A patient intention recognition method and system based on medication trajectory

Abstract

This invention belongs to the technical field of digital medical management, specifically relating to a method and system for patient intention recognition based on medication usage trajectory. The method includes: collecting patient medication and feature data, labeling abnormal medication behavior characteristics such as prescriptions from multiple institutions, self-purchase and prescription overlap, parallel emergency and outpatient visits, duplicate prescriptions, and dosage deviations; constructing a user intention scoring model composed of a general evaluation model and a weighted function for easily mutated behaviors, outputting intention scores for follow-up visits, medication purchases, self-adjustment of medication regimens, and hospitalization; determining the intention type based on the scoring results and generating corresponding prompt information. The system includes modules for data acquisition, behavior recognition, intention scoring, intention determination, and prompt information generation. This invention achieves dynamic perception and classification of complex medication behaviors, and is applicable to intelligent health management, risk warning, and personalized intervention scenarios.

Description

Technical Field

[0001] This invention belongs to the technical field of digital medical management, specifically relating to a method and system for identifying patient intentions based on medication trajectory. Background Technology

[0002] With the continuous improvement of medical informatization, in-depth mining of medical big data has become an important means to assist in diagnosis and treatment decisions and improve the quality of medical services. Among them, patients' medication records during their medical treatment, as a true reflection of their disease development, treatment plan selection, and efficacy response, are increasingly regarded as a type of medical trajectory data with behavioral attributes. Through in-depth analysis of patients' medication trajectories, it is hoped that their potential health intentions, treatment preferences, and adherence characteristics can be identified, providing support for precision medicine and intelligent health management.

[0003] Currently, in the field of patient intention recognition, some studies have attempted to construct patient behavior prediction models based on treatment pathways by combining electronic medical record information, questionnaire results, and behavioral labeling methods. Some of these approaches extract keywords from medical records using natural language processing or analyze patient visit frequency and departmental transfers based on time series models to infer disease progression trends and treatment intentions. However, these methods mostly rely on poorly structured text data or lack continuous behavioral data, making it difficult to fully characterize individual patient medication choices over time. Furthermore, the lack of systematic modeling of multidimensional factors such as drug type, dosage changes, and alternating treatments leads to weak generalization ability and poor interpretability in recognizing patients' true treatment preferences and intentions. In addition, because some methods operate on a single visit basis, they fail to establish a complete medication usage trajectory, making it difficult to effectively capture and analyze long-term treatment patterns in patients with chronic diseases, thus affecting the development of subsequent personalized intervention strategies.

[0004] Therefore, there is an urgent need for a method that can identify patients' treatment intentions and health preferences by comprehensively modeling multi-dimensional features such as drug use sequence, frequency, and dosage based on continuous medication behavior, in order to supplement traditional behavior recognition models based on medical records or medical visit paths, and provide data support and technical assurance for precision medicine and proactive health management for patients. Summary of the Invention

[0005] To address the above problems, the present invention aims to propose a method for patient intention recognition based on medication usage trajectory, comprising the following steps:

[0006] S1. Collect patient medication and characteristic data: Obtain patient medication data and patient characteristic data. The medication data includes prescription source, drug name, medication duration, medication frequency and drug dosage. The patient characteristic data includes patient chief symptoms, medical record description, age and gender.

[0007] S2. Labeling abnormal medication use characteristics: Labeling the medication use data for abnormal medication use characteristics, which include at least the following: multi-institution prescription behavior, overlapping of self-purchased and prescription drugs, concurrent prescription of emergency and outpatient drugs, duplicate prescription behavior, and dosage deviation behavior.

[0008] S3. Construct a user intention scoring model: The user intention scoring model includes a general evaluation model and a weighted function for mutable behaviors. The general evaluation model is used to identify medical treatment intentions, and the weighted function for mutable behaviors is used to perform biased weighting on catastrophic behaviors in abnormal medication characteristics and output the scoring results.

[0009] S4. Determine the patient's intention type: Determine the patient's intention type based on the scoring results output by the user intention scoring model. The intention type includes the tendency to return for a follow-up visit, the tendency to purchase medication, the tendency to adjust medication regimen on one's own, and the tendency to be hospitalized.

[0010] S5. Generate intention prompt information: Generate corresponding prompt information according to the intention type. The prompt information includes not recommending purchasing medicine, recommending seeking medical treatment, not recommending adjusting the medication plan, and recommending hospitalization.

[0011] As a preferred technical solution, the user intention scoring model includes a general evaluation model and a weighted function for volatile behaviors. The general evaluation model is a multi-class probabilistic scoring model constructed based on multi-dimensional medication behavior characteristics. Its input features include at least the number of prescriptions, consultation frequency, proportion of self-purchased drugs, number of medication interruptions, medication duration distribution, and disease severity label within a preset time window. The general evaluation model calculates the input features using a logistic regression model or a shallow neural network model, and outputs basic score values ​​corresponding to follow-up visit tendency, medication purchase tendency, tendency to adjust medication regimen independently, and hospitalization tendency, respectively.

