show probability, late arrival probability, and congestion probability are as follows: Based on the outputs of patient no-show probability, late consultation probability, and consultation congestion probability, a monotonic constraint scoring formula is constructed. While maintaining the original medical resources, priority is given to allocating low-risk, feasible time slots. The monotonic constraint scoring formula is as follows: ， C represents the appointment priority of the candidate time slot T, and T is the time parameter of the candidate time slot. For normalization function, For biased items, As a weight for the reliability of patient arrival, Let T be the probability function of no patient arrival during the candidate time period T. Weighting for on-time arrival, Let T be the probability function of being late for the candidate time period T. Weighting for low congestion preference, Let T be the congestion risk function for the candidate time period T. For buffer sufficiency weights, Let T be the buffer sufficiency function for the candidate time period T. Weighting for inadequate preparation penalties The report-in-readiness risk function is the candidate time period T.

6. The online outpatient intelligent appointment method based on deep learning according to claim 1, characterized in that: The specific steps for constructing link-level constraint modeling and initial orchestration are as follows: By abstracting the patient's testing process into time nodes and setting hard and soft constraints, a computable feasible domain for the link is obtained. By setting hierarchical objectives, the coverage of feasible links is first increased, and then the total consultation time and peak congestion rate are reduced, thereby reducing the rearrangement radius and obtaining an interpretable optimization order.

7. The online outpatient intelligent appointment method based on deep learning according to claim 1, characterized in that: The specific steps for setting up the scrolling and reordering of the visit trigger event are as follows: By subscribing to report-ready events, comprehensive risk scores, and monotonic constraint scores, and identifying affected link subgraphs through the event engine, the target patient and their adjacent time slice set are obtained. Lock-free incremental solving is performed on the affected subgraph and the candidate time slice set, and the event engine is scheduled to generate alternative paths to obtain a new schedule, thereby reducing rescheduling and minimizing secondary queuing. By setting upper limits for the rearrangement radius and notification window, the execution module controls the granularity of message delivery, thus achieving a controllable online correction process.

8. The online outpatient intelligent appointment method based on deep learning according to claim 7, characterized in that: The specific steps for setting up the scrolling reordering of the visit trigger event also include: By writing conflict proofs to the replacement time slices after rearrangement and the reason for the change, a traceable rearrangement loop is obtained, which can improve the verifiability of authorization. By providing a backup for the candidate strategy triggered by high-risk tags, the strategy unit allocates spare time slices to the triggered candidate strategy, resulting in a fault-tolerant chain from failure to rollback to recovery.

9. The online outpatient intelligent appointment method based on deep learning according to claim 1, characterized in that: The specific steps for jointly training and iteratively optimizing the prediction model while protecting user privacy are as follows: By training sub-models on local data for both community and specialist sides, extracting intermediate representations from the trained sub-models, and then using a federated coordinator to securely aggregate the gradients of the extracted representations and sub-model parameters, a global model is obtained without leaking the original data. By employing a secure aggregation protocol on the submitted update packages and adding configurable differential privacy noise after aggregation, a privacy-preserving global update vector is obtained. By performing knowledge distillation on the aggregated global model and generating lightweight inference sub-models that are periodically distributed to hospitals and community institutions, an efficient predictor can be deployed in environments with weak networks and low computing power.

10. The method for intelligent online outpatient appointment based on deep learning according to claim 9, characterized in that: The specific steps for jointly training and iteratively optimizing the prediction model while protecting user privacy also include: After each round of model distribution, coverage and quantile error are calculated in each institution using a reserved validation set, and aggregated metrics are fed back to obtain consistent performance evaluation results across institutions. By periodically feeding back the actual arrival time, report issuance time, and rescheduling results of patients' online procedures as supervision labels and conducting A / B trials in an isolated environment to verify the strategy, an empirically validated parameter adjustment scheme is obtained.

Images