A smart processing and analysis feedback system for gastric cancer PROs data

Abstract

This invention relates to the field of medical data processing technology, specifically to an intelligent processing, analysis, and feedback system for gastric cancer prognostic (PRO) data. The system includes a data acquisition module, a weight reconstruction module, and a field optimization module. It acquires questionnaire information and clinical data with initial weights from patients at different disease stages. Based on the accuracy of field prediction and completion, it performs field classification and candidate field allocation under dual constraints. After reconstructing field weights, it dynamically adjusts the model's generalization ability and iteratively optimizes to obtain questionnaire-appropriate fields for each disease stage. This invention improves the matching degree between gastric cancer PRO questionnaires and disease stages, balances the longitudinal comparability of cross-stage data with the effectiveness of single-stage data, reduces the burden on patients, and improves the efficiency and clinical applicability of PRO data collection.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, specifically to an intelligent processing, analysis, and feedback system for gastric cancer PROs data. Background Technology

[0002] Patient-reported outcomes (PROs) data for gastric cancer are crucial for clinical assessment of patients' conditions and the development of individualized treatment plans. Current methods for collecting gastric cancer PROs data often employ standardized questionnaire templates covering all disease stages without dynamically adapting questionnaire fields to different stages. This results in low data relevance to patients' current clinical circumstances, high questionnaire redundancy, heavy burden on patients, poor compliance, and the generation of significant noise and invalid data, thus hindering the enhancement of the clinical application value of PROs data. Summary of the Invention

[0003] To address the aforementioned problems, this invention provides an intelligent processing, analysis, and feedback system for gastric cancer PROs data.

[0004] The intelligent processing, analysis, and feedback system for gastric cancer PROs data of the present invention adopts the following technical solution: One embodiment of the present invention provides an intelligent processing and analysis feedback system for gastric cancer PROs data, the system comprising the following modules: The data acquisition module is used to acquire questionnaire information and clinical data of patients at different stages of their condition; the questionnaire information contains several fields with initial weights. The weight reconstruction module, based on the model trained with the initial weights of each field, obtains the accuracy of the predicted content for each field at each stage of the disease. Based on this accuracy, all fields are divided into: suitable category, candidate category, and invalid category. Fields in the candidate category are assigned to the suitable and invalid categories to maximize the overlap between the assigned suitable categories at different disease stages, and to maximize the accuracy difference between the fields in the assigned suitable and invalid categories at each disease stage. Finally, the weights of the fields in the assigned suitable and invalid categories at each disease stage are reconstructed. The field optimization module reduces the model's generalization ability by adjusting the weight differences of fields assigned to the appropriate and invalid categories in each disease stage candidate category. The reconstructed weights are then used as the initial weights, and the weight reconstruction module is rerun. During the rerun of the weight reconstruction module, the fields in the appropriate category after reassignment in each disease stage are used as the fields for feedback questionnaire information in each disease stage.

[0005] Preferably, the specific steps for obtaining the adaptation category, candidate category, and invalid category are as follows: For any given stage of the disease, the accuracy of all fields is assessed, with the fields having the highest accuracy assigned to the fit category, the fields having the lowest accuracy assigned to the invalid category, and the remaining fields assigned to the candidate category.

[0006] Preferably, the specific steps of assigning fields from candidate categories to suitable and invalid categories are as follows: For each disease stage, the obtained matching category, candidate category, and invalid category are selected. Several fields are selected from the candidate categories and temporarily assigned to the matching category. The remaining fields in the candidate categories are temporarily assigned to the invalid category. Obtain the overlap between the temporarily assigned fit categories under different disease stages; calculate the first difference between the average accuracy of the temporarily assigned fit categories and the average accuracy of the temporarily assigned invalid categories under any disease stage, and record the average of the first differences obtained under all disease stages as the first accuracy difference; obtain the allocation index, which is positively correlated with the first accuracy difference and overlap; when the allocation index is at its maximum, the fields selected from the candidate categories are allocated to the fit categories, and the remaining fields are allocated to the invalid categories.

[0007] Preferably, the specific steps for obtaining the overlap between the temporarily assigned fit categories under different disease stages are as follows: For all disease stages, the temporarily assigned fitting categories are recorded as the baseline category, and the temporarily assigned fitting categories of two adjacent disease stages are referred to as adjacent fitting categories. Any adjacent adapted category of the baseline category is denoted as a 1-neighbor category. The intersection-union ratio (IU) of the baseline category and the 1-neighbor category is obtained, and the intersection of the baseline category and the 1-neighbor category is denoted as the 1-neighbor intersection. The adjacent adapted categories of the 1-neighbor category are denoted as 2-neighbor categories. The intersection-union ratio of the 1-neighbor intersection and the 2-neighbor category is obtained, and the intersection of the 1-neighbor intersection and the 2-neighbor category is denoted as the 2-neighbor intersection. This process is repeated until all neighbor categories of the baseline category have been traversed. The average of all the obtained IU ratios is denoted as the cross-stage association index of the baseline category. The mean of the cross-stage association index of the appropriate categories after temporary assignment for all disease stages is denoted as the overlap degree.

[0008] Preferably, the specific steps for reconstructing the weights of fields in the assigned adaptive and invalid categories for each disease stage are as follows: The fields from the candidate categories in the assigned adaptive categories are denoted as the first field set, and the fields from the candidate categories in the assigned invalid categories are denoted as the second field set. When the accuracy of any field in the first field set is less than the maximum accuracy of all fields in the second field set, the accuracy of that field is set to the maximum value; when the accuracy of any field in the second field set is greater than the minimum accuracy of all fields in the first field set, the accuracy of that field is set to the minimum value.

[0009] Preferably, the specific steps for reducing the generalization ability of the model are as follows: The generalization ability of the model is reduced by changing the pruning parameters of the model, and the magnitude of the change in the pruning parameters is negatively correlated with the weight difference.