[0012] The mutation-prone behavior weighting function is constructed based on the abnormal medication characteristics, including dose deviation behavior, multi-institution prescription behavior, overlapping self-purchase and prescription drug behavior, and parallel prescription behavior in emergency and outpatient departments. A corresponding mutation weight coefficient is set for each type of abnormal medication behavior, and the basic score value is weighted and corrected according to the frequency of occurrence, duration and drug category of the abnormal behavior to obtain the final patient intention score result.

[0013] As a preferred technical solution, the method for identifying multi-institutional prescription behavior in step S2 includes:

[0014] The prescription sources of multiple medication records are grouped and identified. When there are more than three prescription records from different sources within the same time period, and all of the drugs are covered by medical insurance, they are marked as having no clear intention to take the medication or a low tendency to follow up.

[0015] Statistics were compiled on the duration of medication use. When there were two consecutive medication records from different institutions within a certain period of time, and the time interval between medication use was less than one week, and no characteristic medication purchase behavior was observed, it was marked as a tendency to seek follow-up visits.

[0016] The frequency of prescription records from multiple institutions is statistically analyzed to determine whether the same medication was obtained from different institutions more than three times within a month. If so, it is marked as a high tendency to seek follow-up visits; otherwise, it is marked as no clear intention.

[0017] As a preferred technical solution, the method for identifying overlapping behaviors of self-purchased and prescription drugs in step S2 includes:

[0018] Obtain the patient's purchase records after the most recent prescription drug to determine whether there is any overlap in drug names between self-purchased and prescription drugs;

[0019] If there is overlap, determine whether the overlapping drugs belong to the predefined drugs for which the patient tends to purchase. If they do, mark them as drugs for which the patient tends to purchase or drugs for which the patient tends to adjust their medication regimen.

[0020] If it does not fall under this category, it will be marked as a continuing treatment behavior, corresponding to a tendency to seek follow-up visits.

[0021] As a preferred technical solution, the medicines that the user is inclined to purchase include:

[0022] Over-the-counter medications include antipyretics, analgesics, and antihistamines.

[0023] In addition, prescription drugs include antibiotics, hormones, antidepressants, and anthelmintics.

[0024] As a preferred technical solution, the method for identifying dose shift behavior in step S2 includes:

[0025] Analyze patients' medical records to determine their usual daily medication dosages and establish dose offsets. Statistical formula:

[0026]

[0027] in, Indicates dose offset, Indicates the first Daily drug dosage To assess the number of days within the period, This indicates that the daily dosage of medication is being taken according to the doctor's recommendations.

[0028] when When the absolute value is less than a preset threshold, it is marked as no dose offset behavior;

[0029] when When the dose exceeds the preset positive threshold, it is marked as a dose deviation behavior. If the drug used is a drug that the user tends to buy and the frequency of use decreases, it is further marked as a drug purchase tendency behavior.

[0030] when When the dose is less than the preset negative threshold, it is marked as a dose deviation behavior. If the drug used is a addictive psychotropic or hormonal drug and the dosage is reduced, it is further marked as a tendency to self-adjust medication regimen.

[0031] As a preferred technical solution, the method for identifying concurrent prescription refills in emergency and outpatient departments includes:

[0032] Extract medication data from the patient's medical records and sort it according to timestamp;

[0033] Determine whether each medication record falls under the category of emergency or outpatient care, and indicate the nature of the treatment plan.

[0034] Whether concurrent medication is permitted is determined based on the severity of the patient's condition. If permitted, it is further divided into permitted and non-permitted categories according to the type of medication.

[0035] For drugs that cannot be used in parallel, their frequency and duration of use are recorded. If there is an overlap between the medication time and the duration of the next medical visit or emergency room visit, it is marked as a tendency to seek follow-up visits or hospitalization.

[0036] The present invention also provides a patient intention recognition system based on medication trajectory, comprising:

[0037] The data acquisition module is used to acquire patients' medication data and patient characteristic data. The medication data includes prescription source, drug name, medication duration, medication frequency and drug dosage. The patient characteristic data includes patients' chief symptoms, medical record description, age and gender.

[0038] The behavior recognition module is used to label the medication data with abnormal medication characteristics, which include at least the following: multi-institution prescription behavior, overlapping self-purchased and prescription drugs, concurrent prescriptions in emergency and outpatient departments, duplicate prescriptions, and dosage deviation behavior.

[0039] The intention scoring module is used to construct a user intention scoring model. The user intention scoring model includes a general evaluation model and a weighted function for mutable behaviors. The general evaluation model is used to identify medical treatment intentions, and the weighted function for mutable behaviors is used to perform bias weighting on catastrophic behaviors in abnormal medication characteristics and output the scoring results.

[0040] The intention determination module is used to determine the patient's intention type based on the scoring results. The intention type includes revisit intention, medication purchase intention, self-adjustment of medication regimen intention, and hospitalization intention.

[0041] The prompt information generation module is used to generate prompt information based on the intention type. The prompt information includes advice against purchasing medicine, advice to seek medical treatment, advice against adjusting the medication plan, and advice to be hospitalized.

