Matching method and system for same-disease patients based on pathological characteristics and treatment cycles

Abstract

The application discloses a same-disease patient matching method and system based on pathological characteristics and treatment cycles, and the method comprises the following steps: acquiring multi-dimensional data of objective medicine, social demography and platform behavior of a user; pre-processing and extracting features of the multi-dimensional input data, and constructing a multi-dimensional rehabilitation feature model comprising biological, social demographic and behavior cognitive feature vectors; based on dynamic weight matching calculation of the whole treatment cycle, matching and calculating the multi-dimensional rehabilitation feature model of the target user with a candidate user database; according to the calculation result of the matching calculation, combining the calculation result of the time sequence prognosis guidance logic, generating a sorted patient friend recommendation list, and pushing it to the target user. The application can accurately recommend the same-disease patient friends with the same gene mutation, treatment path coverage and survival guidance value for the critical patients by mining multi-dimensional medical data and dynamic behavior characteristics.

Description

Technical Field

[0001] This invention belongs to the field of medical information data processing and Internet community technology, and specifically relates to a method and system for matching patients with the same disease based on pathological characteristics and treatment cycles. Background Technology

[0002] With the development of internet-based healthcare, the demand for outpatient rehabilitation management and communication among fellow patients is increasing among cancer and critically ill patients. Existing general social platforms primarily match users based on location-based services (LBS) or broad interest tags, which is insufficient to meet the deeper needs of critically ill patients seeking fellow sufferers. The concerns of cancer and critically ill patients are diverse, depending on the stage of cancer and the circumstances, including cancer symptoms, post-treatment discomfort, anxiety about the future, treatment costs, and cosmetic concerns. Because these issues are unique and can lead to feelings of isolation, cancer patients can help alleviate their psychological burden by connecting with supportive peers who have experienced similar situations.

[0003] To address this issue, several methods for recommending healthcare communities based on patient information have emerged in the prior art. For example, Japanese patent application JP2025043565A discloses a system that helps patients match with peer support communities. This system receives patient input regarding their medical background, treatment history, and psychological state, combines this information with the patient's set "important matters" (weights), uses morphological analysis and the TF-IDF algorithm to extract keyword feature vectors, calculates the affinity between the patient and peer support communities or topics within those communities, and then recommends a list of communities or topics with high matching scores to the patient.

[0004] However, the aforementioned existing technologies still have the following limitations in practical applications:

[0005] The lack of in-depth structured matching based on pathological dimensions and the coarse-grained matching based on text similarity in existing technologies can easily lead to the aggregation of patients with drastically different treatment plans and prognoses, resulting in low reference value for information exchange and even causing unnecessary psychological panic due to differences in their conditions.

[0006] Current technologies focus on calculating static affinity or similarity, i.e., finding groups of people with similar current states. However, in the process of cancer recovery, newly diagnosed patients often have a more urgent need to find recovered patients or those with longer survival periods as guides, and current technologies lack time-series guidance mechanisms based on treatment cycles. Summary of the Invention

[0007] This invention provides a method and system for matching patients with the same disease based on pathological characteristics and treatment cycles. It aims to solve the problems of existing technologies that mainly rely on text keyword similarity for general community or topic matching, which cannot distinguish heterogeneous groups based on accurate pathological subtyping and genetic characteristics, and lack a prognostic guidance mechanism based on the time sequence of treatment cycles, resulting in low patient matching accuracy, insufficient reference value of rehabilitation experience, and inability to obtain effective survival guidance.

[0008] To address the aforementioned technical problems, this invention proposes a method for matching patients with the same disease based on pathological characteristics and treatment cycles, comprising the following steps:

[0009] Acquire multidimensional input data from target users, including objective medical data, basic attribute data, and platform behavior data;

[0010] The multidimensional input data is preprocessed and features are extracted to construct a multidimensional rehabilitation feature model that includes biological, sociodemographic, and behavioral cognitive feature vectors.

[0011] Based on dynamic weight matching calculation throughout the entire treatment cycle, the multidimensional rehabilitation feature model of the target user is matched and calculated with the candidate user database.

[0012] Based on the results of the matching calculation and the results of the time-series prognosis guidance logic, a sorted list of recommended patients is generated and pushed to the target users.

