Follow-up visit plan generation method and device, equipment and medium
The follow-up plan generation model, which uses multi-dimensional medical data for supervised fine-tuning, solves the problem of lack of personalization and rationality in existing follow-up plans, and achieves safe and rational follow-up plan generation.
Patent Information
- Application Number
- CN202511783970.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-02-27
AI Technical Summary
Existing follow-up management solutions cannot effectively integrate complex unstructured medical data, resulting in follow-up plans that lack personalization and targeting, and are of poor rationality and reliability.
By acquiring multi-dimensional medical data of the target subjects, and using positive and negative sample sets to perform supervised fine-tuning of the training model, a follow-up plan generation model is generated, ensuring the safety and rationality of the generated follow-up plan.
The resulting follow-up plans are more targeted and personalized, ensuring their medical safety and rationality, and are aligned with the level of top medical experts.
Smart Images
Figure CN121583441A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of medical data processing technology, and in particular to a method, apparatus, device, and medium for generating follow-up plans. Background Technology
[0002] With the development of medical informatization, patient follow-up management has become a key aspect of managing the entire disease lifecycle and improving long-term efficacy. Currently, relevant solutions typically employ standardized follow-up templates based on clinical guidelines, rule-based expert systems, or models to generate follow-up plans for target individuals.
[0003] However, these methods can only process simple, structured medical data and do not further combine complex, unstructured medical data to generate follow-up plans. As a result, the final follow-up plans are not targeted enough, lack personalization, and are difficult to achieve precise follow-up with "one plan for one person". Moreover, the rationality and reliability of the generated follow-up plans are also poor. Summary of the Invention
[0004] This disclosure provides a method, apparatus, device, and medium for generating follow-up plans; it can generate more targeted follow-up plans and ensure the safety and rationality of the generated follow-up plans.
[0005] The technical solution disclosed herein is implemented as follows: Firstly, this disclosure provides a method for generating a follow-up plan, including: Acquire multi-dimensional medical data of the target object; Based on multi-dimensional medical data, a follow-up plan is generated for the target subjects through a follow-up plan generation model. The follow-up plan generation model is obtained by supervised fine-tuning of the model to be trained using a positive sample set and a negative sample set. The positive sample set includes at least one positive sample, and the negative sample set includes at least one negative sample. The positive sample is a verified and reasonable first training follow-up plan, and the negative sample is a verified and unreasonable second training follow-up plan.
[0006] Secondly, this disclosure provides an apparatus for generating a follow-up plan, comprising: The acquisition module is used to acquire multi-dimensional medical data of the target object; The generation module is used to generate a target follow-up plan for the target object based on multi-dimensional medical data and through a follow-up plan generation model. The follow-up plan generation model is obtained by supervised fine-tuning of the model to be trained using a positive sample set and a negative sample set. The positive sample set includes at least one positive sample, and the negative sample set includes at least one negative sample. The positive sample is a verified and reasonable first training follow-up plan, and the negative sample is a verified and unreasonable second training follow-up plan.
[0007] Thirdly, this disclosure provides an electronic device comprising: a processor; and a memory storing a computer program; wherein the processor is configured to, when executing the computer program, implement the method for generating a follow-up plan as described in the first aspect.
[0008] Fourthly, this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for generating a follow-up plan as described in the first aspect.
[0009] This disclosure provides a method, apparatus, device, and medium for generating follow-up plans. Follow-up plans can be generated based on multi-dimensional medical data of the target individual using a follow-up plan generation model, resulting in more targeted follow-up plans. Furthermore, the follow-up plan generation model is obtained through supervised fine-tuning of the training model using positive and negative sample sets, ensuring the safety and rationality of the generated target follow-up plan. Attached Figure Description
[0010] Figure 1 This is a schematic diagram of the components of a follow-up plan generation system provided in this disclosure.
[0011] Figure 2 This is a flowchart illustrating a method for generating a follow-up plan provided in this disclosure.
[0012] Figure 3 This is a flowchart illustrating the process of generating a target follow-up plan based on multi-dimensional medical data, as provided in this disclosure.
[0013] Figure 4 This is a schematic diagram of the architecture of a follow-up plan generation model provided in this disclosure.
[0014] Figure 5 This is a flowchart illustrating an output target follow-up plan provided in this disclosure.
[0015] Figure 6 A schematic diagram illustrating the process of oversight and fine-tuning provided in this disclosure.
[0016] Figure 7 This is a schematic diagram illustrating an optimization process for a follow-up plan generation model provided in this disclosure.
[0017] Figure 8 This is a schematic diagram illustrating the optimization process of another follow-up plan generation model provided in this disclosure.
[0018] Figure 9 This is a schematic diagram of the components of a follow-up plan generation device provided in this disclosure.
[0019] Figure 10 A schematic diagram of the composition of another follow-up plan generation device provided in this disclosure. Detailed Implementation
[0020] The technical solutions in this disclosure will now be clearly and completely described with reference to the accompanying drawings.
[0021] Figure 1 This is a schematic diagram illustrating the composition of an exemplary follow-up plan generation system applicable to the technical solutions of this disclosure. See also... Figure 1 The follow-up plan generation system 100 may include a server 110, a terminal 120, a data subsystem 130, and a network 140. In some examples, at least one component of the follow-up plan generation system 100, such as server 110, terminal 120, or data subsystem 130, may communicate with at least one other component of the follow-up plan generation system 100 via network 140. For example, server 110 may obtain multi-dimensional medical data (including structured and unstructured medical data) of the target object from data subsystem 130 via network 140. In this disclosure, the target object refers to an object that has been treated or has not been treated; for example, the target object may be a patient or a specific part of the patient's body, such as the head, chest, abdomen, or a combination thereof.