[0010] Preferably, the model trained based on the initial weights of each field includes the following specific steps: Clinical data and questionnaire information for each disease stage are used as samples and labels to form a dataset. A model is initialized for each disease stage and trained using the dataset for each disease stage. During training, after the samples are input into the model, the model outputs the predicted content for each field. The difference between the actual content and the predicted content for each field in the label is recorded as the error for each field. The errors of all fields are weighted and summed based on the initial weights of all fields, and the weighted sum is used as the loss function for the training process.

[0011] Preferably, the specific steps for obtaining the accuracy are as follows: At any stage of the disease, the difference between the predicted content of each field output by each patient and the actual content of each patient in each field in the model after inputting the clinical data of each patient into the model is denoted as the first difference of each patient in each field at any stage of the disease; the accuracy is negatively correlated with the median of the first differences of all patients in each field at any stage of the disease.

[0012] Preferably, the specific steps for changing the pruning parameters of the model are as follows: When the pruning parameter of the model is the dropout parameter, the dropout parameter is reduced, and the reduction is negatively correlated with the weight difference.

[0013] Preferably, the specific steps for obtaining the weight difference are as follows: The difference between the average reconstructed weights of the fields assigned to the candidate category and the average reconstructed weights of the fields assigned to the invalid category in each disease stage is denoted as the weight difference.

[0014] The beneficial effects of the technical solution of the present invention are: The intelligent processing and analysis feedback system for gastric cancer PROs data of the present invention acquires patient clinical data and questionnaire information with initial weights in stages, and combines an iterative optimization process of field prediction accuracy classification, dual-constraint candidate field allocation, weight reconstruction and dynamic adjustment of model generalization ability to achieve adaptive optimal configuration of questionnaire fields at each stage of the disease.

[0015] Overall, this approach effectively addresses the technical challenges of low matching rates between PROs (Prognostic Disease Status) data and disease stages, as well as poor longitudinal comparability across the entire disease course in existing technologies. Specifically, it assigns candidate fields based on two constraints: maximizing the overlap of adapted categories across stages and maximizing the difference in accuracy between adapted and invalid categories within a single stage. This ensures, on the one hand, that each questionnaire field within a single stage represents the field with the highest reference value for that corresponding disease stage, reducing questionnaire redundancy, lessening the burden on patients, and improving the effectiveness of data collection. On the other hand, it ensures that adapted fields across different disease stages have sufficient commonality, enabling longitudinal comparability of PROs data throughout the entire disease course. This provides a unified reference dimension for clinical efficacy evaluation and disease progression tracking, avoiding the problem of excessively sacrificing the clinical usability of questionnaire feedback in order to reduce questionnaire redundancy. Secondly, by dynamically adjusting the model's generalization ability and iteratively optimizing field classification based on the weight differences assigned to candidate fields, it improves the classification accuracy of low-discrimination fields at the boundary, avoiding misclassification of fields when using a model with high generalization ability to adapt to high-noise questionnaire information. This further improves the accuracy and clinical applicability of questionnaire field configuration at each stage. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0017] Figure 1 This is a framework diagram of an intelligent processing and analysis feedback system for gastric cancer PROs data provided in one embodiment of the present invention; Figure 2 The flowchart shows the steps of all modules in an intelligent processing and analysis feedback system for gastric cancer PROs data provided in one embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an intelligent processing and analysis feedback system for gastric cancer PROs data proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent processing and analysis feedback system for gastric cancer PROs data provided by this invention.

[0021] Please see Figure 1 The diagram illustrates the framework of an intelligent processing and analysis feedback system for gastric cancer PROs data according to an embodiment of the present invention. The system includes a data acquisition module, a weight reconstruction module, and a field optimization module.

[0022] The data acquisition module collects data in stages, achieving a one-to-one correspondence between questionnaire information and patients' clinical situations. This avoids the mixing of irrelevant data from stages of the illness and provides a data foundation for subsequent analysis and processing.

[0023] The weight reconstruction module assigns high-accuracy fields to the fit category, low-accuracy fields to the invalid category, and fields in the intermediate boundary range to the candidate category. Based on this initial classification, the final assignment of candidate fields is completed through dual-constraint optimization, taking into account both the comparability of cross-stage fields and the discriminability of single-stage fields. The weights are then further reconstructed, and the reconstructed weights are related to the actual stage matching degree of the fields. Subsequent model iterations based on the reconstructed weights help improve the accuracy of the model output and provide a data foundation for improving the clinical applicability of the model output.

[0024] The field optimization module adjusts the model's generalization ability and performs secondary classification and allocation of all fields. Ultimately, it optimizes the output of the unique fields of the best-fit questionnaire for each stage of the disease, achieving a high degree of matching between the questionnaire and the patient's disease stage, and improving the clinical applicability of the questionnaire.

[0025] Figure 2 The specific steps included in all modules are shown.

[0026] The data acquisition module includes the following steps: Step S101: Obtain questionnaire information and clinical data of patients at different stages of their condition; the questionnaire information contains several fields with initial weights.

[0027] This embodiment considers that using a uniform template to collect questionnaire data for all disease stages, without differentiating the data source by disease stage and without setting field priority criteria, will lead to low matching between the collected data and the patient's clinical situation. All patients will complete the same lengthy questionnaires irrelevant to their specific disease stage, resulting in data redundancy, errors, or missing data, heavy patient burden, and low compliance. Therefore, this embodiment will subsequently dynamically configure the fields within the questionnaire for different disease stages, achieving a high degree of matching between questions and the patient's current clinical situation, improving data relevance and collection efficiency.

[0028] In this embodiment, questionnaire information and clinical data are collected according to the different disease stages of the patients (the number of disease stages is greater than 1); in the collected questionnaire information, each field is manually configured with a corresponding initial weight as the basis for field priority.

[0029] In this embodiment, the phased data collection achieves a one-to-one correspondence between questionnaire information and the patient's clinical situation, avoiding the inclusion of irrelevant data that does not correspond to the disease stage, and providing a data foundation for subsequent analysis and processing.

[0030] The weight reconstruction module includes the following steps: Step S102: Based on the initial weights of each field, train the model to obtain the accuracy of the predicted content of each field under each stage of the disease, and divide all fields into: suitable category, candidate category, and invalid category according to the accuracy.