[0013] Preferably, the method further includes a reinforcement learning mechanism based on user interaction feedback, specifically including:

[0014] Track user interaction behavior after a successful match, defining positive behavior of the target user as a positive reward signal and negative behavior as a negative punishment signal;

[0015] The weight parameters of the multidimensional feature vector are updated by backpropagation using the positive reward signal and the negative penalty signal.

[0016] Preferably, the feature extraction specifically includes:

[0017] Based on medical natural language processing technology, core pathological entities are identified and extracted from objective medical data, including primary cancer type, pathological subtype, TNM stage, gene mutation site and current treatment methods, and biological feature vectors are generated.

[0018] The basic attribute data is normalized to generate sociodemographic feature vectors;

[0019] By analyzing platform behavioral data, we can calculate users' knowledge needs, social activity, and psychological state index, and generate behavioral cognitive feature vectors.

[0020] Preferably, the construction process of the biological feature vector includes medical terminology standardization mapping, synonym disambiguation and normalization, and missing value inference;

[0021] The medical terminology standardization mapping is based on a vertical domain medical knowledge graph constructed using ICD-10, NCCN guidelines, and a drug generic name database, which maps medical terms in objective medical data.

[0022] The synonym disambiguation and normalization will uniformly map heterogeneous text in user-uploaded reports to standard drug codes and uniformly map non-standard pathological descriptions to standard pathological subtype codes.

[0023] When missing pathology data is found, the missing immunohistochemical indicators are inferred in reverse based on the identified medication regimen and treatment pathway logic.

[0024] Preferably, the matching calculation specifically includes:

[0025] Using biological feature vectors as a primary hard constraint, a candidate group whose primary cancer type and core pathological subtype are consistent is screened out;

[0026] In the selected candidate groups, the comprehensive similarity distance is calculated by combining sociodemographic feature vectors and behavioral cognitive feature vectors.

[0027] The matching strategy is dynamically adjusted based on the user's current treatment stage. While searching for the first user in the same treatment stage, the system also retrieves the second user who is in the next treatment stage and has a quality of life score higher than a preset threshold based on prognostic data.

[0028] Preferably, the matching calculation further includes a clustering step based on complications:

[0029] Real-time monitoring of structured symptom logs uploaded by target users, and extraction of current radiotherapy and chemotherapy adverse reactions or drug side effects tags;

[0030] In response to adverse reaction or side effect labels, retrieve recovered users from the candidate user database who have previously recorded the same side effect label and whose quality of life score has improved within a preset time period after the occurrence of the side effect.

[0031] The nursing logs and dietary plans published by the rehabilitation users within the corresponding time period will be given high weight and recommended to the target users along with the rehabilitation users themselves.

[0032] Preferably, the timing prediction guidance logic executes the following screening criteria:

[0033] The diagnosis time or current treatment stage of candidate users must be more than N months ahead of the target users, where N is a preset time threshold.

[0034] Among the candidate users who meet the survival time difference, priority is given to users who have not recorded disease progression events in the most recent M months and whose quality of life scores recorded on the platform are in the top preset proportion as guided recommendation targets;

[0035] Verify whether the candidate user's past treatment path covers the target user's current proposed treatment plan.

[0036] Preferably, the calculation of the comprehensive similarity distance also employs a complementary matching strategy for the behavioral cognitive feature vector, specifically:

[0037] Identify the platform role tags of target users. If the target user is identified as low-activity or knowledge-acquiring, increase the matching weight of high-activity or content-contributing candidate users.

[0038] Based on natural language sentiment analysis, if the recent log sentiment of the target user is negative, the matching priority of candidate users with the same negative sentiment will be reduced, and the matching weight of candidate users with positive sentiment and high similarity of condition will be increased.

[0039] Preferably, the method for adjusting the dynamic weights includes:

[0040] The frequency of occurrence of each pathological feature in the full user database is calculated in real time; if the frequency of occurrence of a pathological feature of a target user in the database is lower than a preset threshold, the weight of the corresponding feature in the matching calculation is increased exponentially.

[0041] Construct a real-time intent heatmap for the target user, capture the dwell time and interaction frequency of the target user in different functional sections within the most recent preset time window, and generate intent vectors; calculate the correlation strength between the intent vectors and each feature dimension, and dynamically shift the matching weights corresponding to the intent vectors.