[0022] exist Figure 1 In this context, server 110 can be a server cluster deployed in the cloud or a local server deployed within a medical institution, such as a hospital. Server 110 is the implementing entity for the technical solution disclosed herein, and a follow-up plan generation model is deployed within it.
[0023] In this disclosure, such as Figure 1 As shown, server 110 can be implemented as an electronic device including at least one processor 112, memory 114 and communication connector 116, wherein communication connector 116 is connected to network 140 so that server 110 can communicate with other components in follow-up plan generation system 100, such as terminal 120 and data subsystem 130.
[0024] In this disclosure, memory 114 may include a mass storage device, a removable storage device, a volatile read-write memory, a read-only memory (ROM), or a combination thereof. Memory 114 may store data, instructions, and / or any other information. For example, memory 114 may store unstructured medical data and structured medical data transmitted by data subsystem 130 via network 140. As another example, memory 114 may store data and / or instructions executable by processor 132, which, when executed by processor 132, enable the implementation of the follow-up plan generation method provided in this disclosure.
[0025] In some examples, server 110 can implement the follow-up plan generation method by invoking an internally deployed follow-up plan generation model whose trained parameters are also stored in memory 114.
[0026] Terminal 120 may include a first terminal 121 deployed on the doctor's side and a second terminal 122 deployed on the target's side. The first terminal 121 and the second terminal 122 may be at least one of the following devices: smartphone, smartwatch, desktop computer, laptop, virtual reality terminal, augmented reality terminal, wireless terminal, and laptop computer.
[0027] In this disclosure, the first terminal 121 runs client software. Doctors use the client software on the first terminal 121 to view, review, adjust, and confirm the follow-up plan generated by the server 110 or by the follow-up plan generation model. After the doctor adjusts the follow-up plan on the first terminal 121, the first terminal 121 transmits the plan adjustment data back to the server 110 via network 140 to provide feedback and optimization to the follow-up plan generation model.
[0028] In this disclosure, client software also runs on the second terminal 122. The target subject can view the follow-up plan pushed by the server 110 through the client software on the second terminal 122, and can also upload efficacy data during the execution of the follow-up plan, such as blood pressure, blood glucose, and blood routine values, to the server 110 through the second terminal 122.
[0029] The data subsystem 130 may include a Picture Archiving and Communication System (PACS), a Radiology Information System (RIS), a Laboratory Information System (LIS), a Hospital Information System (HIS), or an Electronic Health Records (EHR) database. This data subsystem 130 stores unstructured medical data such as medical images, laboratory reports, pathology reports, or discharge summaries of the target individual, as well as structured medical data such as demographic information, examination and testing indicators, vital signs, and medication records of the target individual. Medical images may include, for example, computed tomography (CT) images, positron emission tomography (PET) images, digital radiography (DR) images, magnetic resonance imaging (MRI) images, and four-dimensional (4D) images.
[0030] In practice, after a target patient is discharged from the hospital or completes a medical treatment, server 110 can actively or passively obtain multi-dimensional medical data of the target patient from data subsystem 130 via network 140. Based on this multi-dimensional medical data, server 110 generates a follow-up plan for the target patient through a follow-up plan generation model, and pushes the follow-up plan to the first terminal 121 and / or the second terminal 122 via network 140 for review or access by doctors and / or the target patient.
[0031] Figure 2 This is a flowchart illustrating a method for generating a follow-up plan provided in this disclosure.
[0032] See Figure 2 In step S210, multi-dimensional medical data of the target object is obtained.
[0033] Specifically, when it is determined that a specific target subject will be followed up, it is necessary to first obtain multi-dimensional medical data generated during the treatment or diagnosis process. This multi-dimensional medical data comes from medical imaging examinations, blood routine tests, and physiological data tests, such as blood glucose levels, blood pressure levels, and electrocardiograms. It can also include medical imaging data, laboratory report data, medication record data, pathology report data, and discharge summary data.
[0034] See Figure 2 In step S220, based on multi-dimensional medical data, a target follow-up plan for the target object is generated through a follow-up plan generation model.
[0035] Specifically, after obtaining multi-dimensional medical data, the multi-dimensional medical data can be input into the follow-up plan generation model. The follow-up plan generation model can perform time-series data analysis and multi-modal understanding and reasoning on the multi-dimensional medical data, and generate a structured follow-up plan that matches the target object based on the time-series analysis results and reasoning results.
[0036] The follow-up plan generation model is obtained through supervised fine-tuning of the training model using positive and negative sample sets. The positive sample set includes at least one positive sample, which refers to a validated and reasonable first training follow-up plan. This can be understood as a follow-up plan that has been confirmed by medical experts and conforms to clinical guidelines and best practices. For example, for target subjects receiving adjuvant trastuzumab therapy after surgery for human epidermal growth factor receptor-2 (Her-2+) breast cancer, the follow-up plan might be: intravenous infusion of trastuzumab every 3 weeks for 1 year; cardiac function monitoring via echocardiography every 3 months during treatment; and breast ultrasound / mammography every 6 months. This follow-up plan is a positive sample, confirmed by medical experts, conforms to clinical guidelines and best practices, and can serve as a positive sample. The negative sample set includes at least one negative sample, which is a validated but unreasonable second training follow-up plan. This can be understood as follows: negative samples can be unsafe or unreasonable follow-up plans constructed by medical experts for model training, or they can be erroneous, unsafe, or unreasonable follow-up plans extracted from error cases in hospitals. For example, for target subjects with epidermal growth factor receptor (EGFR) mutated lung cancer, a follow-up plan of performing a whole-body PET-CT scan every two days would lead to excessive radiation exposure and is a medically unsafe follow-up plan. For target subjects with Her-2+ breast cancer, a follow-up plan using paclitaxel and carboplatin as adjuvant therapy is a non-standard follow-up plan and is medically unreasonable. For target subjects with Her-2 negative breast cancer, a follow-up plan using Her-2 targeted drugs involves incorrect medication that is inconsistent with the characteristics of the target subjects, and is therefore an erroneous follow-up plan.