[0031] In this embodiment, each disease stage corresponds to a model. The model refers to a model trained using clinical data and questionnaire information of all patients at each disease stage. When training the model, the larger the initial weight, the more attention is paid to the content filled in the corresponding field in the questionnaire information. This helps guide the model to prioritize fitting the field features with high initial weights, making the training process consistent with clinical experience logic, and making the model initially clinically applicable.

[0032] This model can predict the content filled in each field of the questionnaire based on clinical data at each stage of the disease. The accuracy of any field describes the difference between the predicted content and the actual content filled in the document. The higher the accuracy (the smaller the difference), the higher the effectiveness of the field in the corresponding stage of the disease and the higher the degree of matching with clinical needs. The lower the accuracy (the larger the difference), the higher the randomness or missing rate of the field in the corresponding stage of the disease and the lower the degree of matching with clinical needs.

[0033] In this step, the corresponding matching category for each disease stage contains the fields with the highest degree of matching (or fit) with clinical needs; the corresponding invalid category for each disease stage contains the fields with the lowest degree of matching (or fit) with clinical needs; the accuracy of fields in the candidate category is less than the accuracy of fields in the matching category, but greater than the accuracy of fields in the invalid category.

[0034] In this embodiment, considering that the questionnaire information used to train the model has significant noise (e.g., noise due to data redundancy errors or missing data, or low compliance), this embodiment uses a model with high generalization ability for training to avoid the model being contaminated or interfered with by noise errors. However, the drawback of using a model with high generalization ability is that the quantitative representation ability of the accuracy of the above evaluation is limited, resulting in the inability to clearly describe whether the accuracy in the middle range (i.e., the accuracy of fields in the candidate categories) matches the disease stage (i.e., it is impossible to assert the attribution of boundary fields).

[0035] In this embodiment, high-accuracy fields are classified into the matching category, low-accuracy fields into the invalid category, and fields in the middle boundary range are classified into the candidate category. This avoids misclassification of boundary fields and includes fields whose matching degree cannot be clearly determined in the subsequent optimization and allocation process, thereby improving the accuracy of the final field classification. The setting of candidate categories provides room for adjustment in the subsequent cross-stage matching field overlap optimization, meeting the need for field comparability between different disease stages.

[0036] Step S103: Assign fields from candidate categories to suitable and invalid categories to maximize the overlap between suitable categories assigned under different disease stages, and maximize the accuracy difference between fields in suitable and invalid categories assigned under each disease stage; reconstruct the weights of fields in suitable and invalid categories assigned under each disease stage.

[0037] This embodiment takes into account that the accuracy of candidate category fields falls within an intermediate range. Directly retaining or eliminating them cannot guarantee optimal questionnaire configuration. Retaining all fields would increase questionnaire length, while eliminating all would result in the loss of potentially high-value fields and would not support the optimization requirements for cross-stage overlap. This step assigns candidate fields, clarifying their final classification as either suitable or invalid. This process achieves final binary classification of all fields, avoiding the impact of candidate field uncertainty on questionnaire generation. Furthermore, the assignment process can be combined with subsequent global optimization goals to compensate for the limitations of the initial classification based solely on the accuracy of a single stage (i.e., an independent stage of the disease).

[0038] This embodiment further considers that using single-stage accuracy as the allocation criterion would result in excessively low overlap of matching fields across different disease stages, making longitudinal comparison of questionnaire data from different disease courses impossible and failing to support the tracking and analysis of changes in patient conditions. This step uses maximizing the overlap of cross-stage matching categories as the global optimization objective to guide the allocation of candidate fields, ensuring that questionnaires from different disease stages have common comparison fields. This achieves longitudinal comparability of PROs data throughout the entire patient disease course, providing a unified reference dimension for clinical efficacy evaluation and disease progression tracking.

[0039] The greater the overlap, the more common matching fields there are across different disease stages, and the stronger the comparability of the longitudinal data; the smaller the overlap, the more unique matching fields there are across different disease stages, and the weaker the comparability of the longitudinal data.

[0040] Furthermore, using cross-stage overlap as the sole allocation criterion can lead to some low-accuracy fields being assigned to the appropriate category and high-accuracy fields to the invalid category, resulting in decreased field relevance in single-stage questionnaires, increased burden on patients, and increased data noise. This step further optimizes the allocation of candidate fields by maximizing the accuracy difference between the appropriate and invalid categories within a single stage, working in conjunction with the global overlap objective. This ensures that the average accuracy of fields in the appropriate category is significantly higher than that in the invalid category within a single stage, guaranteeing that the final questionnaire fields are those with the highest reference value for that stage. This reduces questionnaire redundancy, lessens the burden on patients, and improves data validity, while also considering the longitudinal comparability of PROs data throughout the entire patient's disease course, avoiding the problem of sacrificing the clinical usability of the questionnaire feedback in order to reduce questionnaire redundancy.

[0041] Furthermore, the initial weights mentioned above are configured solely based on prior knowledge and do not incorporate the accuracy analysis results of actual data. This fails to reflect the differences in actual field matching degrees, causing the model optimization direction to deviate from actual needs. This step reconstructs the weights of each field based on the assigned field category and accuracy results. The reconstructed weights are related to the actual stage matching degree of the field. Subsequent model iterations based on these reconstructed weights help improve the accuracy of the model output, providing a data foundation for enhancing the clinical applicability of the model output.

[0042] At this point, for each stage of the disease, the field has been reweighted based on the assigned suitable and invalid categories. It should be noted that the reweighted values ​​for the same field may differ across different disease stages.

[0043] In summary, this step, based on the initial classification described above, completes the final assignment of candidate fields through dual-constraint optimization, taking into account both the comparability of fields across stages and the discriminative power of fields within a single stage. At the same time, the reconstructed weights provide a data source foundation for the subsequent optimization iterations of the model.

[0044] The field optimization module includes the following steps: Step S104: Reduce the model's generalization ability based on the weight differences of the fields assigned to the appropriate and invalid categories under each disease stage, and use the reconstructed weights as the initial weights to redefine the categories and fields. Use the fields in the appropriate categories reassigned under each disease stage as the fields used to provide feedback on questionnaire information under each disease stage.