[0042] On the other hand, the present invention also proposes a matching system for patients with the same disease based on pathological characteristics and treatment cycles. This system is used to implement the matching method described in the first aspect of the present invention, including:

[0043] The data acquisition module is configured to acquire multidimensional input data from the target user, including objective medical data recognized by OCR or synchronized via API interface, basic attribute data filled in by the user, and platform behavior data collected by embedded points.

[0044] The feature engineering and modeling module is configured to call medical natural language processing algorithms and a pre-built vertical domain medical knowledge graph to clean, standardize, map, and extract features from the multidimensional input data, and construct a multidimensional rehabilitation feature model that includes biological, sociodemographic, and behavioral cognitive dimensions.

[0045] The intelligent matching calculation engine is configured to perform vector space distance calculation between the feature model of the target user and the candidate user database based on a dynamic weight strategy throughout the treatment cycle; the engine executes logic including pathological hard constraint screening, comprehensive similarity calculation, and time-series guidance based on prognostic survival.

[0046] The recommendation generation and push module is configured to sort candidate users based on the calculated comprehensive similarity score and the wizard-style filtering results, generate an accurate list of recommended patients, and display it to the target users through the client interface;

[0047] The feedback optimization module is configured to track user interaction with recommendation results, calculate positive and negative feedback signals, and use reinforcement learning algorithms to update the feature weight parameters in the intelligent matching calculation engine online.

[0048] Compared with the prior art, the present invention has the following technical effects:

[0049] 1. The matching method proposed in this invention introduces biological feature vectors as hard constraints, enabling deep matching based on primary cancer type, pathological subtype, gene mutation site, and TNM stage. This not only filters out invalid social interactions but also avoids information noise and unnecessary psychological panic caused by significant differences in medical conditions, ensuring the medical relevance and reference value of the communication content.

[0050] 2. The matching method proposed in this invention introduces a timing misalignment matching strategy for the treatment cycle. By screening out guide-type users who have been diagnosed earlier and have excellent quality of life, it recommends them to users who are newly diagnosed or in a period of confusion about treatment. This not only achieves the effective transfer of recovery experience, but also builds new patients' confidence in fighting cancer through real long-term survival cases, reflecting the platform's purpose of extending survival and improving quality of life.

[0051] 3. The matching method proposed in this invention dynamically aggregates patients who are experiencing or have successfully overcome the same adverse reactions by monitoring users' symptom logs and side effect tags in real time. This dynamic clustering based on specific rehabilitation pain points transforms social behavior into actual nursing mutual assistance, significantly improving patients' quality of life during out-of-hospital rehabilitation.

[0052] 4. The matching method proposed in this invention comprehensively considers users' psychological state and knowledge needs, and improves the community atmosphere through a complementary matching strategy. Simultaneously, it introduces a reinforcement learning mechanism, utilizing user interaction feedback to correct feature weights in reverse, thus solving the problem of rigidity in traditional rule-based matching. This allows the recommendation system to continuously optimize itself as users' lifecycles change and platform data accumulates. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the matching method described in this invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0055] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0056] OCR stands for Optical Character Recognition. It refers to the process by which electronic devices (such as scanners or digital cameras) examine printed characters on paper, determine their shapes by detecting dark and light patterns, and then translate the shapes into computer text using character recognition methods. In other words, for printed characters, it uses optical methods to convert the text in a paper document into a black-and-white dot matrix image file, and then uses recognition software to convert the text in the image into text format for further editing and processing by word processing software.

[0057] TNM staging: The TNM staging system is a system for staging malignant tumors proposed by French physician Pierre Denoix between 1943 and 1952. The American Joint Committee on Cancer (AJCC) and the Union for International Cancer Control (UICC) jointly developed the international standards for this system. This system divides tumors into three dimensions: T (primary tumor extent), N (regional lymph node metastasis), and M (distant metastasis). T is divided into T1-T4, N into N0-N3, and M into M0-M1. These combined dimensions correspond to clinical stages I-IV. Specific standards vary depending on the tumor type.

[0058] LBS: Short for Location Based Services, it uses various types of positioning technologies to obtain the current location of a positioning device and provides information resources and basic services to the positioning device through the mobile Internet.