[0037] The model to be trained can adopt a multimodal base model that integrates medical knowledge. By using positive and negative sample sets for supervised fine-tuning, the base model can absorb the experience of positive samples and actively avoid the problems of negative samples, ensuring the safety and rationality of the generated follow-up plan in medicine, and aligning its decision-making logic with the level of top medical experts.
[0038] exist Figure 2 The technical solution shown allows for the generation of follow-up plans based on multi-dimensional medical data of the target individual using a follow-up plan generation model, resulting in more targeted follow-up plans. Furthermore, the follow-up plan generation model is obtained through supervised fine-tuning of the training model using both positive and negative sample sets, ensuring the safety and rationality of the generated follow-up plans.
[0039] In some embodiments, the follow-up plan includes at least: follow-up items, follow-up timeline, follow-up frequency, and follow-up risks.
[0040] For example, the follow-up plan generation model can output the target follow-up plan using a structured JSON data format. For instance, for a patient with EGFR-mutant lung adenocarcinoma, the generated target follow-up plan could be: chest CT scans every 8 weeks post-surgery, circulating tumor DNA (ctDNA) testing every 4 weeks, and a 15% probability of disease progression within the next 30 days. A JSON example of this target follow-up plan is shown below: { “patient_id”: "001", “diagnosis”: "EGFR-mutated lung adenocarcinoma", “follow_up_plan”:[{"item": "chest CT", "frequency": "postoperative" "every 8 weeks"}, {"item": "ctDNA monitoring", "frequency": "every 4 weeks"}], “risk_prediction”: { "progression_30d_risk": 0.15 } } In the JSON example above, "item" represents the follow-up items, including chest CT scans and ctDNA monitoring. "frequency" represents the follow-up timeline and / or follow-up frequency, with the timeline including "post-operative" and the frequency including "every 8 weeks" and "every 4 weeks". "risk_prediction" represents the follow-up risk, where {"progression_30d_risk": 0.15} indicates a 15% probability of disease progression within the next 30 days.
[0041] The structured target follow-up plan output above not only provides follow-up items, follow-up frequency, and follow-up timeline to inform the target subjects of the follow-up items, follow-up time, and follow-up frequency that should be done, but also provides follow-up risk prediction, providing comprehensive data support for clinical decision-making.
[0042] Figure 3 This is a flowchart illustrating the process of generating a target follow-up plan for a target subject based on multi-dimensional medical data.
[0043] See Figure 3 In step S222, the multi-dimensional medical data is standardized to generate an embedded vector sequence that can be processed by the follow-up plan generation model.
[0044] Specifically, multi-dimensional medical data includes data of different types, sources, and formats, which follow-up plan generation models cannot directly process. Therefore, preprocessing of multi-dimensional medical data is necessary. For example, structuring the multi-dimensional medical data can be performed to convert it into the target data format, and abnormal and redundant information can be removed. Then, standardization processing can be performed on the multi-dimensional medical data to form an embedded vector sequence that the follow-up plan generation model can process. This standardization process can be a multi-modal data fusion encoding process.
[0045] In some examples, multidimensional medical data may include structured and unstructured medical data, and the embedded vector sequence may include a first vector sequence and a second vector sequence.
[0046] Specifically, structured medical data refers to data defined in a standardized format that can be directly stored and quantitatively analyzed in a database. Examples include: demographic information (e.g., age, gender), examination and testing indicators (e.g., complete blood counts, liver function tests, tumor markers), vital signs, and medication records (e.g., drug name, dosage, and frequency). This structured medical data is typically timestamped, possessing a time attribute, and follow-up plan generation models can analyze its temporal patterns.
[0047] Unstructured medical data refers to data that has not been defined in a standardized format and is difficult to quantify directly. Examples include medical images (such as CT images, MRI images, and PET-CT images), corresponding test reports, pathology reports (such as immunohistochemistry results), and discharge summary texts. By extracting feature vectors from these unstructured medical data, the follow-up plan generation model can perform multimodal inference based on these feature vectors.
[0048] It should be noted that this disclosure does not specifically limit the types or sources of the various data included in the multidimensional medical data. For example, multidimensional medical data may include text-based data (such as test reports) or image-based data (such as medical images). Furthermore, the data in multidimensional medical data may originate from electronic medical record systems used to manage patient records, or from medical imaging equipment, wearable health monitoring devices, etc. In other words, multidimensional medical data may include multimodal data and multi-source heterogeneous data.
[0049] Based on this, step S222 may include the following steps: The structured medical data is time-aligned and then converted into a first vector sequence.
[0050] Feature extraction is performed on unstructured medical data, and the processed unstructured medical data is converted into a second vector sequence.
[0051] Specifically, since structured medical data has a time attribute, in order to enable the follow-up plan generation model to analyze its time series patterns, the structured medical data can be sorted, interpolated, and time aligned according to timestamps. This will match the discrete event data and the continuous monitoring event data contained in the structured medical data to the same time axis. Then, the processed structured medical data is converted into a time series to represent it, namely the first vector sequence.