[0042] As a preferred technical solution, the method for identifying concurrent prescription refills in emergency and outpatient departments includes:

[0043] Extract medication data from the patient's medical records and sort it by timestamp;

[0044] Determine whether each medication record comes from the emergency room or outpatient department, and mark its medical scenario attributes;

[0045] Whether concurrent medication is permissible depends on the severity of the condition;

[0046] If permitted, drugs are classified according to their parallel availability, and the frequency and duration of use of drugs that cannot be used in parallel are statistically analyzed. If there is an overlap with the duration of the next medical visit or emergency room visit, it is marked as a tendency to seek follow-up visits or hospitalization.

[0047] As a preferred technical solution, the user intention scoring model specifically includes a general evaluation model and a weighting function for easily mutated behaviors;

[0048] The general assessment model is used to identify the intention to seek medical treatment, the weighting function is used to weight the acute change behavior, and the base score is corrected according to the weighting result to output the patient's intention type;

[0049] The output results are used to assist medical personnel in personalized medication management, health monitoring, and disease early warning services.

[0050] As a preferred technical solution, the intention scoring module is used to perform weighted processing on medication behavior with specific identifiers based on a weighting function after acquiring patient medication data and feature data, including:

[0051] Reduce the preset weight for behaviors such as prescribing medications from multiple institutions, overlapping self-purchased and prescription drugs, concurrent prescribing in emergency and outpatient departments, and duplicate prescribing;

[0052] Increase the preset weight for dose offset behavior;

[0053] When dose offset When the value is less than zero, the upward intention rating is strengthened;

[0054] When dose offset When the value is greater than zero, the downward intention rating is intensified.

[0055] Beneficial effects

[0056] This invention, a patient intention recognition system based on medication trajectory, solves the problem in existing methods that rely solely on static visit tags and cannot characterize dynamic medication behavior by introducing a two-layer structure: "fusion modeling of multiple abnormal medication behaviors" and "general intention scoring model + weighted function for acute behavior changes." In particular, this system, through multi-dimensional characterization of abnormal behaviors such as dose deviation, multi-institution prescriptions, and overlapping self-purchases, not only identifies the type of behavior but also assesses its intensity, achieving a leap from "behavior recognition" to "intention intensity scoring." This scoring method significantly improves the system's sensitivity to trends in patient behavior and enhances its early identification ability for mild non-compliance behaviors.

[0057] This invention employs a collaborative approach combining behavioral semantic rules and probabilistic models to avoid the bias issues inherent in pure algorithmic models under conditions of insufficient training data or label bias. By constructing a mapping function between behavior and intention and introducing a bias weight correction mechanism, this invention achieves dynamic adjustment of intention determination results, enabling the scoring mechanism to not only rely on historical statistics but also respond to short-term, drastic behavioral fluctuations. This strategy can more accurately reconstruct patients' true medical treatment tendencies and improve classification robustness and interpretability when dealing with scenarios involving behavioral mutations (such as overlapping emergency visits or frequent self-purchases) or ambiguous behavioral boundaries (such as the intersection of follow-up visits and adjusted treatment plans).

[0058] Furthermore, the prompt information generation module of this invention can generate personalized intervention suggestions based on the scoring results and behavioral tags, and supports differentiated information push among doctors, patients, and management institutions. This output mechanism not only improves the practical operability of the system, but also transforms traditional risk classification results into a closed-loop information structure oriented towards intervention execution. It is particularly suitable for complex scenarios such as chronic disease management, medical insurance behavior analysis, and health intervention triggering, thereby achieving complete link optimization from identification and judgment to intervention. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0060] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0061] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0062] Example 1

[0063] To make the technical solution of the present invention clearer and more explicit, the following describes in detail a patient intention recognition method based on medication trajectory of the present invention with reference to specific embodiments.

[0064] like Figure 1 As shown, this method systematically collects and analyzes patients' medication data, characteristic data, and behavioral patterns to establish a classification and assessment model, accurately identify patients' medical intentions, and provide targeted prompts. It is applicable to various hospitals, drug sales platforms, internet medical platforms, and public health management systems.

[0065] S1. Collect patient medication and characteristic data:

[0066] In this step, data interfaces are used to connect to the hospital information system (HIS), electronic medical record system (EMR), drug sales platform, medical insurance settlement platform, and online consultation system to collect patients' historical medication behavior and basic characteristic information. The data collection method supports both automatic batch interface synchronization and manual data import to ensure deployment across different medical institutions.

[0067] The medication data specifically includes the following:

[0068] Prescription source: By identifying the medical institution code, doctor's employee number and prescription type (ordinary prescription, prescription for controlled substances, etc.) in the prescription document, it is determined whether it comes from a tertiary hospital, community clinic, internet medical platform or private clinic, etc.

[0069] Drug Name: Includes the generic name, brand name, ingredient information, administration method, packaging specifications, etc., and is uniformly mapped to the National Drug Standards Database;

[0070] Medication duration and frequency: Record the number of days of medication, the number of times of daily medication, and the interval period for each prescription, and analyze their continuity and regularity;

[0071] Drug dosage: Converted to a uniform dosage unit (such as mg / dose or ml / day) for storage, to be used for subsequent dose offset calculations;

[0072] Purchase time and batch number: used to determine the temporal correlation between self-purchase and prescription drug purchase.