[0045] This step addresses the issue of noise interference during the initial model training, including the unreliability of initial weights based solely on prior knowledge. A high generalization model is used, which can lead to insufficient fitting of the model to boundary fields within candidate categories, resulting in low accuracy and discriminative power of the candidate category fields and hindering precise allocation. This embodiment addresses this by reducing the model's generalization ability based on the weight differences assigned to the two classes during the previous steps. This improves the model's fitting accuracy to boundary fields, thereby increasing the prediction accuracy for low-discriminative fields, amplifying the accuracy differences of boundary candidate fields, and resolving the issue of ambiguous classification of boundary fields under high generalization.

[0046] Furthermore, the generalization ability is dynamically adjusted based on the weight differences. For example, a larger weight difference indicates that the fields within the candidate categories are already clearly separable. In this case, there is no need to significantly reduce the generalization ability, thus avoiding the introduction of significant noise. Conversely, a smaller weight difference indicates that the fields within the current candidate categories cannot be reliably separated. In this case, it is necessary to significantly reduce the generalization ability to improve the model's accuracy, making the prediction results for fields that are difficult to distinguish more accurate, so that subsequent clearly separable candidate categories can be obtained.

[0047] In summary, this embodiment reduces the generalization ability of the model based on weight differences, ensuring that fields within candidate categories can be reliably separated and assigned while avoiding noise interference caused by excessive reduction in generalization ability.

[0048] Furthermore, in the above process, the weights of the fields were reconstructed at each stage of the disease, and the reconstructed weights were used as the initial weights. For each stage of the disease, the model for each stage of the disease was retrained according to steps S102 and S103, and all fields were reclassified (reclassified into suitable categories, candidate categories, and invalid categories), and the fields of the candidate categories were reassigned to the suitable and invalid categories (i.e., secondary reassignment).

[0049] The initial weights during training in this process better reflect the actual stage matching degree of the fields, guiding the model to prioritize fitting the features of fields with high matching degree. In this process, all fields are divided twice and then reassigned to the fit category and invalid category. This avoids the problem that the results of the first division and the first assignment are limited by high generalization ability, and the classification accuracy of boundary fields is insufficient, which makes it impossible to achieve the optimal configuration of fit fields. This makes the field classification more accurate, and the allocation of candidate fields better meets the dual optimization goals of global and single-stage optimization.

[0050] Furthermore, the fields in the newly acquired fitting categories (i.e., the fitting categories after secondary allocation) under each disease stage are used as unique fields for feedback questionnaire information under each disease stage; patient interviews and information surveys are conducted using the questionnaire information containing unique fields under each disease stage.

[0051] This process avoids problems such as misclassification of boundary fields, insufficient cross-stage overlap, and insufficient single-stage discrimination that can occur when using the initial classification of the adaptive fields as questionnaire fields. By using the adaptive category fields after the second iteration as the final questionnaire fields for each stage, it simultaneously meets the requirements of high correlation with the corresponding disease stage, high cross-stage field overlap, and low field redundancy. This can reduce the burden on patients, improve questionnaire compliance, reduce data noise and redundancy, and achieve efficient collection of PROs data.

[0052] In summary, all the steps described above in this embodiment adjust the model's generalization ability, update the initial weights, and incorporate secondary classification assignment with cross-stage constraints to ultimately output the unique fields of the optimally matched questionnaire for each disease stage. This achieves a high degree of matching between the questionnaire and the patient's disease stage, thereby improving the clinical applicability of the questionnaire.

[0053] As a preferred example, different disease stages include, but are not limited to, the following periods: initial diagnosis and adjuvant therapy period: after diagnosis and before surgery, chemotherapy or radiotherapy may be performed to shrink the tumor; surgical period: preoperative preparation, the surgery itself (such as radical gastrectomy), and short-term postoperative recovery; adjuvant therapy period: after surgery, chemotherapy, radiotherapy, or targeted therapy to eliminate potential residual cancer cells; recurrence / metastasis treatment period: when the tumor recurs or metastasizes to distant sites, a systemic treatment phase aimed at controlling the disease and prolonging survival (such as multi-line chemotherapy, targeted therapy, and immunotherapy).

[0054] Clinical data include, but are not limited to: tumor invasion depth (T), lymph node metastasis (N), and distant metastasis (M) during the initial diagnosis and adjuvant therapy period; surgical scope, lymph node dissection scope, and postoperative pathology report during the surgical period; physical condition recovery (ECOG score), adjuvant therapy regimen, and records of specific adverse reactions during the adjuvant therapy period; and medication and dosage during the recurrence / metastasis treatment period.

[0055] It should be noted that there are various pathological types of gastric cancer (such as adenocarcinoma, signet ring cell carcinoma, etc.) and tumor locations are also diverse (such as cardia, gastric body, pylorus, etc.). This embodiment does not distinguish between pathological types and tumor locations to ensure the acquisition of a large amount of clinical data. In other examples, clinical data for the different disease stages mentioned above can be collected only for a specific pathological type (such as adenocarcinoma) or a specific tumor location (such as pylorus).

[0056] In addition, other examples may also be used to classify the specific treatment of gastric cancer into other disease stages or collect other clinical data based on specific circumstances (such as the hospital's medical conditions, medical practice experience, etc.).

[0057] It should be further noted that the clinical data in this embodiment are all in text format. While real-world clinical data may contain missing information, errors, or other anomalies, this embodiment utilizes a model with generalization capabilities for intelligent processing and big data mining of the clinical data, thus providing a degree of compatibility with such anomalies.

[0058] As a preferred example, the questionnaire information includes fields such as: abdominal pain / stomach pain, early satiety, postprandial fullness, nausea and other gastric cancer symptoms; fields also include treatment-related side effects such as loss of appetite, changes in taste, diarrhea, constipation, hand-foot syndrome (numbness and erythema in the hands and feet), oral mucositis and other treatment-related side effects; and fields also include nutritional information such as weight changes (past month), difficulty swallowing, and current food type (regular diet, semi-liquid diet, liquid diet, enteral nutrition).