[0052] Unstructured medical data can be divided into at least two categories: unstructured image data and unstructured text data. To extract key medical features from unstructured medical data, Natural Language Processing (NLP) techniques can be used to extract features from the unstructured text data, and visual models, such as Convolutional Neural Networks (CNNs), can be used to extract features from the unstructured image data. These features can then be fused and concatenated to transform the data into a second vector sequence.
[0053] In detail, the process of processing unstructured text data, such as pathology reports and discharge summaries, using natural language processing techniques can include: First, key medical concepts in the text data are identified and treated as entities, such as diseases ("lung adenocarcinoma"), drugs ("trastuzumab"), surgeries ("radical surgery"), anatomical sites, etc.
[0054] Next, the relationships between the entities are further extracted.
[0055] Finally, the identified entities and relationships are converted into high-dimensional text embedding vectors.
[0056] In detail, for unstructured image data, such as CT images, MRI images, and PET-CT images, a pre-trained visual model is used to extract deep features from the image data (such as tumor volume, metabolic activity value, tumor shrinkage ratio, etc.), and these features are converted into image embedding vectors.
[0057] After obtaining the text embedding vector and the image embedding vector, the text embedding vector and the image embedding vector can be fused by concatenation or cross-attention mechanism to generate a second vector sequence, which represents an unstructured medical data snapshot of the target object at a point in time.
[0058] See Figure 3 In step S224, the embedded vector sequence is input into the follow-up plan generation model so that the follow-up plan generation model outputs the target follow-up plan.
[0059] Specifically Figure 4 This is a schematic diagram of the architecture of a follow-up plan generation model provided in this disclosure. Figure 4 In the follow-up plan generation model, a Temporal Convolutional Network (TCN) 410, an inference network 420, and an output layer 430 are included. Based on... Figure 4 The follow-up plan generation model shown is as follows: Figure 5 As shown, step S224 specifically includes steps S2241 to S2243.
[0060] See Figure 5 In step S2241, based on the first vector sequence, time series analysis is performed through the temporal convolutional network 410 to generate time series analysis results.
[0061] Specifically, TCN 410 can capture the dependencies and patterns of change in the first vector sequence over time. For example, regarding changes in blood routine tests, the temporal convolutional network 410 in the follow-up plan generation model, after learning from a large amount of chemotherapy patient data that "in patients receiving the FOLFOX chemotherapy regimen, the absolute neutrophil count (ANC) typically reaches its lowest point within 7-14 days," allows TCN 410 to perform prospective predictions and generate time-series analysis results when it receives the first vector sequence of a new patient. This first vector sequence contains the patient's current FOLFOX regimen and blood routine indicators. The time-series analysis results include quantitative risk predictions, such as an 85% risk of the patient developing grade 3 neutropenia on day 10 after chemotherapy.
[0062] Understandably, this time series analysis result is a key basis for the follow-up risk in the subsequent generation of the target follow-up plan.
[0063] See Figure 5 In step S2242, based on the second vector sequence, multimodal reasoning is performed through the inference network 420 to generate multimodal reasoning results.
[0064] Specifically, in addition to the time-series analysis performed by TCN 410, the inference network 420 performs multimodal inference on the second vector sequence containing text embedding vectors and image embedding vectors to generate multimodal inference results. It is understood that in practical implementation, the follow-up plan generation model can execute steps S2241 and S2242 simultaneously, or in a specific order. This disclosure does not specifically limit the execution order of steps S2241 and S2242.
[0065] For example, the inference network 420 can employ a multimodal Transformer architecture that integrates a medical knowledge graph. This architecture, through a cross-attention mechanism, is capable of deeply understanding and fusing the relationships between different modalities of data. For instance, the inference network 420 simultaneously processes text embedding vectors and image embedding vectors from a second vector sequence and cross-validates these two embedding vectors to generate more reliable multimodal inference results regarding efficacy. For example, the image embedding vector might include a comparison of the metabolic tumor volume (MTV) extracted from the current PET-CT image with the MTV extracted from the previous PET-CT image, calculating a "45% reduction"; the text embedding vector might include a conclusion of partial response (PR). The inference network 420 generates multimodal inference results by cross-validating these two embedding vectors.
[0066] See Figure 5 In step S2243, based on the time series analysis results and the multimodal reasoning results, the target follow-up plan is generated through the output layer 430.
[0067] Specifically, the follow-up plan generation model can first form a joint feature vector by concatenating the time series analysis results from TCN 410 and the multimodal inference results from inference network 420; then, it can reduce the dimensionality of the concatenated joint feature vector to prevent overfitting; finally, it can generate the target follow-up plan by passing the dimensionality-reduced joint feature vector through output layer 430.
[0068] In this disclosure, the follow-up plan generation model is obtained by supervised fine-tuning (SFT) of the model to be trained using positive and negative sample sets. This supervised fine-tuning process includes a contrastive learning training process and a safety boundary training process. Figure 6 A schematic diagram illustrating the process of oversight and fine-tuning provided in this disclosure. Figure 6 In the image, the left side of the model to be trained 600 shows the contrastive learning training process, and the right side of the model to be trained shows the safety boundary training process.
[0069] See Figure 6 In the comparative learning training process, the training model 600 is trained 630 based on the positive sample set 610 and the negative sample set 640 to narrow the vector distance between the training follow-up plan generated by the training model 600 and the first training follow-up plan, and widen the vector distance between the training follow-up plan and the second training follow-up plan, until the comparative loss function meets the first preset condition, and the follow-up plan generation model is obtained.