[0059] NLP: Short for Natural Language Processing, it is an important research direction in the field of artificial intelligence. It integrates knowledge from multiple disciplines such as linguistics, computer science, machine learning, mathematics, and cognitive psychology. It includes two main aspects: natural language understanding and natural language generation. The research content includes multiple levels such as characters, words, phrases, sentences, paragraphs, and texts.

[0060] ICD-10: International Statistical Classification of Diseases and Related Health Problems (10th Revision), or International Statistical Classification of Diseases and Related Health Problems, tenth edition, is a system developed by the World Health Organization (WHO) to classify diseases according to certain characteristics and using a coding method to represent them.

[0061] NCCN: National Comprehensive Cancer Network.

[0062] Example 1

[0063] This embodiment describes a method for matching patients with the same disease based on pathological characteristics and treatment cycles. Figure 1 As shown, it includes the following steps one through four:

[0064] Step 1: Obtain multidimensional input data from the target user, including objective medical data, basic attribute data, and platform behavior data.

[0065] In this embodiment, the acquisition and analysis of objective medical data is the core of building an accurate matching model and has the highest matching weight. Sources include, but are not limited to, photos of paper pathology reports uploaded by users, electronic reports returned by gene testing companies, and discharge summaries photographed by users. The system calls an OCR engine equipped with a vertical domain medical thesaurus to perform text recognition and structured extraction on the image files.

[0066] Basic attribute data serves as a secondary filtering condition to improve the comfort and practicality of social matching, with a weight set to medium. Data sources include the user's personal profile entered on the registration guide page, anonymized information returned from the real-name authentication interface, and the user's selected medical insurance type.

[0067] Platform behavioral data is dynamic and time-varying, used to fine-tune the ranking of the recommendation list, with weights dynamically adjusted throughout the user's lifecycle. Data sources include end-to-end tracking data from the platform's app / mini-program, including clickstream logs, search terms entered in the search box, content browsing duration, and posted / comment content.

[0068] Step 2: Preprocess and extract features from the multidimensional input data to construct a multidimensional rehabilitation feature model that includes biological, sociodemographic, and behavioral cognitive feature vectors.

[0069] In this embodiment, the feature engineering module deployed on the server side is used for execution, specifically including the construction and deep processing of biological feature vectors, the generation of sociodemographic feature vectors, and the generation process of behavioral cognitive feature vectors.

[0070] The construction and deep processing of biological feature vectors are the core steps in building an accurate matching model. The system first uses medical natural language processing technology to identify and extract core pathological entities from objective medical data, including primary cancer type, pathological subtype, TNM stage, gene mutation site, and current treatment methods, generating biological feature vectors.

[0071] To ensure that the extracted features can be accurately understood and computed by the computer, the system needs to address the issues of ambiguity, heterogeneity, and missing values ​​in medical data. Specifically, the construction of the biological feature vectors includes medical terminology standardization mapping, synonym disambiguation and normalization, and missing value inference.

[0072] Regarding standardized mapping: The system performs mapping operations through a pre-built knowledge base engine. The standardized mapping of medical terms maps medical terms in objective medical data based on a vertical domain medical knowledge graph constructed from ICD-10, NCCN guidelines, and a generic drug name database; for example, mapping "C34.1" identified in a user-uploaded report to "malignant tumor of the upper lobe of the lung".

[0073] Regarding synonym disambiguation and normalization: For non-standard expressions in user input or OCR recognition, the synonym disambiguation and normalization uniformly map heterogeneous text in user-uploaded reports to standard drug codes, and uniformly map non-standard pathological descriptions to standard pathological subtype codes; for example, the brand name "Iressa" or chemical name "Gefitinib" entered by the user is uniformly mapped to a unique universal drug code; and "signet ring cell carcinoma" is uniformly classified as a high-grade adenocarcinoma subtype code.

[0074] Regarding missing value inference: For cases where user medical data is incomplete, the system employs a logical reasoning algorithm to complete the feature vector. When missing pathology data is found, the missing immunohistochemical markers are inferred backwards based on the identified medication regimen and treatment pathway logic. For example, if the system identifies that a user is using trastuzumab for treatment, but the HER-2 status is missing in the pathology data, the system infers that the user's HER-2 marker is positive based on the NCCN guidelines and automatically fills in the feature vector.