[0059] Other examples also include emotional state fields such as anxiety and depression, and fields for daily activities such as whether they are limited in performing household chores, work, shopping, and going out.

[0060] Other fields can be set in other examples, or fields can be added or deleted according to the specific pathological type of gastric cancer. This embodiment does not impose specific limitations.

[0061] It should be noted that the information entered in each field of the questionnaire in this embodiment is in text or number format. Furthermore, although the questionnaire information may contain redundancy, missing data, errors, or other anomalies, this embodiment considers using a model with a certain degree of generalization ability to intelligently process the questionnaire information and perform big data mining, which can accommodate such anomalies to a certain extent.

[0062] As a preferred example, the initial weights of the fields in the questionnaire information are initialized by setting the initial weights of all fields to the same value, for example, all to 1.

[0063] As a preferred example, the model trained based on the initial weights of each field obtains the accuracy of the predicted content for each field at each stage of the disease. The methods include: Each patient's clinical data at each stage of their condition is used as a sample, and each patient's questionnaire information (including the content of all fields filled in) at each stage of their condition is used as the sample's label. The label refers to an array consisting of the actual content filled in all fields of the questionnaire information. The samples and labels corresponding to all patients at each stage of the condition constitute the dataset for that stage.

[0064] A model is initialized for each stage of the disease, and trained using the dataset for each stage. During training, samples are input into the model, and the model outputs an array of the same length as the labels. Each element in the array represents the predicted content for each field. The difference between each element in the output array and the element with the same index in the labels is recorded as the error for each field. The initial weights of all fields are normalized using the softmax formula, and the errors of all fields are weighted and summed using the normalized initial weights. The weighted sum is used as the loss function for the training process.

[0065] That is, the loss function ,in This represents the initial weight of the i-th field after normalization. This represents the error of the i-th field, and N represents the number of fields.

[0066] Specifically, in some examples, patients may be surveyed multiple times for each stage of the disease. In this case, this embodiment only retains the most recent questionnaire information for each stage. In other examples, each stage of the disease can be further subdivided, for example, the adjuvant therapy period can be considered as a subdivided stage every month. In each stage, only the most recent questionnaire information for that stage is retained, and the clinical data for each patient across these subdivided stages remains identical. In some examples, stages with fewer questionnaire visits (e.g., less than 100) may also be deleted.

[0067] As an example, all disease stages of the model employ fully connected neural networks. Compared to convolutional networks or temporal networks used for processing images and large texts, the fully connected neural network used in this embodiment is suitable for scenarios with relatively small input and output data volumes (i.e., fewer features) (i.e., the scenario of big data mining of clinical data and questionnaire information in this embodiment). Considering the complex nonlinear relationship between clinical data and questionnaire information, this example uses a fully connected neural network with multiple intermediate layers. The number of neurons in the input layer equals the length of the sample, i.e., the dimension of the sample; the number of neurons in the output layer equals the length of the array corresponding to the label, i.e., the dimension of the label. The length of the sample and the length of the array corresponding to the label may differ in different embodiments and are not specifically limited in this embodiment. As an example of this invention, the fully connected neural network used in this invention has 7 layers: the first layer is the input layer, the last layer is the output layer, and it includes 5 intermediate layers with the following neuron counts: 25, 15, 5, 20, and 30, respectively.

[0068] It should also be noted that the aforementioned dataset is quite noisy. To avoid its interference with the model, the dropout parameter of the fully connected neural network is set to 0.5 to ensure its high generalization ability. During training, the learning rate is set to 0.02, the optimizer is Adam, the test set to training set ratio is 3:7, and the maximum number of training iterations is set to 10. 4 Since the specific training parameters need to be combined with clinical data and questionnaire information at specific disease stages, this embodiment does not limit the training parameters. The training process is also well known, and this embodiment will not describe it in detail.

[0069] As an example, before training the model using samples and labels, the samples and labels are vectorized. Specifically, for each sample, it represents multiple indicator data contained in clinical data at the same stage of the disease. The text content of each indicator data is encoded into a normalized vector using word2Vec. In this embodiment, considering that the text content of clinical data is concise text rather than lengthy narrative text, the encoded vector has a smaller dimension, for example, a 5-dimensional vector. Similarly, the text content filled in each field of the label is also encoded into a 5-dimensional vector.

[0070] Specifically, for indicator data or field content in numerical format, the numbers of the same indicator data or field in the dataset are normalized (for example, using the softmax formula for normalization; in other examples, minimum-maximum normalization can also be used); the vectors or normalized scalars corresponding to all indicator data (or field content) in each sample (or label) are merged together (merged into an array or sequence) and input into the model for training.

[0071] It should be noted that missing data in clinical data or questionnaire information is considered random and uncertain data, and is therefore replaced by randomly generated unit vectors or filled with random numbers in the range [0, 1].

[0072] At this point, the error of each field mentioned above refers to the square of the Euclidean distance between each vector element in the array output by the model and the vector element with the same index in the label, or the square of the difference between scalar elements.

[0073] In other examples, the model can be replaced with a random forest model, which is trained using the dataset. The optimization function of the random forest model is obtained in the same way as the loss function described above. As an example, in order to make the random forest model have greater generalization ability, the maximum depth of the decision tree in the random forest model is set to a small value, such as 6.

[0074] In other examples, the model can also be set to other neural network models or machine learning models. In this embodiment, the purpose of using the model is only to perform big data mining on clinical data and questionnaire information. It only focuses on the performance of fields of questionnaire information under different generalization abilities of the same model at each stage of the disease (e.g., the relative distribution of accuracy). There is no need to accurately and reliably predict questionnaire information based on clinical data. Therefore, this embodiment does not specifically limit the model.

[0075] As a preferred example, the methods for obtaining the accuracy of the predicted content for each field at each stage of the disease include: At any stage of the disease, each patient's clinical data is input into the trained model, which outputs the predicted content to be filled in for each field. The difference between the predicted content to be filled in for each field and the actual content to be filled in by each patient in the corresponding questionnaire information for that field is calculated and recorded as the first difference for each patient in each field at any stage of the disease.