[0070] Specifically, during the comparative learning training process 630, the model to be trained 600 generates a training follow-up plan based on the embedding vector sequence. For example, for Her-2+ breast cancer targets, a training follow-up plan is generated based on the embedding vector sequence, and this training follow-up plan is compared with positive and negative samples in the feature space. For example, the first training follow-up plan as a positive sample uses trastuzumab, and the second training follow-up plan as a negative sample uses paclitaxel and carboplatin.
[0071] The goal of the contrastive learning training process is to narrow the vector distance between the training follow-up plan generated by the model to be trained and the first training follow-up plan, and to widen the vector distance between the training follow-up plan and the second training follow-up plan. This disclosure employs a contrastive loss function to measure spatial distance to achieve this goal. Specifically, in the high-level feature space, let the training follow-up plan generated by the model to be trained be A, the first training follow-up plan be P, and the second training follow-up plan be N. The vector distance between the training follow-up plan generated by the model to be trained and the first training follow-up plan is... The vector distance between the training follow-up plan generated by the model to be trained and the second training follow-up plan is: The distance between the two vectors can be either Euclidean or cosine distance. Compare the loss functions. As shown in the following formula:
[0072] in, This represents a preset margin used to ensure that the distance between positive and negative samples is at least greater than a certain value. .
[0073] During the contrastive learning training process, the model parameters in the model to be trained are continuously optimized through the backpropagation algorithm until the contrastive loss function is optimized. The first preset condition is met, for example, The model converges to below a set threshold, or is trained for a preset number of rounds, thus obtaining the follow-up plan generation model.
[0074] See Figure 6 During the safety boundary training process, the training model 600 is trained on the safety boundary based on the negative sample set 640. If the similarity between the training follow-up plan generated by the training model 600 and the second training follow-up plan exceeds a threshold, the training model 600 is penalized by the risk loss function until the risk loss function meets the second preset condition, and the follow-up plan generation model is obtained.
[0075] Specifically, the purpose of training the safety boundary for 650 is to enable the follow-up plan generation model to proactively avoid unsafe, unreasonable, and erroneous follow-up plans, thereby ensuring the medical safety and rationality of the generated follow-up plans. The risk loss function is an additional loss term besides the contrastive loss function, calculated for negative samples.
[0076] In detail, during the safety boundary training process 650, when the training follow-up plan generated by the model to be trained 600 is judged by a discriminator to have a similarity exceeding a set threshold with a second training follow-up plan in the negative sample set, the risk loss function increases sharply. For example, the similarity is measured by cosine similarity in high-dimensional space, with a set threshold of 0.8. If the training follow-up plan generated by the model to be trained 600 for EGFR-mutant lung cancer targets PET-CT scans every 2 days, and the discriminator judges that this training follow-up plan has a cosine similarity exceeding 0.8 with a second training follow-up plan in the negative sample library concerning "radiation overdose" in high-dimensional space, the risk loss function will increase sharply. In this disclosure, this risk loss function can be a binary cross-entropy loss.
[0077] As the risk loss function increases dramatically, the huge risk loss value will generate a huge penalty gradient during backpropagation, forcing the model to be trained to adjust its internal parameters to avoid generating suggestions similar to known errors.
[0078] In this way, safety boundary training is continuously performed until the risk loss function meets the second preset condition. For example, when the similarity probability between the training follow-up plan generated by the training model 600 and the second training follow-up plan in the negative samples is lower than a set threshold, the follow-up plan generation model can be obtained.
[0079] Through the dual SFT training of the comparative learning training process and the safety boundary training process, the follow-up plan generation model can imitate excellent follow-up plans while avoiding unsafe and unreasonable follow-up plans. This enables the follow-up plan generated by the model to achieve both optimization and deterioration in medicine, thereby improving the safety and reliability of the generated follow-up plan in medicine.
[0080] In this disclosure, in accordance with Figure 6 After obtaining the follow-up plan generation model through the supervised fine-tuning process shown, in order to avoid the defects of knowledge solidification and inability to continuously evolve after the model is trained, the technical solution disclosed herein will also perform feedback optimization on the follow-up plan generation model. Figure 7 and Figure 8 This is a schematic diagram illustrating the process of feedback optimization of the follow-up plan generation model provided in this disclosure.
[0081] See Figure 7 In step S230, feedback data is obtained when the target object performs the target follow-up plan.
[0082] Specifically, such as Figure 8 As shown, once the follow-up plan generation model 800 is trained and deployed, it can continuously collect feedback from clinical practice regarding the target subjects' execution of the target follow-up plan 810. This feedback data may include at least one of the following: Adjustment data for the target follow-up plan, such as Figure 8 As shown in 830, the plan adjustment data refers to the record of physicians’ actual adjustments, adoptions, or rejections of the target follow-up plan 810; Treatment efficacy data during the target follow-up program for the target subjects, such as Figure 8 As shown in 870, the efficacy data refers to the long-term tracking of the target subjects' efficacy, such as progression-free survival (PFS), overall survival (OS), and incidence of adverse reactions.
[0083] And, updated medical knowledge data, such as Figure 8 As shown in 872, the updated medical knowledge data refers to the updated content of authoritative clinical guidelines in the medical field, such as the clinical practice guidelines of the National Comprehensive Cancer Network (NCCN) and the Chinese Society of Clinical Oncology (CSCO).
[0084] See Figure 7 In step S240, the follow-up plan generation model is optimized based on the feedback data to update the follow-up plan generation model.