[0075] The purpose of the sociodemographic feature vector generation process is to measure the similarity of users in real-life dimensions. The system normalizes the basic attribute data to generate sociodemographic feature vectors. In specific implementation, continuous values ​​such as age are mapped to the [0,1] interval, and discrete data such as geographical location and medical insurance type are transformed into sparse vectors to facilitate subsequent calculation of Euclidean distance.

[0076] The process of generating behavioral cognitive feature vectors aims to achieve dynamic complementary matching based on user psychology and cognition. The system analyzes platform behavioral data to calculate users' knowledge needs, social activity levels, and psychological state indices, generating behavioral cognitive feature vectors. Specifically, the system weights the frequency of users' search keywords, the sentiment of their posts, and browsing time to quantify users' current psychological resilience and social willingness, thus providing dynamic weighting for the ranking algorithm.

[0077] Step 3: Based on dynamic weight matching calculation throughout the treatment cycle, the multidimensional rehabilitation feature model of the target user is matched with the candidate user database. In this embodiment, the matching calculation is not a simple linear weighting, but rather employs a strategy combining hierarchical filtering and multidimensional dynamic adjustment.

[0078] The system first performs a funnel-shaped screening process, the matching calculation specifically being:

[0079] Using biological feature vectors as a primary hard constraint, a candidate population with the same primary cancer type and core pathological subtype is screened out; to ensure that lung cancer patients will not be matched with gastric cancer patients, and EGFR mutation patients will be preferentially matched with the same mutation carriers;

[0080] In the selected candidate group, the comprehensive similarity distance is calculated by combining sociodemographic feature vectors and behavioral cognitive feature vectors; thus, potential patients of similar age and family background are found among many patients with the same disease.

[0081] To prevent the matching algorithm from becoming rigid, the system introduces adaptive weight adjustment to ensure that rare cases and immediate intentions are given priority.

[0082] The method for adjusting the dynamic weights includes:

[0083] The system calculates the frequency of occurrence of various pathological features in the full user database in real time. If the frequency of occurrence of a pathological feature of a target user in the database is lower than a preset threshold, the weight of the corresponding feature in the matching calculation is increased exponentially. For example, if a user has an extremely rare "ROS1 fusion", the system will ignore differences such as age and region and force priority matching with rare users who also have ROS1 fusion.

[0084] Construct a real-time intent heatmap for the target user, capture the dwell time and interaction frequency of the target user in different functional sections within the most recent preset time window, and generate intent vectors; calculate the correlation strength between the intent vector and each feature dimension, and dynamically shift the matching weight corresponding to the intent vector; for example, if the user frequently browses the "Out-of-town Medical Insurance Reimbursement" section, the system will temporarily increase the matching weight of the "Geographical Region" and "Medical Insurance Type" features.

[0085] To build a healthy community ecosystem, the system not only seeks similar individuals but also complementary ones. The calculation of the comprehensive similarity distance also employs a complementary matching strategy for behavioral cognitive feature vectors, specifically:

[0086] Identify the platform role tags of target users. If the target user is identified as low-activity or knowledge-acquiring, increase the matching weight of high-activity or content-contributing candidate users; thereby driving the participation of low-activity users.

[0087] Based on natural language sentiment analysis, if the recent log sentiment of a target user is negative, the matching priority of candidate users with the same negative sentiment will be reduced, while the matching weight of candidate users with positive sentiment and high similarity of symptoms will be increased. Through this sentiment hedging mechanism, negative energy resonance is avoided, and psychological healing within the community is achieved.

[0088] The system also utilizes time-based data to implement a mentorship mechanism for experienced patients. It dynamically adjusts the matching strategy based on the user's current treatment stage, simultaneously identifying the first user in the same stage and, based on prognostic data, searching for a second user in the next stage with a quality of life score above a preset threshold. This allows users to receive both peer support and guidance from rehabilitation role models.

[0089] The system has an independent response mechanism for sudden physical discomfort during treatment. In this embodiment, the matching calculation also includes a clustering step based on complications:

[0090] Real-time monitoring of structured symptom logs uploaded by target users, and extraction of current radiotherapy and chemotherapy adverse reactions or drug side effects tags;

[0091] In response to adverse reaction or side effect labels, retrieve recovered users from the candidate user database who have previously recorded the same side effect label and whose quality of life score has improved within a preset time period after the occurrence of the side effect.