[0076] The median of the first difference in each field for all patients at any disease stage is denoted as x. The accuracy of each field at that disease stage is negatively correlated with x. A smaller x indicates that the predicted content for each field is closer to the actual content, and thus the greater the accuracy.

[0077] It should be noted that in this example, both the predicted content and the actual content are encoded as vectors or normalized scalar data, and the first difference refers to the square of the Euclidean distance between scalar elements or the square of the difference between scalar elements.

[0078] As an example, accuracy Q = exp(-x), where exp() represents an exponential function with the natural constant as the base. Specifically, if the actual content entered for a certain field by a patient at any stage of the disease is missing, then the first difference for that patient in that field is deleted and no longer included in the median statistics.

[0079] As a preferred example, based on the stated accuracy, all fields are divided into: fit category, candidate category, and invalid category, including the following methods: For any given stage of the disease, the accuracy of all fields obtained is assessed. The n1 fields with the highest accuracy are assigned to the appropriate category, the n2 fields with the lowest accuracy are assigned to the invalid category, and the remaining fields are assigned to the candidate category.

[0080] As an example, n1 and n2 are each equal to 30% of all fields (rounded up).

[0081] As a preferred example, assigning fields from candidate categories to suitable and invalid categories includes the following methods: For any disease stage, the appropriate category, candidate category, and invalid category are obtained. Several fields are selected from the candidate category and temporarily assigned to the appropriate category. The remaining fields in the candidate category are temporarily assigned to the invalid category.

[0082] For all disease stages, the appropriate categories, candidate categories, and invalid categories obtained are temporarily assigned to the appropriate and invalid categories according to the above method.

[0083] Furthermore, for all disease stages, the overlap between the temporarily assigned suitable and invalid categories is obtained. The first difference between the average accuracy of all fields within a temporarily assigned suitable category and the average accuracy of all fields within a temporarily assigned invalid category is calculated for any disease stage. The average of these first differences across all disease stages is recorded as the first accuracy difference. An allocation index is obtained, which is positively correlated with both the first accuracy difference and the overlap. That is, the larger the first accuracy difference and the greater the overlap, the larger the allocation index. This indicates a greater overlap between the temporarily assigned suitable categories under different disease stages, and a greater accuracy difference between the fields in each temporarily assigned suitable category and the invalid category under each disease stage.

[0084] It should be noted that there are multiple selection methods when choosing several fields from candidate categories under each disease stage. The selection methods under all disease stages can be combined with each other. Each combination result means that the above calculation process is performed using one of the selection methods for each disease stage, and thus a corresponding allocation index is obtained.

[0085] The maximum combination of allocation indicators is obtained, which represents the selection method of fields within the candidate categories corresponding to all disease stages when the allocation indicator is maximum. Fields within the candidate categories are then assigned to suitable and invalid categories according to this selection method.

[0086] As an example, the allocation metric is equal to the sum of the first accuracy difference and the overlap.

[0087] As an example, the selection of several fields from candidate categories is fixed at 50% of the total number of fields in the candidate categories (rounded up).

[0088] As an optional example, assigning fields from candidate categories to suitable and invalid categories includes the following methods: The above preferred example may face situations with high computational load (e.g., when there are many fields within the disease stage or candidate category). This optional example provides a method to optimize computational load: For each disease stage, the candidate categories are clustered into two clusters based on the accuracy of the fields using K-Means. The first average of the accuracy of all fields in each cluster is obtained, and the absolute value of the difference between the first averages of the two clusters is recorded as the first indicator for each disease stage. The n3 disease stages with the largest first indicator are obtained for all disease stages.

[0089] Furthermore, in accordance with the above preferred example, when selecting several fields from candidate categories under n3 disease stages, the fields are selected and assigned (or temporarily assigned) according to a unified and fixed selection method. That is, for any candidate category under n3 disease stages, the selected fields are always the fields in each of the above clusters. The selection and assignment (or temporary assignment) of fields specifically means: the fields in the cluster with the largest first average value are assigned to the corresponding suitable category, and the fields in the cluster with the smallest first average value are assigned to the corresponding invalid category.

[0090] This optional example significantly reduces computation by implementing a selection method for a given subset of candidate categories and then following the preferred example.

[0091] As an example, n3 is set to half the number of all disease stages (rounded down).

[0092] As a preferred example, the method for obtaining the overlap between temporarily assigned fit categories under different disease stages includes: For all disease stages, the temporarily assigned fit categories are designated as the baseline category, and the temporarily assigned fit categories of two adjacent disease stages are referred to as adjacent fit categories.

[0093] Any adjacent adapted category of the baseline category is denoted as a 1-neighbor category. The intersection-union ratio (i.e., the ratio of the number of fields in the intersection to the number of fields in the union) of the baseline category and the 1-neighbor category is obtained. The intersection of the baseline category and the 1-neighbor category is denoted as the 1-neighbor intersection. The adjacent adapted categories of a 1-neighbor category are denoted as 2-neighbor categories (excluding the baseline category). The intersection-union ratio of the intersection of 1-neighbor categories and 2-neighbor categories is obtained, and the intersection of the intersection of 1-neighbor categories and 2-neighbor categories is denoted as the 2-neighbor intersection. The adjacent fitting categories of a 2-neighbor category are denoted as 3-neighbor categories (excluding the baseline category and 1-neighbor categories). The intersection-union ratio of the 2-neighbor intersection and the 3-neighbor category is obtained, and the intersection of the 2-neighbor intersection and the 3-neighbor category is denoted as the 3-neighbor intersection. Following this process, after traversing all neighboring categories of the baseline category, the mean of all intersection-union ratios is recorded as the cross-stage association index of the baseline category. The larger the cross-stage association index, the more the fields within the baseline category can encompass or associate multiple disease stages, the more matching fields the disease stage corresponding to the baseline category shares with other disease stages, and the stronger the comparability of the longitudinal data.

