A hierarchical diagnosis and treatment based regional coordination follow-up method and system for tumor patients

Abstract

The application discloses a tumor patient regional coordination follow-up method and system based on hierarchical diagnosis and treatment, belongs to the technical field of medical health information processing, and comprises the following steps: acquiring original clinical data of tumor patients, extracting an encrypted local gradient vector and an organization identifier; inputting the encrypted local gradient vector into a federal aggregation server to output a global risk feature vector; calling a target hierarchical evaluation model according to the global risk feature vector and real-time regional resource parameters to output a diagnosis and treatment level identifier; calling a target referral path optimizer according to the diagnosis and treatment level identifier and real-time medical institution service capability data to output a minimum cost referral path; generating a referral scheduling instruction based on the minimum cost referral path and the diagnosis and treatment level identifier and issuing the referral scheduling instruction; collecting execution feedback data after the referral scheduling instruction is executed, inputting the execution feedback data into the federal aggregation server, and updating encrypted local gradient vector generation parameters and model weight parameters. The application improves the utilization efficiency of medical resources.

Description

Technical Field

[0001] This invention relates to the field of medical and health information processing technology, specifically to a method and system for regional collaborative follow-up of cancer patients based on hierarchical diagnosis and treatment. Background Technology

[0002] Follow-up management of cancer patients is a key link in the implementation of the hierarchical medical system. In recent years, with the popularization of electronic health records, the construction of regional medical information platforms and the development of artificial intelligence-assisted decision-making technology, some regions have initially realized the electronic archiving of cancer patient information and the function of basic follow-up reminders. Some tertiary hospitals have also tried to build in-hospital follow-up systems, combined with AI models to perform static scoring of recurrence risk, thus promoting the digitalization process of chronic cancer management.

[0003] However, the existing technology system still has significant structural defects: the grading standards are static and rigid, detached from resource realities: existing follow-up grading is mostly based on TNM staging or fixed scoring tables, without integrating real-time load parameters such as regional beds, experts, and equipment, often resulting in high-risk patients being stranded at the primary care level or low-risk patients crowding out specialist resources; referral paths rely on manual coordination and lack intelligent optimization: referrals are mostly designated by doctors' experience or administrative agreements, without quantifying the multi-dimensional costs such as waiting costs, transportation burdens, medical insurance compatibility, and expert matching, resulting in low path efficiency, poor patient experience, and high rate of implementation disconnect; there is no closed-loop feedback mechanism, and the model continuously degrades: after system deployment, there is a lack of model self-optimization capabilities based on real referral execution data, and evaluation biases accumulate over time, path recommendations become invalid, and the system cannot adapt to disease evolution and resource fluctuations. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: how to achieve cross-institutional privacy and security collaboration for cancer patients, accurate allocation of follow-up responsibilities, intelligent scheduling of referral paths, and continuous evolution of model parameters through federated gradient aggregation, resource-aware dynamic hierarchical classification, multi-objective path optimization, and closed-loop feedback mechanisms, while ensuring efficient utilization of regional medical resources and controllable closed-loop clinical decision-making.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for regional collaborative follow-up of tumor patients based on hierarchical diagnosis and treatment, which includes: acquiring the original clinical data of tumor patients, and extracting encrypted local gradient vectors and institutional identifiers;

[0007] The encrypted local gradient vector is input into the federated aggregation server, and the global risk feature vector is output.

[0008] Based on the global risk feature vector and real-time regional resource parameters, the target hierarchical assessment model is invoked to output the diagnosis and treatment level identifier;

[0009] Based on the diagnosis and treatment level identifier and real-time medical institution service capacity data, the target referral path optimizer is invoked to output the minimum cost referral path.

[0010] Based on the minimum cost referral path and the treatment level identifier, a referral scheduling instruction is generated and issued;

[0011] The execution feedback data after the referral scheduling instruction is completed is collected, and the execution feedback data is input into the federated aggregation server to update the encrypted local gradient vector generation parameters and model weight parameters.

[0012] As a preferred embodiment of the regional collaborative follow-up method for tumor patients based on hierarchical diagnosis and treatment described in this invention, the step of extracting encrypted local gradient vectors and institutional identifiers includes:

[0013] Structured clinical data is collected from the electronic health records of local medical institutions. The structured clinical data includes TNM staging, ECOG score, treatment plan, laboratory indicators and follow-up adherence records.

[0014] The structured clinical data is standardized and mapped to a numerical feature vector within a unified feature space.

[0015] The numerical feature vector is input into the local pre-trained risk prediction model, the model gradient is calculated, and homomorphic encryption is performed on the model gradient to generate the encrypted local gradient vector.

[0016] Extract the unique code of the local medical institution in the regional collaborative network and use it as the institution identifier.

[0017] As a preferred embodiment of the regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment described in this invention, the output global risk feature vector includes:

[0018] Receive multiple encrypted local gradient vectors and their corresponding organization identifiers, and batch aggregate them according to a preset time window;

[0019] Perform homomorphic addition aggregation on all the encrypted local gradient vectors within the same batch to generate an encrypted global gradient vector;

[0020] The encrypted global gradient vector is sent to a private key decryptor deployed in a secure execution environment to perform private key decryption and obtain the plaintext global gradient vector.

[0021] The plaintext global gradient vector is input into the central risk feature extraction model, and the global risk feature vector is output through forward propagation.