[0073] The patient characteristic data includes:

[0074] Chief complaint symptoms and diagnosis results: Extract standardized ICD-10 codes or natural language text descriptions from outpatient records and inpatient summaries to construct a disease dictionary for subsequent model label matching;

[0075] Medical record summary: Includes background information such as past medical history, allergy history, family history, and surgical history, used to help construct a label for the severity of the condition;

[0076] Demographic information: such as gender, date of birth, height, weight, marital and reproductive status, etc.

[0077] Medical insurance information and drug reimbursement ratio: used to distinguish between drug purchasing behavior driven by economic behavior and disease behavior.

[0078] The system preprocesses the raw data, including field standardization, missing value completion, outlier removal, time series normalization, entity deduplication, and feature cleaning, ultimately forming a structured data table. Each patient's medication record is recorded as an independent behavioral event in the behavioral database for subsequent modeling and behavior recognition.

[0079] S2, Marking abnormal medication behavior characteristics

[0080] After completing the collection and structuring of patient medication trajectory data, the system jointly analyzes each medication behavior with the context behavior sequence, identifies abnormal behavior features through predefined rules and behavior templates, assigns labels, and constructs a high-dimensional behavior feature set.

[0081] The annotation of abnormal medication behavior features adopts a combined strategy based on rule-based and semantic feature extraction:

[0082] (1) Prescribing behavior by multiple institutions

[0083] This feature identifies whether a patient has prescription records from three or more different medical institutions within a specific time window (e.g., 7 days), and whether the medications are covered by medical insurance. If such records exist and the medications have overlapping indications or similar treatment pathways, they are marked as "suspected duplicate visits" or "institutional migration behavior."

[0084] Further assessment of whether there is medication continuity (such as multiple institutions continuously prescribing the same drug with an interval of less than 7 days) can preliminarily determine that it is a high-risk characteristic of a tendency to seek follow-up visits.

[0085] (2) Overlapping behavior of self-purchased and prescription drugs

[0086] By comparing patients' self-purchase records at pharmacies with their prescription drug purchase lists, the system identifies whether there is any overlap in drug names or ingredients. If the overlapping drugs are easily accessible over-the-counter medications (such as antipyretics and analgesics) or medications that patients tend to purchase (such as antibiotics and hormones), they are marked as "self-purchased medications".

[0087] If there is a continuous self-purchase of the same prescription drugs at a lower dosage than prescribed, it is presumed to be a behavior of "reduction dependence" or "economic control", which may indicate that the patient has a tendency to actively adjust the medication regimen.

[0088] (3) Prescribing medications in both emergency and outpatient departments

[0089] Identify records of medication prescriptions obtained from both the emergency room and outpatient clinic within a single day or two consecutive days. If the types of medications highly overlap or belong to categories with conflicting treatment logic (such as two different antihypertensive drugs), it is considered that the patient has a behavioral tendency of "inconsistency between urgent medical needs and self-diagnosis," which may indicate an increased risk of hospitalization.

[0090] If such behavior is combined with a record of acute changes in the patient's condition (such as fever, sudden onset of hypertension, etc.), it is further labeled as "high mutation behavior" and used for model weighting.

[0091] (4) Repeated prescription behavior

[0092] If the combination of drug name, dosage form, dosage, and frequency appears repeatedly, and the time interval is shorter than the recommended follow-up visit cycle for the disease (for example, it is usually recommended not to re-prescribe medication within 30 days for chronic diseases), it can be regarded as "repeated drug acquisition". If the repeated behavior occurs in different institutions or different doctors, it is inferred to be "high dependence on follow-up visit behavior".

[0093] (5) Dose offset behavior

[0094] Based on doctor's recommended dosage ( ) and actual purchased dosage ( Calculate dose offset:

[0095]

[0096] If Q is greater than the positive offset threshold (e.g., +20%), it is considered that there is "dose increase" behavior, which may be self-increase in dosage or self-prolongation of efficacy.

[0097] If Q is less than the offset threshold (e.g., -20%), it is considered that there is a "dose reduction" behavior, which may be due to adverse reactions or patient self-judgment.

[0098] If deviation behavior occurs in combination with medication purchase behavior, it suggests a tendency toward "adjustment of treatment plan" or "non-compliance" behavior.

[0099] The system performs one-hot encoding on the above-mentioned abnormal behaviors and combines them into a feature matrix for model training and real-time prediction.

[0100] S3. Construct a user intention scoring model

[0101] This step is the core of the entire methodology. The intention scoring model consists of two main components: a basic general assessment model and a weighted function for abnormal behavior.

[0102] (1) General evaluation model

[0103] The general model is constructed using logistic regression, and its input variables include: the number of prescriptions in the past 30 days; the frequency of visits in the past 30 days (visits / institution); the ratio of self-purchased drugs to prescription drugs; the number of medication interruptions (days); and the standard deviation of the number of days of medication use.