[0085] Specifically, by continuously optimizing the follow-up plan generation model 800 through feedback data, the model 800 acquires the ability to continuously learn and self-optimize, preventing its decision-making capabilities from deteriorating over time. Based on Figure 8 The three types of feedback data shown in this disclosure illustrate the process of optimizing the follow-up plan generation model 800 through the following three examples to update the follow-up plan generation model.
[0086] In some examples, when the feedback data includes plan adjustment data 830, new positive samples and / or new negative samples 850 are generated based on the plan adjustment data 830; and the follow-up plan generation model 800 is incrementally trained 860 using the new positive samples and / or new negative samples 850 to update the follow-up plan generation model.
[0087] In the above example, in detail, combined with Figure 8 As shown in the left feedback loop, when a doctor adjusts the target follow-up plan 810, the difference comparison algorithm 840 is triggered. This difference comparison algorithm 840 is used to compare the target follow-up plan 810 and the doctor's modified follow-up plan, and identify the differences between the two. For example, the algorithm can pinpoint the "fields" and "content" modified by the doctor based on the textual or structural differences between the two.
[0088] For example, the follow-up plan generation model 800 might generate a target follow-up plan 810 for a breast cancer patient that omits the "echocardiography" item. The doctor could manually add this item, possibly noting the reason as "monitoring cardiac function is necessary due to the use of anthracyclines." The difference comparison algorithm 840 would accurately identify this addition and generate new positive and / or new negative samples 850. Specifically, the doctor-modified follow-up plan, which includes the "echocardiography" item, would be stored as a new high-quality positive sample in the positive sample set. Furthermore, the original follow-up plan that omitted the key monitoring item, i.e., the target follow-up plan 810 generated by the follow-up plan generation model 800, would be stored as a new negative sample in the negative sample set.
[0089] Subsequently, the follow-up plan generation model 800 will be incrementally trained 860 periodically, for example, every night, using this batch of new positive / negative samples 850. In the above example, this incremental training will enhance the follow-up plan generation model 800's understanding of the association between "anthracycline drugs" and "cardiotoxicity monitoring".
[0090] pass Figure 8 The feedback loop shown on the left in the diagram allows the follow-up plan generation model 800 to learn from doctors' daily clinical experience in real time; therefore, this feedback loop is also called the real-time feedback loop. Through the real-time feedback loop, the follow-up plan generation model 800 can perform self-correction and evolution in real time, and its decision-making ability will continue to optimize over time.
[0091] In some examples, when the feedback data includes efficacy data, the rationality of the target follow-up plan is determined based on the efficacy data; if the target follow-up plan is determined to be rational (e.g., a significant prolongation of PFS is found), the target follow-up plan is used as a new positive sample; if the target follow-up plan is determined to be unreasonable (e.g., a significant increase in adverse reaction rate is found), the target follow-up plan is used as a negative sample; and reinforcement learning is performed on the follow-up plan generation model using the new positive or new negative samples to update the follow-up plan generation model 800.
[0092] For the above example, specifically, in combination Figure 8 As shown in the right-hand feedback loop, efficacy data 870 can be tracked over a long period during the execution of the target follow-up plan 810, such as obtaining efficacy data like progression-free survival (PFS), overall survival (OS), and adverse event rates, and correlating them with a clinical outcome database. Next, efficacy analysis 875 is performed on the efficacy data to determine the rationality of the target follow-up plan 810, and to generate new positive or negative samples, as well as reinforcement learning reward signals 880 (including positive and negative reward signals).
[0093] For example, for target subjects with stage II colorectal cancer, when the target follow-up plan 810 was "ctDNA monitoring every 4 weeks," their progression-free survival (PFS) was significantly extended by 4.2 months compared to the group receiving the standard follow-up plan of "ctDNA monitoring every 12 weeks." Based on this conclusion, the target follow-up plan 810 can be deemed reasonable, and this target follow-up plan 810 can be used as a new positive sample to generate a positive reward signal.
[0094] Conversely, if efficacy analysis 875 reveals that the target follow-up plan 810 leads to a shortened PFS or an increased adverse reaction rate, then the target follow-up plan 810 can be determined to be unreasonable. The target follow-up plan 810 is then used as a new negative sample, and a negative reward signal is generated.
[0095] Based on these new positive or negative samples and the corresponding reinforcement learning reward signals 880, the follow-up plan generation model 800 is periodically retrained 890, thereby globally optimizing the follow-up plan generation strategy of the follow-up plan generation model 800, so that the target follow-up plan generated by the follow-up plan generation model 800 is always consistent with the most effective clinical practice.
[0096] pass Figure 8The feedback loop shown on the right side of the diagram allows the follow-up plan generation model 800 to be optimized based on long-term (e.g., six months, one year, or even longer) efficacy data of the target subjects. Therefore, this feedback loop is also called the long-term feedback loop. Through this long-term feedback loop, the follow-up plan generation model 800 can be optimized based on real efficacy data, ensuring that the target follow-up plan generated by the follow-up plan generation model 800 is always consistent with the most effective clinical practice.
[0097] In some examples, in addition to optimization based on efficacy data, the long-term feedback loop may also include feedback optimization based on guideline updates. Therefore, when the feedback data includes updated medical knowledge data, new positive samples are generated based on the updated medical knowledge data; and the follow-up plan generation model is retrained using these new positive samples to update the follow-up plan generation model.
[0098] In the example above, specifically, it is possible to continuously monitor medical knowledge updates, particularly authoritative clinical guidelines such as clinical practice guidelines published and updated by NCCN and CSCO.