[0022] As a preferred embodiment of the regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment described in this invention, wherein: the invocation of the target hierarchical assessment model includes:

[0023] Obtain real-time regional resource parameters, including bed occupancy rate, remaining number of specialist appointment slots, equipment queuing time, and doctor saturation.

[0024] Based on the dimensional distribution of the global risk feature vector and the load range of the real-time regional resource parameters, the target hierarchical assessment model is matched from the hierarchical assessment model set.

[0025] The global risk feature vector is input into the target classification and assessment model, and the original risk score is output.

[0026] Based on the original risk score and the preset grading threshold table, the diagnosis and treatment level identifier is mapped and output.

[0027] As a preferred embodiment of the regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment described in this invention, wherein: the invocation of the target referral pathway optimizer includes:

[0028] Acquire real-time medical institution service capacity data, which includes the number of available beds, expert scheduling status, equipment idle time, medical insurance compatibility level, and geographical accessibility score;

[0029] Based on the service efficiency range of the treatment level identifier and the service capability data, the target referral path optimizer is matched from the referral path optimizer set;

[0030] Construct a weighted directed referral graph, where nodes represent medical institutions and edge weights are the comprehensive referral costs calculated by weighting waiting time, transportation costs, medical insurance compatibility, and expert matching.

[0031] The weighted directed referral graph is input into the target referral path optimizer, and the improved Dijkstra algorithm is used to output the minimum cost referral path.

[0032] As a preferred embodiment of the regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment described in this invention, the step of generating and issuing referral scheduling instructions includes:

[0033] The minimum cost referral path is analyzed to extract the source institution, target institution, transit institution, and the estimated time nodes for each segment of the referral;

[0034] Based on the diagnosis and treatment level identifier and path node role, a structured referral scheduling instruction is generated. The referral scheduling instruction includes the patient identifier, target institution, execution time limit, examination item list and data sharing permissions.

[0035] The referral scheduling instruction is sent to the local collaborative terminal of the corresponding institution through the regional collaborative message bus, and the instruction status listener is started to collect execution receipts.

[0036] As a preferred embodiment of the regional collaborative follow-up method for tumor patients based on hierarchical diagnosis and treatment described in this invention, the step of updating the encrypted local gradient vector generation parameters and model weight parameters includes:

[0037] After the referral scheduling instruction is marked as "completed", structured execution feedback data is collected. The execution feedback data includes the actual referral time, the receipt confirmation time, the examination completion status, the satisfaction score, and the complication marker.

[0038] The execution feedback data is input into the federated aggregation server to perform feature alignment and backpropagation, generating a global feedback gradient.

[0039] The global feedback gradient is encrypted and distributed to the local computing nodes of each participating institution to update the encrypted local gradient vector generation parameters, the weight parameters of the hierarchical evaluation model, and the weight parameters of the referral path optimizer.

[0040] This invention provides a regional collaborative follow-up system for cancer patients based on hierarchical diagnosis and treatment.

[0041] To address the aforementioned technical problems, the present invention further provides the following technical solution: a regional collaborative follow-up system for cancer patients based on hierarchical diagnosis and treatment, comprising: a data extraction module, used to acquire the original clinical data of cancer patients and extract encrypted local gradient vectors and institutional identifiers;

[0042] The federated aggregation module is used to input the encrypted local gradient vector into the federated aggregation server and output a global risk feature vector.

[0043] The hierarchical assessment module is used to call the target hierarchical assessment model based on the global risk feature vector and real-time regional resource parameters, and output the diagnosis and treatment level identifier.

[0044] The path optimization module is used to call the target referral path optimizer based on the diagnosis and treatment level identifier and real-time medical institution service capacity data, and output the minimum cost referral path;

[0045] The instruction scheduling module is used to generate and issue referral scheduling instructions based on the minimum cost referral path and the treatment level identifier.

[0046] The feedback update module is used to collect execution feedback data after the referral scheduling instruction is completed, input the execution feedback data into the federated aggregation server, and update the encrypted local gradient vector generation parameters and model weight parameters.

[0047] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the aforementioned method for regional collaborative follow-up of tumor patients based on hierarchical diagnosis and treatment.

[0048] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the aforementioned method for regional collaborative follow-up of tumor patients based on hierarchical diagnosis and treatment.

[0049] The beneficial effects of this invention are as follows: By extracting encrypted gradients from local clinical data and binding them with institutional identifiers, privacy-preserving preprocessing is achieved without the original data leaving the domain, laying a compliant foundation for cross-institutional collaboration; by generating global risk features through federated aggregation and secure decryption, lossless fusion and semantic enhancement of multi-source heterogeneous data in the encrypted space are achieved, constructing a unified risk perception; by combining a dynamic matching and hierarchical model with real-time resource parameters, adaptive coupling between patient risk assessment and regional load capacity is achieved, ensuring accurate allocation of follow-up responsibilities and avoiding resource misallocation; by constructing a multi-dimensional cost referral graph and calling the optimizer to calculate the minimum cost path, the referral decision-making process is upgraded from experience-driven to data-driven, taking into account efficiency, cost, quality, and patient experience; by generating structured role instructions and monitoring execution status, the referral task is accurately transformed from planning to implementation and the process is controllable, preventing responsibility gaps and execution disconnects; by collecting real execution feedback to update model parameters, the system is transitioned from static deployment to continuous evolution, ensuring long-term stability and adaptive optimization capabilities. Through a six-step closed-loop collaboration, a safe, compliant, accurately graded, intelligently scheduled, closed-loop executed, and self-evolving regional collaborative follow-up technology system for cancer patients is ultimately achieved, significantly improving the efficiency of medical resource utilization, the quality of patient management, and the level of regional collaborative governance. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 The following is an overall flowchart of a regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment, provided as an embodiment of the present invention.