[0104] The input variables include: number of prescriptions in the past 30 days; frequency of visits in the past 30 days (times / institution); ratio of self-purchased drugs to prescription drugs; number of medication interruptions (days); standard deviation of medication duration; and patient disease severity level label (mapped to L1~L3 levels according to ICD-10 or doctor's label).

[0105] The model outputs four intention scores: P1 (tendency to follow up on previous visits); P2 (tendency to purchase medication); P3 (tendency to adjust medication); and P4 (tendency to be hospitalized).

[0106] The output value is the sum of normalized probability values ​​of 1 (Softmax layer output). The training data comes from a dataset of previously identified cases with clear labels, and the model's stability and generalization ability are evaluated through cross-validation.

[0107] The multidimensional medication behavior feature vector extracted by the general assessment model for a given patient within a preset time period is as follows:

[0108]

[0109] in: Number of prescriptions Frequency of visits The proportion of self-purchased medicines, Number of medication interruptions Medication duration distribution characteristics Severity level label;

[0110] The general assessment model is a multi-class probability scoring model, and its output is a basic score vector corresponding to the four patient intention types:

[0111]

[0112] in: Basic score of tendency to follow up Basic score of drug purchase tendency Self-adjustment of medication regimen tendencies baseline score, Hospitalization propensity score.

[0113] When using a logistic regression model, its calculation form is as follows:

[0114]

[0115] in, For the first Feature weight vector corresponding to class intention This is a bias term.

[0116] Formulaic expression of the weighting function for mutagenic behavior:

[0117] Let the set of detected abnormal medication behaviors be:

[0118]

[0119] Abnormal medication behaviors include at least: dosage deviation, prescriptions from multiple institutions, overlapping of self-purchased and prescription drugs, and simultaneous prescriptions from emergency and outpatient departments.

[0120] For each type of abnormal medication behavior Set the corresponding mutation weight coefficient as And define its behavior intensity function as:

[0121]

[0122] in, This indicates the frequency of the abnormal behavior. Indicates the duration of the abnormal behavior; Indicates the drug category factor associated with this abnormal behavior; This is a mapping function used to comprehensively characterize the intensity of abnormal behavior;

[0123] Then for the first The patient intention, its weighted correction term can be expressed as:

[0124]

[0125] in, Indicates the first Abnormal behavior of type 1 The influence coefficient of class intention rating.

[0126] The weighted score vector is:

[0127]

[0128] When multiple abnormal medication behaviors are detected simultaneously, the following preset fusion rules are applied to the correction item:

[0129] (1) Weighting rules:

[0130] (2) Maximum weight rule:

[0131] That is, only the abnormal behaviors that have the greatest impact on the current intention score are retained as the basis for correction.

[0132] Final Patient Intent Score: After combining the output of the general assessment model with weighted corrections for mutagenic behavior, the final patient intent score is obtained.

[0133]

[0134] The This is the final score used to determine a patient's tendency to seek follow-up visits, purchase medication, adjust medication regimens independently, and be hospitalized.

[0135] S4. Determine the patient's intention type

[0136] Based on the resulting four-dimensional scoring vector (P1–P4), the system determines the dominant intention type according to the following strategy:

[0137] If a certain rating accounts for more than 40%, it is considered as the dominant intention;

[0138] If the score differences are not significant, then the behavioral label trigger rules are further combined (e.g., if there is a dose reduction + self-purchase behavior, it tends to be an intention to adjust the plan).

[0139] If mutations are observed, such as emergency room visit combined with high dose, then hospitalization risk is prioritized.

[0140] In addition, the system will compare the results of this intention recognition with the patient's past three behavioral intentions to identify whether a change in behavior has occurred, thereby providing support for subsequent chronic disease intervention or early warning of disease recurrence.

[0141] S5. Generate intention prompt information

[0142] Finally, based on the determined dominant intention type and behavioral characteristics combination, the system generates structured prompt information, including:

[0143] Behavioral attribution: such as "high frequency of self-purchased medication was detected, suspected of self-treatment";

[0144] Risk levels: Low, Medium, and High;

[0145] Intervention recommendations include: "It is recommended to schedule an outpatient follow-up visit" and "It is not recommended to continue taking medications not prescribed by a doctor."

[0146] Notification recipients include: doctors, family members, the patient, and insurance platforms.

[0147] The notification messages are generated using templates and pushed in real time through the hospital app, SMS interface, and medical system assistant. For high-risk patients, the system can also trigger automatic phone reminders or guide a manual follow-up process.

[0148] Example 2

[0149] like Figure 2 As shown, this embodiment provides a patient intention recognition system based on medication trajectory, which is used to collect, model, score and determine the intention of patients' multi-source medication behavior data and personal characteristic information, and finally generate personalized prompt information for medical intervention and health management. It has the characteristics of strong versatility, adjustable model and high processing accuracy.

[0150] The system mainly consists of five core functional modules, including: data acquisition module, behavior recognition module, intention scoring module, intention determination module, and prompt information generation module. The system structure is as follows.