[0094] For each of the temporary assigned fit categories under all disease stages, the above method is applied, and the mean of the cross-stage correlation index of all fit categories is recorded as the overlap degree.

[0095] In other examples, considering that when a neighborhood intersection is empty, and that all subsequent intersection-union ratios permanently return to zero, cross-stage association analysis is not possible; this example provides an alternative method for calculating overlap: The cross-stage correlation index of the baseline category obtained according to the above example is denoted as the first-order cross-stage correlation index; next: (1) The benchmark category is directly compared with the 2 neighboring categories. The intersection of the benchmark category and the 2 neighboring categories is called the 2 neighboring intersection. The difference between this process and the example above is that instead of analyzing and associating the benchmark category with the 1 neighboring category, it is directly analyzed and associating with the 2 neighboring categories, that is, skipping the 1 neighboring category to perform analysis and association.

[0096] Furthermore, similarly to the above, the adjacent adapted categories of the 2-neighbor category are denoted as the 3-neighbor category. The intersection-union ratio of the 2-neighbor intersection and the 3-neighbor category is obtained, and the intersection of the 2-neighbor intersection and the 3-neighbor category is denoted as the 3-neighbor intersection. And so on, after traversing all the neighbor categories of the base category, the average of all obtained intersection-union ratios is denoted as the second-order cross-stage association index of the base category.

[0097] (2) The benchmark category is directly compared with the 3 neighboring categories. The intersection of the benchmark category and the 3 neighboring categories is called the 3 neighboring intersection. The difference between this process and the example above is that instead of analyzing and associating the benchmark category with the 1 and 2 neighboring categories, it is directly analyzed and associating with the 3 neighboring categories. That is, the 1 and 2 neighboring categories are skipped for analysis and association.

[0098] Furthermore, similarly to the above, the adjacent fitting categories of the 3-neighbor category are denoted as the 4-neighbor category. The intersection-union ratio of the 3-neighbor intersection and the 4-neighbor category is obtained, and the intersection of the 3-neighbor intersection and the 4-neighbor category is denoted as the 4-neighbor intersection. And so on, after traversing all the neighbor categories of the base category, the average of all obtained intersection-union ratios is denoted as the third-order cross-stage association index of the base category.

[0099] (3) Skip the 1 neighbor category to perform the analysis and association to obtain the second-order cross-stage association index. Skip the 1 and 2 neighbor categories to perform the analysis and association to obtain the third-order cross-stage association index. And so on, until all the intermediate neighbor categories are skipped, so that the benchmark category directly intersects and compares with the last neighbor category. Then, in the same way as above, the last-order cross-stage association index is obtained.

[0100] (4) Calculate the average of the first-order cross-stage correlation index, the second-order cross-stage correlation index, the third-order cross-stage correlation index, ..., the last-order cross-stage correlation index. The average result is used as the final cross-stage correlation index of the benchmark category. Finally, the average of the final cross-stage correlation index of all the matching categories is recorded as the overlap.

[0101] This example implements association analysis across multiple stages, sacrificing computational efficiency to make the obtained overlap more reliable.

[0102] As a preferred example, the methods for reconstructing the weights of fields in the assigned fit category and invalid category for each disease stage include: The fields from the candidate categories in the assigned adaptive categories are denoted as the first field set, and the fields from the candidate categories in the assigned invalid categories are denoted as the second field set. For any field in the first field set, if the accuracy of that field is less than the maximum accuracy of all fields in the second field set, then the accuracy of that field is set to the maximum value. For any field in the second field set, if the accuracy of that field is greater than the minimum accuracy of all fields in the first field set, then the accuracy of that field is set to the minimum value.

[0103] The purpose of this process is to ensure that the accuracy of the assigned fit categories is greater than the accuracy of the assigned invalid categories.

[0104] For the accuracy of all fields in the assigned adapted category and the assigned invalid category, the accuracy of all fields is normalized (using the softmax formula) to obtain the weight of each field (that is, the total weight after reconstruction).

[0105] As a preferred example, reducing the model's generalization ability by differentiating the weights of fields assigned to the fit and invalid categories for each disease stage can be achieved through the following methods: For the model trained at each stage of the disease, the generalization ability of the model is reduced by changing the pruning parameters. The reduction in the pruning parameters is negatively correlated with the weight difference. That is, when the weight difference is larger, there is no need to over-adjust the pruning parameters to avoid significantly reducing the model's generalization ability; when the weight difference is smaller, the pruning parameters need to be adjusted significantly to ensure the model's accuracy or its feature mining ability for clinical data and questionnaire information.

[0106] As an example, when the model uses a fully connected neural network, the pruning parameter is also the dropout parameter (i.e., the regularization parameter). Adjusting the pruning parameter means reducing the dropout parameter, and the reduction is negatively correlated with the weight difference. For example, the dropout parameter can be reduced by r1 × 100%, where r1 = 0.5 × exp(-r0), where r1 represents the reduction magnitude, r0 represents the weight difference, and exp() represents an exponential function with the natural constant as the base. Specifically, when the reduced dropout parameter is less than 0.2, the dropout parameter is set to 0.2.

[0107] As an example, when using a random forest model, the pruning parameter refers to the maximum depth of the decision trees in the random forest. Adjusting the pruning parameter means increasing the maximum depth, and the increase is negatively correlated with the weight difference. For example, increasing the maximum depth by r2 × 100% (rounding up the maximum depth), r2 = 2 × exp(-r0), where r2 represents the increase. Specifically, when the increased maximum depth is greater than 15, the maximum depth is set to 15.