[0052] Figure 2 A flowchart illustrating a regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment, provided as an embodiment of the present invention;

[0053] Figure 3 This is a computer device diagram of a regional collaborative follow-up method for tumor patients based on hierarchical diagnosis and treatment, provided as an embodiment of the present invention. Detailed Implementation

[0054] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0055] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment, including:

[0056] S1: Obtain raw clinical data from cancer patients and extract encrypted local gradient vectors and institutional identifiers;

[0057] S2: Input the encrypted local gradient vector into the federated aggregation server and output the global risk feature vector;

[0058] S3: Based on the global risk feature vector and real-time regional resource parameters, call the target classification assessment model and output the diagnosis and treatment level identifier;

[0059] S4: Based on the diagnosis and treatment level identifier and real-time medical institution service capacity data, call the target referral path optimizer and output the minimum cost referral path;

[0060] S5: Generate and issue referral scheduling instructions based on the minimum cost referral path and treatment level identifier;

[0061] S6: Collect execution feedback data after the referral scheduling instruction is completed, input the execution feedback data into the federated aggregation server, and update the encrypted local gradient vector generation parameters and model weight parameters.

[0062] It should be noted that follow-up management of cancer patients involves collaboration among multiple levels of medical institutions. Patient clinical data is scattered across different institutions and is strictly restricted by privacy regulations, making it impossible to directly aggregate cross-domain data. Traditional centralized modeling is insufficient to achieve accurate risk assessment. At the same time, the dynamic evolution of patients' conditions and the real-time fluctuation of regional medical resource load make it easy to misallocate follow-up responsibilities, inefficient referral pathways, and miss high-risk patients if static grading standards or manual referral allocation are used. In addition, the existing system lacks an execution feedback loop and cannot dynamically optimize the evaluation model and scheduling strategy based on the actual referral effect, resulting in a continuous decline in regional collaboration efficiency.

[0063] Therefore, addressing the four core issues of data silos, static evaluation, resource mismatch, and lack of feedback optimization, the following steps (S1-S6) are implemented: While protecting the original data from leaving the domain, a federated mechanism is used to extract and aggregate encrypted local gradient vectors to construct a global risk feature vector, overcoming the conflict between privacy and collaboration; a hierarchical model is dynamically invoked in conjunction with real-time regional resource parameters to output a diagnosis and treatment level identifier adapted to the current resource environment, achieving "environment-aware" intelligent hierarchical classification; based on the diagnosis and treatment level and the institution's service capabilities, a referral path optimizer is automatically matched to calculate the minimum-cost referral path, ensuring optimal resource allocation; and feedback data from referral execution is collected to feed back into the federated model and evaluation weights, forming a closed loop of "evaluation → scheduling → execution → optimization," continuously improving the accuracy of regional collaborative follow-up and the system's adaptive capabilities.

[0064] Example 2, refer to Figure 1 and Figure 2 This is the second embodiment of the present invention, which provides a regional collaborative follow-up method for tumor patients based on hierarchical diagnosis and treatment.

[0065] S1: Obtain raw clinical data from cancer patients and extract encrypted local gradient vectors and institutional identifiers;

[0066] It should be noted that tumor follow-up involves collaboration among multiple institutions, with data being scattered and subject to privacy restrictions. Traditional methods rely on static rules, which can easily lead to inaccurate assessments, inefficient referrals, and gaps in accountability. Patient conditions evolve dynamically, and resources fluctuate in real time. Therefore, it is necessary to achieve accurate risk assessment, intelligent pathway optimization, and closed-loop feedback without sharing raw data, in order to support safe, efficient, and adaptive regional collaborative follow-up under the tiered medical system.

[0067] Specifically, in this embodiment, the original clinical data of cancer patients to be followed up are obtained from the electronic health records of primary community health service centers connected to the regional cancer collaborative follow-up platform. Primary community health service centers are legally designated primary healthcare institutions responsible for the initial screening, basic follow-up, and pre-referral management of cancer patients. Their electronic health records contain the patient's past medical visits, laboratory examinations, treatment records, and follow-up registration data.

[0068] Further, the original clinical data included: pathological diagnosis results (including TNM staging), performance status score (ECOGPS score), current treatment plan (surgery / chemotherapy / targeted therapy), blood routine and liver and kidney function indicators for the past three months, date of the most recent follow-up, and compliance markers (whether follow-up visits were on time and whether examinations were completed).

[0069] In practical applications, raw clinical data is automatically collected through local federated node agents deployed on the medical institution's intranet, according to preset data extraction rules. The collection frequency is once every 24 hours or triggered in real time when the patient's status is updated. During the collection process, unstructured free text is automatically filtered, and only predefined fields are extracted. These fields are then mapped to unified feature codes using the FHIRR4 standard to ensure semantic consistency across institutions. After collection, the local federated node agent immediately calls a lightweight risk prediction model (such as logistic regression or a small neural network) to calculate the gradient and uses the Paillier homomorphic encryption algorithm to encrypt the gradient vector, generating an encrypted local gradient vector. At the same time, the unique code (such as "SH-PD-C003") of the primary community health service center in the regional medical collaboration network is read from the institution's registry and used as the institution identifier. This code is then encapsulated and uploaded to the federated aggregation server.