[0151] 1. Data Acquisition Module

[0152] The data acquisition module is the foundational module of this system, responsible for extracting patient-related medication information and basic characteristic information from multiple data sources. This module supports integration with hospital information systems (HIS), electronic medical record systems (EMR), medical insurance settlement platforms, drug retail terminals, remote consultation systems, etc., and has the capability to integrate multi-source heterogeneous data.

[0153] The medication data shall include at least the following fields:

[0154] Prescription source: including the medical institution code that issued the prescription, the doctor's identity, and the prescription type (ordinary prescription, special drug prescription, internet prescription, etc.);

[0155] Drug names include generic name, brand name, and international nonproprietary name (INN), which are standardized through pharmacopoeia numbers;

[0156] Duration of medication: The number of consecutive days of medication recommended in each prescription;

[0157] Medication frequency: the number of times a medication is taken daily, such as once a day (qd), twice a day (bid), three times a day (tid), etc.

[0158] Drug dosage: Calculated based on daily dosage, with standardized units of mg, IU, or ml.

[0159] The patient characteristic data includes at least:

[0160] Chief complaint and diagnostic information: Extract structured ICD-10 codes or natural language encoding from medical records to categorize them into standard disease types;

[0161] Medical record description: includes past medical history, allergy history, medication adherence record, etc.;

[0162] Demographic information: basic parameters such as age, gender, height, and weight;

[0163] Medical insurance status and payment method: used to identify whether there are cost-sensitive driving factors in patient behavior.

[0164] This module features two operating modes: real-time synchronization and periodic batch collection. It supports encrypted data transmission and anonymization, ensuring that data preprocessing, field normalization, missing value imputation, and time standardization are completed under the premise of privacy protection, laying a reliable data foundation for subsequent behavioral analysis and model calculation.

[0165] 2. Behavior recognition module

[0166] This module is responsible for matching behavioral templates and identifying abnormal behavioral patterns in patients' medication behavior based on the cleaned data, and forming behavioral feature labels for intention assessment.

[0167] This module supports the identification of the following types of abnormal medication use behavior:

[0168] Multi-institution prescription behavior identification

[0169] Identify whether there are prescriptions from more than three different medical institutions within the same week;

[0170] Semantic matching of different prescription drug names and categories;

[0171] If a prescription is detected to be repeated frequently across institutions, with the same ingredients, it will be marked as a "high-risk behavior for repeat visits".

[0172] Identification of overlap between self-purchased and prescription drug use:

[0173] Identify whether there is overlap in the ingredients of medication purchased from pharmacies and those prescribed by hospitals;

[0174] Determine whether the drugs involved belong to the list of sensitive drugs such as antipyretics, analgesics, antibiotics, and hormones;

[0175] If there is frequent overlap and no doctor's advice is recorded, it will be marked as "predisposing behavior to purchase medication".

[0176] Emergency and outpatient medication dispensing behavior recognition

[0177] Extract all medication records and sort them by timestamp;

[0178] Mark the visit type (emergency / outpatient) for each record;

[0179] If two types of medical visits occur on the same day, and there is a conflict between the medications, it will be marked as "hospitalization risk behavior".

[0180] Identification of duplicate prescriptions

[0181] Compare whether the types and dosages of drugs in two consecutive prescriptions are consistent;

[0182] Determine if there is any behavior of repeatedly obtaining the same medication within the follow-up visit cycle;

[0183] If the criteria are met, it will be labeled as "dependent follow-up visit behavior".

[0184] Dose offset behavior identification:

[0185] Based on recommended dosage Compared with actual dose Construct the formula for calculating dose offset:

[0186]

[0187] in, Indicates dose offset, Indicates the first Daily drug dosage To assess the number of days within the period, This indicates that the daily dosage of medication is being taken according to the doctor's recommendations.

[0188] when When the (positive offset threshold) is reached, it is marked as "overdose behavior";

[0189] when When the (negative offset threshold) is reached, it is marked as "reduction adjustment behavior".

[0190] All identified behavioral features will be converted into label vectors as input to the intention scoring module, and can also be used for later backtracking analysis.

[0191] 3. Intention Scoring Module:

[0192] This module is the core computational unit of this system, constructing a scoring model for determining patient behavioral tendencies. The scoring model consists of the following two parts:

[0193] (1) General assessment model: Based on the aforementioned structured data input, variables including prescription quantity, frequency of visits, self-purchase ratio, number of interruptions, and severity of illness are extracted;

[0194] The model uses multi-class logistic regression or shallow neural networks (such as single-hidden-layer MLP) to construct the score vector;

[0195] The output is a four-dimensional basic score vector:

[0196]

[0197] in: Basic score of tendency to follow up Basic score of drug purchase tendency Self-adjustment of medication regimen tendencies baseline score, Hospitalization propensity score.

[0198] These represent the basic scores for follow-up visits, medication purchases, treatment plan adjustments, and hospitalizations, respectively.

[0199] (2) Weighting function for mutagenic behavior

[0200] For rapidly changing behaviors such as dose deviation and multi-institutional prescriptions, weighting coefficients are set. ;

[0201]

[0202] The final score result after correction is output as follows:

[0203]

[0204] (3) Weight adjustment mechanism:

[0205] Actions such as multiple institutions, self-purchase, and parallel prescriptions are given negative weights to reduce credibility.