[0092] The nursing logs and dietary plans published by the rehabilitation users within the corresponding time period will be given high weight and recommended to the target users along with the rehabilitation users themselves.

[0093] Step 4: Based on the calculation results of the matching calculation (i.e. the comprehensive similarity distance calculated in Step 3), combined with the calculation results of the temporal prognosis guidance logic, generate a sorted list of recommended patients and push it to the target users.

[0094] To provide valuable rehabilitation benchmarks for target users who are in the initial diagnosis or treatment uncertainty stage, the system runs a specialized screening algorithm. The time-series prognostic guidance logic executes the following screening criteria:

[0095] The system reads the candidate users' medical record timestamps and sets the candidate users' diagnosis time or current treatment stage to be at least N months ahead of the target user's, where N is a preset time threshold. If the target user is in the first month after surgery, the system sets N=6, filtering out candidate users who have been post-surgery for at least six months. This ensures that the recommended candidates have already passed through the acute phase that the target user is currently experiencing and have the experience of someone who has been through it.

[0096] Simply having a time lead is not enough; the system must ensure that the recommended candidates are in a good progression-free survival state. Among the candidate users who meet the survival time difference requirement, priority is given to users who have not recorded any disease progression events in the most recent M months and whose quality of life scores recorded on the platform are in the top preset proportion as guided recommendation candidates; the system retrieves the follow-up records of candidate users to confirm that no keyword disease progression, recurrence, or metastasis has occurred in the most recent M (e.g., M=3) months; at the same time, it reads their most recently completed quality of life questionnaire score and only retains users whose scores rank in the top 20% (preset proportion) of the entire platform.

[0097] To ensure the replicability of experience, the system compares the treatment profiles of both users. It verifies whether the candidate user's past treatment history covers the target user's current proposed regimen. If the target user's proposed regimen is docetaxel chemotherapy, the system checks the candidate user's historical treatment logs. The verification passes only if the candidate user's history explicitly includes docetaxel and the course of treatment has been completed.

[0098] The system performs a weighted fusion of the comprehensive similarity distance calculated in step three with the guide-type recommendation objects that emphasize experience transmission and are selected from the vertical matching above.

[0099] The generated recommendation list typically consists of two parts: the first 30% are high-weighted guide users, and the latter 70% are highly similar patients from the same period.

[0100] Final score .in, This is the horizontal similarity weight coefficient, used to adjust the proportion of similarity in the final score; This is the longitudinal guidance weighting coefficient, used to adjust the proportion of rehabilitation guidance value in the final score, satisfying... ; The overall similarity is calculated in step three; This is the timing prognostic guidance score for this step.

[0101] The server will push the generated recommendation list to the target user's terminal through an encrypted channel and display it on the target user's App interface.

[0102] Existing matching algorithms often face problems such as cold start or static stagnation. To enable the system to self-evolve as platform data accumulates, this embodiment further includes step five: a reinforcement learning mechanism based on user interaction feedback, specifically:

[0103] After the recommendation list is pushed out, the system activates a targeted behavior listener to collect real-time feedback from target users on the recommendation results. The system tracks user interactions after a successful match, defining positive behaviors as positive reward signals and negative behaviors as negative penalty signals.

[0104] To ensure the computational feasibility of the algorithm, this embodiment provides a specific quantization definition for the aforementioned signals:

[0105] Positive reward signal (Reward, ):

[0106] Strong signals (+1.0): For example, the target user clicks to request to add as a friend, actively initiates a private message conversation with more than 3 rounds in the first round of conversation, or stays on the candidate user's homepage for a long time (stay time > preset threshold);

[0107] Weak signal (+0.5): For example, the target user clicked to view the candidate user's detailed information, but did not initiate further interaction.

[0108] Negative penalty signal (Penalty, ):

[0109] Strong signal (-1.0): such as the target user clicking "not interested", blocking the candidate user, or reporting the candidate user;

[0110] Weak signal (-0.5): such as the target user quickly swiping past the recommended card (exposure time < 1 second), or ending the conversation within 10 seconds after initiating it.