[0070] S2: Input the encrypted local gradient vector into the federated aggregation server and output the global risk feature vector;

[0071] It should be noted that the original clinical data is scattered across various institutions and plaintext aggregation is prohibited, making traditional centralized training infeasible. It is necessary to securely aggregate gradients in a encrypted space, and then extract global features after decryption. This approach ensures privacy compliance while providing a unified, high-dimensional, cross-domain risk characterization basis for dynamic grading.

[0072] S21: Receive encrypted local gradient vectors and corresponding institution identifiers uploaded from local nodes of multiple medical institutions, and perform batch aggregation according to a preset time window;

[0073] Specifically, the federated aggregation server sets a fixed periodic time window (e.g., every 6 hours). At the end of the window, the received encrypted local gradient vectors are packaged into the current batch according to the order of upload by the institution, and a unique batch ID is attached.

[0074] Furthermore, the window length is dynamically adjusted based on the data upload completion rate of regional institutions: if more than 80% of institutions complete the upload within 30 minutes, the window is shortened to 2 hours; if the number of delayed institutions exceeds the threshold, the window is extended and a re-upload mechanism is triggered to ensure batch integrity.

[0075] S22: Perform secure aggregation operation on all encrypted local gradient vectors within the same batch, and use homomorphic addition to calculate the encrypted global gradient vector without performing decryption operation during the process;

[0076] Specifically, the federated aggregation server calls the built-in homomorphic computation engine, performs ciphertext addition on the encrypted local gradient vector based on the Paillier scheme, and outputs an encrypted global gradient vector without ever touching the plaintext data.

[0077] Furthermore, the homomorphic computing engine verifies the organization's digital signature and whitelist status before accumulation, and assigns dynamic weights based on the organization's historical data quality score, performing weighted homomorphic aggregation to improve the model's representativeness.

[0078] Furthermore, parallel threads are started to perform homomorphic addition in blocks according to gradient feature dimensions. After merging the results, the contribution of each institution and the calculation time are recorded. All intermediate ciphertext is processed in an isolated memory area and cleared immediately after the operation to prevent memory residue and side-channel attacks.

[0079] S23: Send the encrypted global gradient vector to the private key decryptor deployed in the secure execution environment. The private key decryptor deployed in the secure execution environment performs private key decryption to obtain the plaintext global gradient vector.

[0080] Specifically, the private key decryptor deployed in the secure execution environment runs in an Intel SGX enclave. After the encrypted global gradient vector is passed in, the private key is loaded and decryption is performed within the enclave. After decryption is completed, the private key cache is destroyed immediately, and the plaintext global gradient vector is output.

[0081] Furthermore, before decryption, the aggregate integrity hash and service signature of the input data are verified, and decryption is only initiated after verification is passed; the decryption result is written to an isolated memory area outside the enclave, and an access control policy is set to restrict the processing unit used for subsequent risk feature extraction to read it within a specified time window.

[0082] S24: Input the plaintext global gradient vector into the central risk feature extraction model, calculate it through model forward propagation, output the global risk feature vector, and cache it in the feature shared memory for subsequent use.

[0083] Specifically, the central risk feature extraction model is a 3-layer fully connected neural network that receives the plaintext global gradient vector, outputs a 128-dimensional global risk feature vector through forward propagation, and writes it to Redis shared memory with the key name "GlobalFeature_<batch ID>".

[0084] Furthermore, the model performs Z-score standardization before the input layer to eliminate differences in data scale between institutions. When outputting, a timestamp, a list of institutions, and the model version number are added simultaneously, and a 72-hour automatic expiration policy is set for subsequent grading and evaluation processes to trace and call upon the data.

[0085] S3: Based on the global risk feature vector and real-time regional resource parameters, call the target classification assessment model and output the diagnosis and treatment level identifier;

[0086] It should be noted that both patient risk and regional resources are dynamic, making static grading models prone to inaccuracy. Therefore, it is necessary to combine real-time load parameters with model adaptation to achieve intelligent grading driven by both risk and resources. This ensures accurate allocation of follow-up responsibilities and prevents high-risk patients from being relegated to lower-risk areas or low-risk patients from crowding out high-quality resources.

[0087] S31: Obtain real-time regional resource parameters, including the current bed occupancy rate of each medical institution, the remaining number of expert appointment slots, the queuing time for examination equipment, and the saturation index of primary care physicians.

[0088] Specifically, through the regional medical resource scheduling interface, the HIS and LIS of each access institution are polled every 15 minutes to collect real-time regional resource parameters and store them in the resource status table.

[0089] Furthermore, the collected data is filtered by moving average to eliminate instantaneous fluctuations, and outliers (such as queuing time > 24 hours) are marked as "resource anomalies" to trigger a manual review process.

[0090] Furthermore, dynamic weight coefficients are assigned to each type of resource parameter (e.g., expert appointment weight = 0.4, bed weight = 0.3, equipment queuing weight = 0.2, doctor saturation weight = 0.1), and a weighted composite "institutional comprehensive load index" is synthesized, which serves as the basis for matching the subsequent hierarchical evaluation model.

[0091] S32: Match the target hierarchical assessment model from the set of hierarchical assessment models based on the dimensional distribution of the global risk feature vector and the load range of real-time regional resource parameters.

[0092] Specifically, the hierarchical assessment model set pre-stores multiple models, each model corresponding to a combination of a risk feature principal component interval and a resource load interval; calculates the L2 norm and principal component direction of the global risk feature vector, and combines it with the institution's comprehensive load index to determine the target hierarchical assessment model through nearest neighbor matching.