[0206] Severe dose deviation behavior is given a positive weight to improve the sensitivity of behavior determination;

[0207] When dose offset (Reduced behavior) will increase the likelihood of positive feedback (such as follow-up visits).

[0208] when (Increased dosage behavior) leads to a worsening of the rating (such as hospitalization tendency).

[0209] 4. Intention Determination Module:

[0210] Based on the scoring results output by the intention scoring module, this module categorizes multiple behavioral intentions and determines the dominant intention.

[0211] If one component in the four-dimensional score is more than 15% higher than the others, it is directly determined as the dominant intention;

[0212] If the difference is not significant, behavioral characteristics are introduced to assist in the judgment logic (for example, if there is duplicate drug purchase and dosage deviation, the treatment plan is inclined to be adjusted).

[0213] Output intention type tags include: tendency to seek follow-up visits; tendency to purchase medication; tendency to adjust medication regimen on one's own; and tendency to be hospitalized.

[0214] This module has the ability to compare historical behaviors, and can track the same patient's multiple behavioral paths to form a time-series trend of intention evolution.

[0215] 5. Prompt Message Generation Module:

[0216] This module transforms the intent assessment results into human-readable prompts for doctors, patients, or family members to reference. The prompt template consists of intent type + risk level + behavioral summary, for example:

[0217] "The patient has repeatedly requested medications from both the emergency room and outpatient clinic, with a high rate of overlap in the types of medications. A follow-up visit and assessment of the patient's condition are recommended."

[0218] "There is overlap between medication purchases and prescriptions, suggesting self-medication. Doctors are advised to monitor medication adherence."

[0219] "A dose deviation was detected, suggesting a possible adjustment to the treatment plan. We recommend consulting with your doctor."

[0220] This module supports customized push strategies, such as pushing high-risk behaviors to doctors, medium-risk behaviors to patients, and low-risk behaviors to archived observations. The prompts are displayed in text, charts, risk indices, etc., and can be generated as PDF or electronic medical record attachments.

[0221] The system described in this embodiment can be integrated into the server, embedded in hospital diagnosis and treatment terminals, or deployed as a standalone application to achieve automated and intelligent patient intention recognition and medication behavior assistance management. It has significant practical application value and is especially suitable for chronic disease management platforms, medical insurance supervision systems, Internet hospitals, and artificial intelligence-based digital health platforms.

[0222] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying patient intentions based on medication usage patterns, characterized in that, Includes the following steps: S1. Collect patient medication and characteristic data: Obtain patient medication data and patient characteristic data. The medication data includes prescription source, drug name, medication duration, medication frequency and drug dosage. The patient characteristic data includes patient chief symptoms, medical record description, age and gender. S2. Labeling abnormal medication use characteristics: Labeling the medication use data for abnormal medication use characteristics, which include at least the following: multi-institution prescription behavior, overlapping of self-purchased and prescription drugs, concurrent prescription of emergency and outpatient drugs, duplicate prescription behavior, and dosage deviation behavior. Methods for identifying overlap between self-purchased and prescription drug use include: Obtain the patient's purchase records after the most recent prescription drug to determine whether there is any overlap in drug names between self-purchased and prescription drugs; If there is overlap, determine whether the overlapping drugs belong to the predefined drugs for which the patient tends to purchase. If they do, mark them as drugs for which the patient tends to purchase or drugs for which the patient tends to adjust their medication regimen. If it does not belong to this category, it will be marked as a continuous treatment behavior, corresponding to a tendency to seek follow-up visits. Methods for identifying dose offset behavior include: Analyze patients' medical records to determine their usual daily medication dosages and establish dose offsets. Statistical formula: ； in, Indicates dose offset, Indicates the first Daily drug dosage To assess the number of days within the period, This indicates that the daily dosage of medication is being taken according to the doctor's recommendations. when When the absolute value is less than a preset threshold, it is marked as no dose offset behavior; when When the dose exceeds the preset positive threshold, it is marked as a dose deviation behavior. If the drug used is a drug that the user tends to buy and the frequency of use decreases, it is further marked as a drug purchase tendency behavior. when When the dose is less than the preset negative threshold, it is marked as a dose deviation behavior. If the drug used is a psychotropic or hormonal drug with addictive properties and the dosage is reduced, it is further marked as a tendency to self-adjust medication regimen. S3. Construct a user intention scoring model: The user intention scoring model includes a general evaluation model and a weighted function for mutable behaviors. The general evaluation model is used to identify medical treatment intentions, and the weighted function for mutable behaviors is used to perform biased weighting on catastrophic behaviors in abnormal medication characteristics and output the scoring results. The general assessment model is a multi-class probability scoring model constructed based on multi-dimensional medication behavior characteristics. Its input features include at least the number of prescriptions, frequency of visits, proportion of self-purchased drugs, number of medication interruptions, distribution of medication duration, and disease severity label within a preset time window. The general assessment model calculates the input features through a logistic regression model and outputs basic score values ​​corresponding to follow-up visit tendency, medication purchase tendency, tendency to adjust medication regimen on one's own, and hospitalization tendency, respectively. The mutation-prone behavior weighting function is constructed based on the abnormal medication characteristics, including dose deviation behavior, multi-institution prescription behavior, overlapping self-purchase and prescription drug behavior, and parallel prescription behavior in emergency and outpatient departments. A corresponding mutation weight coefficient is set for each type of abnormal medication behavior, and the basic score value is weighted and corrected according to the frequency of occurrence, duration and drug category of the abnormal behavior to obtain the final patient intention score result. S4. Determine the patient's intention type: Determine the patient's intention type based on the scoring results output by the user intention scoring model. The intention type includes the tendency to return for a follow-up visit, the tendency to purchase medication, the tendency to adjust medication regimen on one's own, and the tendency to be hospitalized. S5. Generate intention prompt information: Generate corresponding prompt information according to the intention type. The prompt information includes not recommending purchasing medicine, recommending seeking medical treatment, not recommending adjusting the medication plan, and recommending hospitalization.