[0111] The system constructs a reinforcement learning agent based on policy gradient or deep Q-network, whose goal is to maximize long-term cumulative reward. At the end of each time step or batch processing cycle, the system calculates the loss function between the predicted matching degree and the actual user feedback. The system then uses the positive reward signal and negative penalty signal to backpropagate and update the weight parameters of the multidimensional feature vector.

[0112] The specific implementation process is as follows:

[0113] When a negative penalty signal is received, the system analyzes the dominant feature that caused the match to fail. For example, if a target user (City A, Phase II) repeatedly rejects candidates recommended by the system in the same city but at different phases (City A, Phase IV), the system identifies that the TNM phase feature is more important than geographical location in the user's decision-making.

[0114] The system dynamically adjusts the weight vector of the feature combination using the gradient descent algorithm.

[0115]

[0116] In the example above, the system will automatically reduce the weight of sociodemographic characteristics. And accordingly increase the weight of biological characteristics (such as TNM staging). Other characteristics such as The weights remain unchanged.

[0117] After multiple iterations, the system solidifies a set of personalized preference parameters for each user, thereby providing more accurate recommendations that better meet the user's subconscious needs in the matching calculation in step three.

[0118] Example 2

[0119] This embodiment is a patient matching system based on pathological characteristics and treatment cycles. The system is used to implement the matching method as described in the first aspect of the present invention, including:

[0120] The data acquisition module is configured to acquire multidimensional input data from the target user, including objective medical data recognized by OCR or synchronized via API interface, basic attribute data filled in by the user, and platform behavior data collected by embedded points.

[0121] The feature engineering and modeling module is configured to call medical natural language processing algorithms and a pre-built vertical domain medical knowledge graph to clean, standardize, map, and extract features from the multidimensional input data, and construct a multidimensional rehabilitation feature model that includes biological, sociodemographic, and behavioral cognitive dimensions.

[0122] The intelligent matching calculation engine is configured to perform vector space distance calculation between the feature model of the target user and the candidate user database based on a dynamic weight strategy throughout the treatment cycle; the engine executes logic including pathological hard constraint screening, comprehensive similarity calculation, and time-series guidance based on prognostic survival.

[0123] The recommendation generation and push module is configured to sort candidate users based on the calculated comprehensive similarity score and the wizard-style filtering results, generate an accurate list of recommended patients, and display it to the target users through the client interface;

[0124] The feedback optimization module is configured to track user interaction with recommendation results, calculate positive and negative feedback signals, and use reinforcement learning algorithms to update the feature weight parameters in the intelligent matching calculation engine online.

[0125] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for matching patients with the same disease based on pathological characteristics and treatment cycles, characterized in that, Includes the following steps: Acquire multidimensional input data from target users, including objective medical data, basic attribute data, and platform behavior data; The multidimensional input data is preprocessed and features are extracted to construct a multidimensional rehabilitation feature model that includes biological, sociodemographic, and behavioral cognitive feature vectors. Based on dynamic weight matching calculation throughout the entire treatment cycle, the multidimensional rehabilitation feature model of the target user is matched and calculated with the candidate user database. Based on the calculation results of the matching calculation, combined with the calculation results of the time-series prognosis guidance logic, a sorted list of recommended patients is generated and pushed to the target users. Specifically, the feature extraction involves: identifying and extracting core pathological entities from objective medical data based on medical natural language processing technology, including primary cancer type, pathological subtype, TNM stage, gene mutation site, and current treatment methods, to generate a biological feature vector; normalizing the basic attribute data to generate a sociodemographic feature vector; and calculating the user's knowledge demand, social activity, and psychological state index by analyzing platform behavior data to generate a behavioral cognitive feature vector. The construction and processing of the biological feature vectors includes medical terminology standardization mapping, synonym disambiguation and normalization, and missing value inference. The medical terminology standardization mapping maps medical terms in objective medical data based on a vertical domain medical knowledge graph constructed from ICD-10, NCCN guidelines, and a generic drug name database. The synonym disambiguation and normalization uniformly maps heterogeneous text in user-uploaded reports to standard drug codes and non-standard pathological descriptions to standard pathological subtype codes. When missing pathological data exists, the missing immunohistochemical indicators are inferred in reverse based on the identified medication regimen and treatment path logic. The matching calculation specifically involves: using biological feature vectors as a primary hard constraint to screen out candidate groups whose primary cancer type and core pathological subtype are consistent; among the screened candidate groups, combining sociodemographic feature vectors and behavioral cognitive feature vectors to calculate the comprehensive similarity distance; dynamically adjusting the matching strategy according to the user's current treatment stage, while searching for the first user in the same treatment stage, and based on prognostic data retrieving the second user in the next treatment stage with a quality of life score higher than a preset threshold.