[0108] As an example, the method for obtaining the weight differences of fields assigned to the suitable and invalid categories for each disease stage includes: The difference between the mean (reconstructed) weights of the fields assigned to the appropriate category by the candidate category at each disease stage and the mean (reconstructed) weights of the fields assigned to the invalid category by the candidate category is denoted as the weight difference. As can be seen from the above process, this weight difference is greater than or equal to 0.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent processing, analysis, and feedback system for gastric cancer PROs data, characterized in that, The system includes the following modules: The data acquisition module is used to acquire questionnaire information and clinical data of patients at different stages of their condition; the questionnaire information contains several fields with initial weights. The weight reconstruction module, based on the model trained with the initial weights of each field, obtains the accuracy of the predicted content of each field under each stage of the disease. According to the accuracy, all fields are divided into: suitable category, candidate category, and invalid category. Fields in the candidate category are assigned to the suitable category and invalid category, so that the overlap between the assigned suitable categories under different stages of the disease is maximized, and the fields in the assigned suitable category and invalid category under each stage of the disease have the greatest accuracy difference. Reconstruct the weights of fields in the appropriate and invalid categories assigned for each disease stage; The field optimization module reduces the model's generalization ability based on the weight differences of the fields assigned to the candidate categories in the appropriate and invalid categories under each disease stage. The reconstructed weights are used as the initial weights, and the weight reconstruction module is rerun. During the rerun of the weight reconstruction module, the fields in the appropriate categories after reassignment under each disease stage are used as the fields for feedback questionnaire information under each disease stage. The specific steps involved in assigning fields from candidate categories to suitable and invalid categories are as follows: For each disease stage, the obtained matching category, candidate category, and invalid category are selected. Several fields are selected from the candidate categories and temporarily assigned to the matching category. The remaining fields in the candidate categories are temporarily assigned to the invalid category. Obtain the overlap between the temporarily assigned fit categories under different disease stages; calculate the first difference between the average accuracy of the temporarily assigned fit categories and the average accuracy of the temporarily assigned invalid categories under any disease stage, and record the average of the first differences obtained under all disease stages as the first accuracy difference; obtain the allocation index, which is positively correlated with the first accuracy difference and overlap; when the allocation index is at its maximum, the fields selected from the candidate categories are allocated to the fit categories, and the remaining fields are allocated to the invalid categories.

2. The intelligent processing and analysis feedback system for gastric cancer PROs data according to claim 1, characterized in that, The specific steps for obtaining the adaptation category, candidate category, and invalid category are as follows: For any given stage of the disease, the accuracy of all fields is assessed, with the fields having the highest accuracy assigned to the fit category, the fields having the lowest accuracy assigned to the invalid category, and the remaining fields assigned to the candidate category.

3. The intelligent processing and analysis feedback system for gastric cancer PROs data according to claim 2, characterized in that, The specific steps involved in obtaining the overlap between the temporarily assigned fit categories at different disease stages are as follows: For all disease stages, the temporarily assigned fitting categories are recorded as the baseline category, and the temporarily assigned fitting categories of two adjacent disease stages are referred to as adjacent fitting categories. Any adjacent adapted category of the baseline category is denoted as a 1-neighbor category. The intersection-union ratio of the baseline category and the 1-neighbor categories is obtained, and the intersection of the baseline category and the 1-neighbor categories is denoted as the 1-neighbor intersection. The adjacent adapted categories of a 1-neighbor category are denoted as 2-neighbor categories. The intersection-union ratio of the intersection of 1-neighbor categories and 2-neighbor categories is obtained, and the intersection of the intersection of 1-neighbor categories and 2-neighbor categories is denoted as 2-neighbor intersection. Similarly, after traversing all neighboring categories of the baseline category, the mean of all intersection-union ratios is recorded as the cross-stage association index of the baseline category. The mean of the cross-stage association index of the appropriate categories after temporary assignment for all disease stages is denoted as the overlap degree.

4. The intelligent processing and analysis feedback system for gastric cancer PROs data according to claim 1, characterized in that, The specific steps involved in reconstructing the weights of fields in the assigned appropriate and invalid categories for each disease stage are as follows: The fields from the candidate categories in the assigned adaptive categories are denoted as the first set of fields, and the fields from the candidate categories in the assigned invalid categories are denoted as the second set of fields. When the accuracy of any field in the first set of fields is less than the maximum accuracy of all fields in the second set of fields, the accuracy of that field is set to be equal to that maximum value. If the accuracy of any field in the second set of fields is greater than the minimum accuracy of all fields in the first set of fields, then set the accuracy of that field to be equal to that minimum value.

5. The intelligent processing and analysis feedback system for gastric cancer PROs data according to claim 1, characterized in that, The specific steps for reducing the generalization ability of the model are as follows: The generalization ability of the model is reduced by changing the pruning parameters of the model, and the magnitude of the change in the pruning parameters is negatively correlated with the weight difference.

6. The intelligent processing and analysis feedback system for gastric cancer PROs data according to claim 1, characterized in that, The specific steps involved in training the model based on the initial weights of each field are as follows: Clinical data and questionnaire information for each stage of the disease are used as samples and labels to form a dataset; a model is initialized for each stage of the disease, and the model is trained using the dataset for each stage of the disease. During training, after the samples are input into the model, the model outputs the predicted content to be filled in for each field. The difference between the actual content and the predicted content in each field of the label is recorded as the error of each field. The errors of all fields are weighted and summed based on the initial weights of all fields, and the weighted sum is used as the loss function of the training process.

7. The intelligent processing and analysis feedback system for gastric cancer PROs data according to claim 6, characterized in that, The specific steps for obtaining the accuracy are as follows: At any stage of the disease, the difference between the predicted content of each field output by each patient and the actual content of each patient in each field in the model after inputting the clinical data of each patient into the model is denoted as the first difference of each patient in each field at any stage of the disease; the accuracy is negatively correlated with the median of the first differences of all patients in each field at any stage of the disease.

8. The intelligent processing and analysis feedback system for gastric cancer PROs data according to claim 5, characterized in that, The specific steps involved in changing the pruning parameters of the model are as follows: When the pruning parameter of the model is the dropout parameter, the dropout parameter is reduced, and the reduction is negatively correlated with the weight difference.

9. The intelligent processing and analysis feedback system for gastric cancer PROs data according to claim 5, characterized in that, The specific steps for obtaining the weight differences are as follows: The difference between the average reconstructed weights of the fields assigned to the candidate category and the average reconstructed weights of the fields assigned to the invalid category in each disease stage is denoted as the weight difference.

Images