[0099] When updated medical knowledge data is detected (e.g., the NCCN clinical practice guidelines adjust the adjuvant targeted therapy cycle for breast cancer from 12 months to 6 months), a guideline analysis (e.g., within 48 hours) will be performed to generate new positive samples (e.g., 882) that are adapted to the updated content. In some examples, new negative samples can also be generated based on older medical knowledge.
[0100] Then, these new positive samples 882 can be used to trigger periodic retraining 890 of the follow-up plan generation model 800, thereby ensuring that the target follow-up plan generated by the follow-up plan generation model is always consistent with the latest medical guidelines.
[0101] Based on the same inventive concept as the aforementioned technical solutions, this disclosure also provides a follow-up plan generation device. Figure 9 This is a schematic diagram of the components of the follow-up plan generation device 900. In this disclosure, the follow-up plan generation device 900 can be implemented in hardware, software, or firmware. For example, the follow-up plan generation device 900 can be… Figure 1 The physical entity of server 110 or the functional program running on server 110.
[0102] See Figure 9 The follow-up plan generation device 900 may include: Module 901 is configured to acquire multi-dimensional medical data of the target object. The generation module 902 is configured to generate a target follow-up plan for the target object based on multi-dimensional medical data and through a follow-up plan generation model. The follow-up plan generation model is obtained by supervised fine-tuning of the model to be trained using a positive sample set and a negative sample set. The positive sample set includes at least one positive sample, and the negative sample set includes at least one negative sample. The positive sample is a verified and reasonable first training follow-up plan, and the negative sample is a verified and unreasonable second training follow-up plan.
[0103] In some examples, generation module 902 is configured as follows: Standardize multi-dimensional medical data to generate embedded vector sequences that can be processed by the follow-up plan generation model; and The embedded vector sequence is input into the follow-up plan generation model so that the follow-up plan generation model outputs the target follow-up plan.
[0104] In some examples, multidimensional medical data includes structured and unstructured medical data, and the embedded vector sequence includes a first vector sequence and a second vector sequence; Module 902 is configured as follows: The structured medical data is time-aligned, and the processed structured medical data is converted into a first vector sequence; and Feature extraction is performed on unstructured medical data, and the processed unstructured medical data is converted into a second vector sequence.
[0105] In some examples, the follow-up plan generation model integrated in the generation module 902 includes a temporal convolutional network, an inference network, and an output layer; Module 902 is configured as follows: Based on the first vector sequence, temporal analysis is performed through a temporal convolutional network to generate temporal analysis results. Based on the second vector sequence, multimodal inference is performed through an inference network to generate multimodal inference results; and Based on the time series analysis results and multimodal reasoning results, a target follow-up plan is generated through the output layer.
[0106] In some examples, such as Figure 10 As shown, the follow-up plan generation device 900 may further include a training module 903, configured to perform supervised fine-tuning of the model to be trained using a positive sample set and a negative sample set to generate the follow-up plan generation model.
[0107] In some examples, the training module 903 is configured to perform comparative learning training on the model to be trained based on positive and negative sample sets, so as to narrow the vector distance between the training follow-up plan generated by the model to be trained and the first training follow-up plan, and widen the vector distance between the training follow-up plan and the second training follow-up plan, until the comparative loss function meets the first preset condition, and the follow-up plan generation model is obtained.
[0108] In some examples, the training module 903 is configured to perform safety boundary training on the model to be trained based on a negative sample set, so that if the similarity between the training follow-up plan generated by the model to be trained and the second training follow-up plan exceeds a threshold, the model to be trained is penalized by a risk loss function until the risk loss function meets a second preset condition, and the follow-up plan generation model is obtained.
[0109] In some examples, training module 903 is also configured to acquire feedback data when the target subjects execute the target follow-up plan; this feedback data includes at least one of the following: plan adjustment data for adjusting the target follow-up plan, efficacy data after the target subjects execute the follow-up plan, and medical knowledge update data; and Based on the feedback data, the follow-up plan generation model is optimized to update the follow-up plan generation model.
[0110] In some examples, training module 903 is also configured as follows: If the feedback data includes plan adjustment data, new positive samples and / or new negative samples are generated based on the plan adjustment data; The follow-up plan generation model is incrementally trained using new positive and / or new negative samples to update the follow-up plan generation model.
[0111] In some examples, training module 903 is also configured as follows: When feedback data includes efficacy data, is it reasonable to determine the target follow-up plan based on the efficacy data? If the target follow-up plan is deemed reasonable, the target follow-up plan will be used as the new positive sample. If the target follow-up plan is deemed unreasonable, the target follow-up plan will be used as a negative sample. The follow-up plan generation model is updated by using new positive or negative samples for reinforcement learning.
[0112] In some examples, training module 903 is also configured as follows: If the feedback data includes updated medical knowledge data, new positive samples are generated based on the updated medical knowledge data. The follow-up plan generation model was retrained using new positive samples to update the follow-up plan generation model.
[0113] This disclosure also provides a computer-readable storage medium storing at least one instruction that is executed by a processor to implement the follow-up plan generation method as described in the various embodiments above.
[0114] This disclosure also provides a computer program product including computer instructions stored in a computer-readable storage medium; a processor of a computing device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computing device to perform the follow-up plan generation method described in the above embodiments.
[0115] Those skilled in the art will recognize that the functions described in this disclosure in one or more of the examples above can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of a computer program from one place to another. Storage media can be any available medium accessible to a general-purpose or special-purpose computer.
[0116] It should be noted that the technical solutions described in this disclosure can be combined arbitrarily as long as they do not conflict.