2. The patient intention recognition method based on medication trajectory according to claim 1, characterized in that, The method for identifying multi-institution prescription behavior in step S2 includes: The prescription sources of multiple medication records are grouped and identified. When there are more than three prescription records from different sources within the same time period, and all of the drugs are covered by medical insurance, they are marked as having no clear intention to take the medication or a low tendency to follow up. Statistics were compiled on the duration of medication use. When there were two consecutive medication records from different institutions within a certain period of time, and the time interval between medication use was less than one week, and no characteristic medication purchase behavior was observed, it was marked as a tendency to seek follow-up visits. The frequency of prescription records from multiple institutions is statistically analyzed to determine whether the same medication was obtained from different institutions more than three times within a month. If so, it is marked as a high tendency to seek follow-up visits; otherwise, it is marked as no clear intention.

3. The patient intention recognition method based on medication trajectory according to claim 2, characterized in that, The drugs that the purchase preference includes: Over-the-counter medications include antipyretics, analgesics, and antihistamines. In addition, prescription drugs include antibiotics, hormones, antidepressants, and anthelmintics.

4. The patient intention recognition method based on medication trajectory according to claim 1, characterized in that, The method for identifying concurrent prescriptions issued in both emergency and outpatient departments includes: Extract medication data from the patient's medical records and sort it according to timestamp; Determine whether each medication record falls under the category of emergency or outpatient care, and indicate the nature of the treatment plan. Whether concurrent medication is permitted is determined based on the severity of the patient's condition. If permitted, it is further divided into permitted and non-permitted categories according to the type of medication. For drugs that cannot be used in parallel, their frequency and duration of use are recorded. If there is an overlap between the medication time and the duration of the next medical visit or emergency room visit, it is marked as a tendency to seek follow-up visits or hospitalization.

5. A patient intention recognition system based on medication trajectory, used to implement the method as described in any one of claims 1-4, characterized in that, include: The data acquisition module is used to acquire patients' medication data and patient characteristic data. The medication data includes prescription source, drug name, medication duration, medication frequency and drug dosage. The patient characteristic data includes patients' chief symptoms, medical record description, age and gender. The behavior recognition module is used to label the medication data with abnormal medication characteristics, which include at least the following: multi-institution prescription behavior, overlapping self-purchased and prescription drugs, concurrent prescriptions in emergency and outpatient departments, duplicate prescriptions, and dosage deviation behavior. The intention scoring module is used to construct a user intention scoring model. The user intention scoring model includes a general evaluation model and a weighted function for mutable behaviors. The general evaluation model is used to identify medical treatment intentions, and the weighted function for mutable behaviors is used to perform bias weighting on catastrophic behaviors in abnormal medication characteristics and output the scoring results. The intention determination module is used to determine the patient's intention type based on the scoring results. The intention type includes revisit intention, medication purchase intention, self-adjustment of medication regimen intention, and hospitalization intention. The prompt information generation module is used to generate prompt information based on the intention type. The prompt information includes advice against purchasing medicine, advice to seek medical treatment, advice against adjusting the medication plan, and advice to be hospitalized.

6. The patient intention recognition system based on medication trajectory according to claim 5, characterized in that, The method for identifying concurrent prescriptions issued in both emergency and outpatient departments includes: Extract medication data from the patient's medical records and sort it by timestamp; Determine whether each medication record comes from the emergency room or outpatient department, and mark its medical scenario attributes; Whether concurrent medication is permissible depends on the severity of the condition; If permitted, drugs are classified according to their parallel availability, and the frequency and duration of use of drugs that cannot be used in parallel are statistically analyzed. If there is an overlap with the duration of the next medical visit or emergency room visit, it is marked as a tendency to seek follow-up visits or hospitalization.

7. The patient intention recognition system based on medication trajectory according to claim 6, characterized in that, The user intention scoring model specifically includes a general evaluation model and a weighting function for easily mutable behaviors; The general assessment model is used to identify the intention to seek medical treatment, the weighting function is used to weight the acute change behavior, and the base score is corrected according to the weighting result to output the patient's intention type; The output results are used to assist medical personnel in personalized medication management, health monitoring, and disease early warning services.

Images