[0093] Furthermore, before matching, the model set is cached and warmed up, and frequently used models are loaded into memory; during matching, a KD-Tree index is used to accelerate the query, and the response time is controlled within 50ms.

[0094] Furthermore, if there is no exact matching model, the interpolation mechanism is activated: two models in adjacent load intervals are selected, and temporary model parameters are generated by linear interpolation according to the resource load ratio, and the interpolation events are recorded for subsequent training supplementation.

[0095] S33: Input the global risk feature vector into the target hierarchical assessment model, perform forward inference calculation, and output the original risk score;

[0096] Specifically, the target classification assessment model is a lightweight multilayer perceptron. The input layer receives a 128-dimensional global risk feature vector, which is then processed through two hidden layers (ReLU activation) and a Softmax output layer to generate an original risk score of 0 to 100.

[0097] Furthermore, before inference, the input vector is normalized to Min-Max and scaled to the [0,1] interval to avoid the impact of feature scale differences on scoring stability; the inference result is appended with the model version number and confidence interval for audit traceability.

[0098] Furthermore, multiple candidate models are loaded in parallel to perform inference, and an integrated voting mechanism (such as weighted average or maximum confidence) is used to determine the final original risk score, thereby improving the robustness of the assessment.

[0099] S34: Based on the original risk score and combined with the preset grading threshold table, the diagnosis and treatment level identifier is mapped and output. The diagnosis and treatment level identifier includes three levels: primary care follow-up, regional center follow-up, and specialist hospital follow-up.

[0100] Specifically, the built-in tiered threshold table maps the original risk score of 0-30 points to "primary care follow-up", 31-70 points to "regional center follow-up", and 71-100 points to "specialist hospital follow-up", directly outputting the corresponding treatment level identifier.

[0101] Furthermore, the grading threshold table supports dynamic adjustment: the threshold boundaries are automatically updated monthly based on referral execution feedback data and patient outcome data using the ROC curve optimization method to ensure a balance between grading sensitivity and specificity.

[0102] Furthermore, when the original risk score is within ±5 points of the threshold boundary, the "dual-institution collaborative assessment" process is initiated, and the patient data is simultaneously pushed to the adjacent level of institution, where doctors from both sides will consult and confirm the final treatment level identification within 24 hours.

[0103] S4: Based on the diagnosis and treatment level identifier and real-time medical institution service capacity data, call the target referral path optimizer and output the minimum cost referral path;

[0104] It should be noted that referrals are not simple jumps; they involve a comprehensive consideration of factors such as waiting time, distance, medical insurance coverage, and specialist resources. Traditional manual assignment is inefficient and provides a poor user experience. A path map that perceives the service efficiency of institutions needs to be constructed, and an intelligent optimizer should be used to calculate the minimum-cost path to achieve optimal resource allocation and a balance between patient experience.

[0105] S41: Obtain real-time medical institution service capacity data, which includes the target institution’s current number of available beds, expert scheduling status, available time slots for examination equipment, medical insurance reimbursement compatibility level, and geographical accessibility score.

[0106] Specifically, through the regional medical collaboration platform API, HIS, PACS, and medical insurance data of each target medical institution are synchronized every 10 minutes, and real-time service capability data of medical institutions is extracted and cached in the service capability database.

[0107] Furthermore, the validity of service capability data is validated: if an organization returns null values ​​or times out three times in a row, it is marked as "service unavailable" and its participation in path calculation is suspended until manual confirmation is received.

[0108] Furthermore, dynamic weights are assigned to each service capability indicator (bed weight = 0.25, expert weight = 0.3, equipment weight = 0.2, medical insurance weight = 0.15, geographical weight = 0.1), and these weights are combined to form the "institutional service efficiency index," which serves as the core factor for path cost calculation.

[0109] S42: Match the target referral path optimizer from the referral path optimizer set based on the service efficiency range of the diagnosis and treatment level identifier and real-time medical institution service capacity data;

[0110] Specifically, the referral path optimizer set pre-stores multiple optimizers, each of which corresponds to a combination of treatment level and service efficiency. Based on the treatment level identifier (such as "specialist hospital follow-up") and the target institution's "service efficiency index" range, the target referral path optimizer is quickly located through hash mapping.

[0111] Furthermore, a memory cache pool is established for the high-frequency combination optimizer. When a match is found, the cache is read first, and the response latency is less than 20ms when the hit rate is >95%. When a miss occurs, the data is loaded from the disk and automatically cached.

[0112] Furthermore, if there is no exact matching optimizer currently available, the "policy migration mechanism" is activated: an optimizer with adjacent performance intervals is selected, its cost function parameters are adjusted according to the performance difference ratio, a temporary optimizer is generated, and the migration event is recorded for subsequent training.

[0113] S43: Construct a weighted directed referral graph, where nodes represent medical institutions and edge weights represent the comprehensive referral cost. The comprehensive referral cost is calculated by weighting waiting time, transportation costs, medical insurance compatibility, and expert matching.

[0114] Specifically, a weighted directed referral graph is constructed using the current patient's institution as the source node and institutions that meet the diagnostic and treatment level identifiers as candidate target nodes. The initial values ​​of the edge weights are calculated based on the straight-line distance between institutions, the average waiting time for appointments, the difference in medical insurance reimbursement ratios, and the matching degree of expert expertise.