2. The method according to claim 1, characterized in that, The method also includes a reinforcement learning mechanism based on user interaction feedback, specifically including: Track user interaction behavior after a successful match, defining positive behavior of the target user as a positive reward signal and negative behavior as a negative punishment signal; The weight parameters of the multidimensional feature vector are updated by backpropagation using the positive reward signal and the negative penalty signal.

3. The method according to claim 1, characterized in that, The matching calculation also includes a clustering step based on complications: Real-time monitoring of structured symptom logs uploaded by target users, and extraction of current radiotherapy and chemotherapy adverse reactions or drug side effects tags; In response to adverse reaction or side effect labels, retrieve recovered users from the candidate user database who have previously recorded the same side effect label and whose quality of life score has improved within a preset time period after the occurrence of the side effect. The nursing logs and dietary plans published by the rehabilitation users within the corresponding time period will be given high weight and recommended to the target users along with the rehabilitation users themselves.

4. The method according to claim 1, characterized in that, The calculation of the comprehensive similarity distance also employs a complementary matching strategy for behavioral cognitive feature vectors, specifically: Identify the platform role tags of target users. If the target user is identified as low-activity or knowledge-acquiring, increase the matching weight of high-activity or content-contributing candidate users. Based on natural language sentiment analysis, if the recent log sentiment of the target user is negative, the matching priority of candidate users with the same negative sentiment will be reduced, and the matching weight of candidate users with positive sentiment and high similarity of condition will be increased.

5. The method according to claim 1, characterized in that, The timing prediction guidance logic executes the following screening criteria: The diagnosis time or current treatment stage of candidate users must be more than N months ahead of the target users, where N is a preset time threshold. Among the candidate users who meet the survival time difference, priority is given to users who have not recorded disease progression events in the most recent M months and whose quality of life scores recorded on the platform are in the top preset proportion as guided recommendation targets; Verify whether the candidate user's past treatment path covers the target user's current proposed treatment plan.

6. The method according to claim 1, characterized in that, The method for adjusting the dynamic weights includes: The frequency of occurrence of each pathological feature in the full user database is calculated in real time; if the frequency of occurrence of a pathological feature of a target user in the database is lower than a preset threshold, the weight of the corresponding feature in the matching calculation is increased exponentially. Construct a real-time intent heatmap for the target user, capture the dwell time and interaction frequency of the target user in different functional sections within the most recent preset time window, and generate intent vectors; calculate the correlation strength between the intent vectors and each feature dimension, and dynamically shift the matching weights corresponding to the intent vectors.

7. A patient matching system based on pathological characteristics and treatment cycles, characterized in that, The system is used to implement the matching method as described in any one of claims 1-6, including: The data acquisition module is configured to acquire multidimensional input data from the target user, including objective medical data recognized by OCR or synchronized via API interface, basic attribute data filled in by the user, and platform behavior data collected by embedded points. The feature engineering and modeling module is configured to call medical natural language processing algorithms and a pre-built vertical domain medical knowledge graph to clean, standardize, map, and extract features from the multidimensional input data, and construct a multidimensional rehabilitation feature model that includes biological, sociodemographic, and behavioral cognitive dimensions. The intelligent matching calculation engine is configured to perform vector space distance calculation between the feature model of the target user and the candidate user database based on a dynamic weight strategy throughout the treatment cycle; the engine executes logic including pathological hard constraint screening, comprehensive similarity calculation, and time-series guidance based on prognostic survival. The recommendation generation and push module is configured to sort candidate users based on the calculated comprehensive similarity score and the wizard-style filtering results, generate an accurate list of recommended patients, and display it to the target users through the client interface; The feedback optimization module is configured to track user interaction with recommendation results, calculate positive and negative feedback signals, and use reinforcement learning algorithms to update the feature weight parameters in the intelligent matching calculation engine online.

Images