[0117] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure.
Claims
1. A method for generating a follow-up plan, characterized in that, The generation method includes: Acquire multi-dimensional medical data of the target object; Based on the multi-dimensional medical data, a target follow-up plan for the target object is generated through a follow-up plan generation model. The follow-up plan generation model is obtained by supervised fine-tuning of the model to be trained using a positive sample set and a negative sample set; the positive sample set includes at least one positive sample, the negative sample set includes at least one negative sample, the positive sample is a verified reasonable first training follow-up plan, and the negative sample is a verified unreasonable second training follow-up plan.
2. The method according to claim 1, characterized in that, The step of generating a target follow-up plan for the target object based on the multi-dimensional medical data and through a follow-up plan generation model includes: The multidimensional medical data is standardized to generate an embedded vector sequence that the follow-up plan generation model can process; The embedded vector sequence is input into the follow-up plan generation model so that the follow-up plan generation model outputs the target follow-up plan.
3. The method according to claim 2, characterized in that, The multi-dimensional medical data includes structured medical data and unstructured medical data, and the embedded vector sequence includes a first vector sequence and a second vector sequence; The standardization process of the multi-dimensional medical data to generate an embedded vector sequence that the follow-up plan generation model can process includes: The structured medical data is time-aligned, and the processed structured medical data is converted into the first vector sequence. The unstructured medical data is subjected to feature extraction processing, and the processed unstructured medical data is converted into the second vector sequence.
4. The method according to claim 3, characterized in that, The follow-up plan generation model includes a temporal convolutional network, an inference network, and an output layer; The step of inputting the embedded vector sequence into the follow-up plan generation model so that the follow-up plan generation model outputs the target follow-up plan includes: Based on the first vector sequence, temporal analysis is performed through the temporal convolutional network to generate temporal analysis results. Based on the second vector sequence, multimodal reasoning is performed through the inference network to generate multimodal reasoning results; Based on the time series analysis results and the multimodal reasoning results, the target follow-up plan is generated through the output layer.
5. The method according to any one of claims 1 to 4, characterized in that, The supervised fine-tuning of the model to be trained using positive and negative sample sets includes: The model to be trained is subjected to comparative learning training based on the positive sample set and the negative sample set, so as to narrow the vector distance between the training follow-up plan generated by the model to be trained and the first training follow-up plan, and widen the vector distance between the training follow-up plan and the second training follow-up plan, until the comparative loss function meets the first preset condition, and the follow-up plan generation model is obtained.
6. The method according to any one of claims 1 to 4, characterized in that, The supervised fine-tuning of the model to be trained using positive and negative sample sets includes: The model to be trained is trained on a safety boundary based on the negative sample set. If the similarity between the training follow-up plan generated by the model to be trained and the second training follow-up plan exceeds a threshold, the model to be trained is penalized by a risk loss function until the risk loss function meets a second preset condition, thereby obtaining the follow-up plan generation model.
7. The method according to claim 1, characterized in that, The method further includes: Obtain feedback data when the target subject executes the target follow-up plan; the feedback data includes at least one of the following: plan adjustment data for adjusting the target follow-up plan, efficacy data of the target subject during the execution of the target follow-up plan, and medical knowledge update data. Based on the feedback data, the follow-up plan generation model is optimized to update the follow-up plan generation model.
8. The method according to claim 7, characterized in that, The step of optimizing the follow-up plan generation model based on the feedback data to update the follow-up plan generation model includes: If the feedback data includes the plan adjustment data, new positive samples and / or new negative samples are generated based on the plan adjustment data; The follow-up plan generation model is incrementally trained using the new positive samples and / or the new negative samples to update the follow-up plan generation model.
9. The method according to claim 7, characterized in that, The step of optimizing the follow-up plan generation model based on the feedback data to update the follow-up plan generation model includes: If the feedback data includes the efficacy data, determine whether the target follow-up plan is reasonable based on the efficacy data; If the target follow-up plan is deemed reasonable, the target follow-up plan will be used as a new positive sample. If the target follow-up plan is determined to be unreasonable, the target follow-up plan will be used as a new negative sample. The follow-up plan generation model is updated by using the new positive or negative samples to perform reinforcement learning.
10. The method according to claim 7, characterized in that, The step of optimizing the follow-up plan generation model based on the feedback data to update the follow-up plan generation model includes: If the feedback data includes the updated medical knowledge data, a new positive sample is generated based on the updated medical knowledge data. The follow-up plan generation model is retrained using the new positive samples to update the follow-up plan generation model.
11. The method according to claim 1, characterized in that, The follow-up plan shall include at least the following: follow-up items, follow-up timeline, follow-up frequency, and follow-up risks.
12. An apparatus for generating a follow-up plan, characterized in that, include: The acquisition module is used to acquire multi-dimensional medical data of the target object; The generation module is used to generate a target follow-up plan for the target object based on the multi-dimensional medical data and through a follow-up plan generation model. The follow-up plan generation model is obtained by supervised fine-tuning of the model to be trained using a positive sample set and a negative sample set; the positive sample set includes at least one positive sample, the negative sample set includes at least one negative sample, the positive sample is a verified reasonable first training follow-up plan, and the negative sample is a verified unreasonable second training follow-up plan.
13. An electronic device, characterized in that, The electronic device includes: processor; and a memory, in which computer programs are stored; When the processor is configured to execute the computer program, it implements the method for generating a follow-up plan as described in any one of claims 1 to 11.
14. 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 method for generating a follow-up plan as described in any one of claims 1 to 11.