[0115] Furthermore, a dynamic cost factor is introduced: traffic cost = inter-institution driving time × regional congestion coefficient, expert matching degree = cosine similarity between patient pathology type and doctor's expertise vector, and medical insurance compatibility degree = 1 − (difference in co-payment ratio / benchmark co-payment ratio), to ensure that the cost is close to the actual experience.

[0116] Furthermore, normalization compression is performed on the edge weights: after scaling the costs of each dimension to the [0,1] range, the weights are summed according to preset weights (waiting 0.4, transportation 0.2, medical insurance 0.2, expert 0.2) to generate the final edge weight value, and a timestamp is attached to support dynamic cost updates.

[0117] S44: Input the weighted directed referral graph into the target referral path optimizer, use the improved Dijkstra algorithm to calculate the minimum cost referral path from the current institution to the target institution, and output the minimum cost referral path.

[0118] Specifically, the target referral path optimizer calls the improved Dijkstra algorithm, starting from the current institution, traversing the weighted directed referral graph, and outputting the single-hop or multi-hop referral path with the minimum total cost as the minimum cost referral path.

[0119] Furthermore, the algorithm introduces an "institution service capacity decay factor" in the relaxation operation: if the target institution service capacity index is less than 0.5, the path cost is multiplied by a penalty coefficient of 1.5 to avoid recommending overloaded institutions.

[0120] Furthermore, it supports multi-objective optimization: when there are multiple paths with a cost difference of less than 5%, they are sorted a second time according to the strategy of "expert matching degree priority" or "medical insurance compatibility priority", and the Top-3 paths are output for doctors to choose from, thereby improving the flexibility of clinical decision-making.

[0121] S5: Generate and issue referral scheduling instructions based on the minimum cost referral path and treatment level identifier;

[0122] It should be noted that referral pathways need to be translated into executable instructions and that institutional collaboration must be ensured. Vague instructions, lack of role definition, and absence of status tracking can easily lead to disjointed execution. Structured scheduling instructions with permissions, time limits, and sub-tasks must be generated and distributed encrypted with status monitoring to ensure a closed-loop, controllable referral process and clear accountability.

[0123] S51: Analyze the minimum cost referral path and extract the source institution, transit institution, target institution, and the estimated time nodes for each segment of the referral process.

[0124] Specifically, the minimum cost referral path is subjected to structured parsing to identify the path node sequence, extract the source institution (currently affiliated institution), the target institution (final destination institution), and the transit institutions in the path (if cross-regional referral is required), and read the preset referral window time for each segment of the path (such as "transfer completed within T+1 day").

[0125] Furthermore, each path node is labeled with an "institutional role tag" (such as "transferor", "recipient", "transfer coordinator"), and the estimated time node is dynamically adjusted based on the institution's service capacity data. If the target institution's current queue is >3 days, the referral window will be automatically postponed and marked with "resource delay warning".

[0126] Furthermore, a path topology summary is generated, including inter-organizational dependencies, the longest path bottleneck node, and parallel executable segments, and a unique path ID and version number are attached for subsequent instruction distribution and status tracking.

[0127] S52: Generate structured referral scheduling instructions based on the diagnosis and treatment level identifier and path node role. The referral scheduling instructions include patient identifier, referral type, target institution, execution time limit, examination item list and data sharing permissions.

[0128] Specifically, the instruction template engine is invoked to match the preset instruction structure based on the diagnosis and treatment level identifier (such as "specialist hospital follow-up"), and to fill in the patient identifier, target institution, execution time limit (from the path parsing results), and mandatory test items (such as "tumor markers + enhanced CT"), and to set the data sharing permission to "only the target institution can read the current medical record".

[0129] Furthermore, the instruction engine dynamically adds sub-instructions based on the roles of the institutions in the path: for the "transferor", add "upload the latest medical records and image package"; for the "receiver", add "reserve beds and specialist appointments"; and for the "transferor", add "coordinate transfer vehicles and green channels".

[0130] Furthermore, each instruction is given a digital signature and an execution priority label (such as "urgent: response within 2 hours"), and bound to a path ID to ensure that the instructions are traceable, tamper-proof, and scheduled for execution according to priority.

[0131] S53: Send the referral scheduling instruction to the local collaborative terminal of each corresponding institution in the path through the regional collaborative message bus, and start the instruction status listener.

[0132] Specifically, the referral scheduling instructions are distributed to the local collaborative terminals of the corresponding institutions according to their roles through the message middleware (such as RabbitMQ or Kafka) of the regional medical collaboration platform, ensuring that the instructions are delivered within 30 seconds.

[0133] Furthermore, an independent status listener is activated for each instruction to collect real-time status feedback from the organization, such as "read", "confirmed", and "completed", and the execution progress is visualized on the collaborative dashboard; if a node fails to respond within a timeout period, an upgrade reminder is automatically triggered to the organization administrator.

[0134] Furthermore, a two-way encrypted "instruction-receipt" channel is adopted. When an instruction is issued, it carries a temporary session key, and the institution's receipt must be signed with this key to ensure that the instruction transmission process is tamper-proof and non-repudiable. All communication logs are stored in the blockchain for evidence preservation, supporting post-event auditing.

[0135] S6: Collect execution feedback data after the referral scheduling instruction is completed, input the execution feedback data into the federated aggregation server, and update the encrypted local gradient vector generation parameters and model weight parameters.

[0136] It should be noted that the model will degenerate if there is no feedback. The referral execution results contain real-world costs and risk signals, and feedback data needs to be collected to feed back into the federated model, dynamically update the gradient generation parameters and the evaluation / scheduling model weights, forming a closed loop of "execution → learning → optimization" to drive the system to continuously self-evolve.

[0137] S61: After the referral scheduling instruction is marked as "completed", structured execution feedback data is collected from the local collaborative terminals of each participating institution. The execution feedback data includes the actual referral time, the time of receipt and confirmation by the target institution, the completion status of the examination items, the patient satisfaction score, and the referral complication marker.

[0138] Specifically, after receiving "completed" status receipts from all institutions in the path, a feedback collection task is automatically triggered. The task retrieves execution feedback data from the local terminals of each institution via API and stores it in the feedback data pool according to the institution's role.

[0139] Furthermore, integrity checks are performed on the feedback data: if an institution fails to return key fields (such as "actual referral time"), it is marked as "feedback missing" and the collection request is automatically resent; if it fails three times in a row, it is recorded as "abnormal institution collaboration" and included in the institution's credit score.

[0140] Furthermore, each piece of feedback data is appended with a time stamp (collection time + institution GPS coordinates), instruction path ID, and executor's digital ID card to ensure data traceability and tamper-proofing. Standardized feedback records are generated after automatically desensitizing patient identity information.

[0141] S62: Input the execution feedback data into the federated aggregation server, perform feature alignment and gradient backpropagation, and calculate the global feedback gradient used to optimize the local model;

[0142] Specifically, after receiving the execution feedback data, the federated aggregation server maps it to a unified feedback feature space (such as "time consumption → 0~1 normalization" and "satisfaction → 5-point encoding"), and inputs it as a supervision signal into the central evaluation-scheduling joint model to perform backpropagation to calculate the loss gradient and generate a global feedback gradient.

[0143] Furthermore, a multi-task learning framework is adopted to jointly weight the referral time prediction loss, satisfaction classification loss, and complication risk loss to generate a multi-objective global feedback gradient, thereby improving the model's generalization ability.

[0144] Furthermore, importance sampling is performed on the feedback data before backpropagation: samples with "high-risk referral failure" or "excessive time consumption" are given double the weight to accelerate the model's learning efficiency for abnormal events and avoid optimization bias caused by long-tail distribution.

[0145] S63: Distribute the global feedback gradient to the local computing nodes of each participating institution to update the parameters for generating encrypted local gradient vectors and the model weight parameters for the hierarchical evaluation model and the referral pathway optimizer.

[0146] Specifically, the federated aggregation server encrypts the global feedback gradient and distributes it to the local computing nodes of each participating institution through a secure channel. The nodes perform gradient descent updates locally and simultaneously adjust the encrypted local gradient vector generation parameters (such as the local model learning rate and feature selection mask) as well as the model weight parameters of the locally cached hierarchical evaluation model and referral path optimizer.

[0147] Furthermore, an asynchronous update mechanism is adopted: each node performs parameter updates during idle periods (such as at night) to avoid affecting daytime business; after the update, the model version hash is automatically uploaded to the central registry for consistency verification during the next aggregation.

[0148] Furthermore, an "incremental parameter fusion" strategy is introduced: the local node weights and smooths the new and old parameters at a ratio of 0.7:0.3 to prevent parameter oscillation; at the same time, the parameter update trajectory is recorded, and if the evaluation index drops for three consecutive rounds, it will automatically roll back to the previous stable version and trigger manual review.

[0149] Example 3, the third embodiment of the present invention, provides a regional collaborative follow-up system for cancer patients based on hierarchical diagnosis and treatment, comprising:

[0150] The data extraction module is used to acquire raw clinical data from cancer patients and extract encrypted local gradient vectors and institutional identifiers.

[0151] The federated aggregation module is used to input encrypted local gradient vectors into the federated aggregation server and output a global risk feature vector.

[0152] The tiered assessment module is used to call the target tiered assessment model based on the global risk feature vector and real-time regional resource parameters, and output the diagnosis and treatment level identifier.

[0153] The route optimization module is used to call the target referral route optimizer based on the diagnosis and treatment level identifier and real-time medical institution service capacity data, and output the referral route with minimum cost.

[0154] The instruction scheduling module is used to generate and issue referral scheduling instructions based on the minimum cost referral path and treatment level identifier.

[0155] The feedback update module is used to collect execution feedback data after the referral scheduling instruction is completed, input the execution feedback data into the federated aggregation server, and update the encrypted local gradient vector generation parameters and model weight parameters.

[0156] Example 4, refer to Figure 3 This is the fourth embodiment of the present invention, which differs from the previous three embodiments in that: if the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0157] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0158] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0159] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0160] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment, characterized in that: include, Obtain raw clinical data from cancer patients and extract encrypted local gradient vectors and institutional identifiers; The encrypted local gradient vector is input into the federated aggregation server, which outputs a global risk feature vector, including: Receive multiple encrypted local gradient vectors and their corresponding organization identifiers, and batch aggregate them according to a preset time window; Perform homomorphic addition aggregation on all the encrypted local gradient vectors within the same batch to generate an encrypted global gradient vector; The encrypted global gradient vector is sent to a private key decryptor deployed in a secure execution environment to perform private key decryption and obtain the plaintext global gradient vector. The plaintext global gradient vector is input into the central risk feature extraction model, and the global risk feature vector is output after forward propagation. Based on the global risk feature vector and real-time regional resource parameters, the target hierarchical assessment model is invoked to output the diagnosis and treatment level identifier; Based on the diagnosis and treatment level identifier and real-time medical institution service capacity data, the target referral path optimizer is invoked to output the minimum cost referral path. Based on the minimum cost referral path and the treatment level identifier, a referral scheduling instruction is generated and issued; The execution feedback data after the referral scheduling instruction is completed is collected, and the execution feedback data is input into the federated aggregation server to update the encrypted local gradient vector generation parameters and model weight parameters.

2. The method for regional collaborative follow-up of cancer patients based on hierarchical diagnosis and treatment as described in claim 1, characterized in that: The extraction of encrypted local gradient vectors and mechanism identifiers includes: Structured clinical data is collected from the electronic health records of local medical institutions. The structured clinical data includes TNM staging, ECOG score, treatment plan, laboratory indicators and follow-up adherence records. The structured clinical data is standardized and mapped to a numerical feature vector within a unified feature space. The numerical feature vector is input into the local pre-trained risk prediction model, the model gradient is calculated, and homomorphic encryption is performed on the model gradient to generate the encrypted local gradient vector. Extract the unique code of the local medical institution in the regional collaborative network and use it as the institution identifier.

3. The method for regional collaborative follow-up of cancer patients based on hierarchical diagnosis and treatment as described in claim 2, characterized in that: The target hierarchical evaluation model for invocation includes: Obtain real-time regional resource parameters, including bed occupancy rate, remaining number of specialist appointment slots, equipment queuing time, and doctor saturation. Based on the dimensional distribution of the global risk feature vector and the load range of the real-time regional resource parameters, the target hierarchical assessment model is matched from the hierarchical assessment model set. The global risk feature vector is input into the target classification and assessment model, and the original risk score is output. Based on the original risk score and the preset grading threshold table, the diagnosis and treatment level identifier is mapped and output.

4. The regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment as described in claim 3, characterized in that: The invocation of the target referral path optimizer includes: Acquire real-time medical institution service capacity data, which includes the number of available beds, expert scheduling status, equipment idle time, medical insurance compatibility level, and geographical accessibility score; Based on the service efficiency range of the treatment level identifier and the service capability data, the target referral path optimizer is matched from the referral path optimizer set; Construct a weighted directed referral graph, where nodes represent medical institutions and edge weights are the comprehensive referral costs calculated by weighting waiting time, transportation costs, medical insurance compatibility, and expert matching. The weighted directed referral graph is input into the target referral path optimizer, and the improved Dijkstra algorithm is used to output the minimum cost referral path.

5. A method for regional collaborative follow-up of cancer patients based on hierarchical diagnosis and treatment as described in claim 4, characterized in that: The generation and issuance of referral scheduling instructions includes: The minimum cost referral path is analyzed to extract the source institution, target institution, transit institution, and the estimated time nodes for each segment of the referral; Based on the diagnosis and treatment level identifier and path node role, a structured referral scheduling instruction is generated. The referral scheduling instruction includes the patient identifier, target institution, execution time limit, examination item list and data sharing permissions. The referral scheduling instruction is sent to the local collaborative terminal of the corresponding institution through the regional collaborative message bus, and the instruction status listener is started to collect execution receipts.

6. The method for regional collaborative follow-up of cancer patients based on hierarchical diagnosis and treatment as described in claim 5, characterized in that: The updating of the encrypted local gradient vector generation parameters and model weight parameters includes: After the referral scheduling instruction is marked as "completed", structured execution feedback data is collected. The execution feedback data includes the actual referral time, the receipt confirmation time, the examination completion status, the satisfaction score, and the complication marker. The execution feedback data is input into the federated aggregation server to perform feature alignment and backpropagation, generating a global feedback gradient. The global feedback gradient is encrypted and distributed to the local computing nodes of each participating institution to update the encrypted local gradient vector generation parameters, the weight parameters of the hierarchical evaluation model, and the weight parameters of the referral path optimizer.

7. A regional collaborative follow-up system for cancer patients based on hierarchical diagnosis and treatment, employing the regional collaborative follow-up method for cancer patients based on hierarchical diagnosis and treatment as described in any one of claims 1 to 6, characterized in that, include: The data extraction module is used to acquire raw clinical data from cancer patients and extract encrypted local gradient vectors and institutional identifiers. The federated aggregation module is used to input the encrypted local gradient vector into the federated aggregation server and output a global risk feature vector, including: receiving multiple encrypted local gradient vectors and their corresponding institution identifiers, and performing batch aggregation according to a preset time window; Perform homomorphic addition aggregation on all the encrypted local gradient vectors within the same batch to generate an encrypted global gradient vector; The encrypted global gradient vector is sent to a private key decryptor deployed in a secure execution environment to perform private key decryption and obtain the plaintext global gradient vector. The plaintext global gradient vector is input into the central risk feature extraction model, and the global risk feature vector is output after forward propagation. The hierarchical assessment module is used to call the target hierarchical assessment model based on the global risk feature vector and real-time regional resource parameters, and output the diagnosis and treatment level identifier. The path optimization module is used to call the target referral path optimizer based on the diagnosis and treatment level identifier and real-time medical institution service capacity data, and output the minimum cost referral path; The instruction scheduling module is used to generate and issue referral scheduling instructions based on the minimum cost referral path and the treatment level identifier. The feedback update module is used to collect execution feedback data after the referral scheduling instruction is completed, input the execution feedback data into the federated aggregation server, and update the encrypted local gradient vector generation parameters and model weight parameters.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the regional collaborative follow-up method for tumor patients based on hierarchical diagnosis and treatment, as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of any one of the following claims: a regional collaborative follow-up method for tumor patients based on hierarchical diagnosis and treatment.

Images