Internet hospital weight management two-way referral triggering method based on hierarchical diagnosis and treatment

Abstract

The invention relates to the field of medical information, and discloses an internet hospital weight management bidirectional referral triggering method based on hierarchical diagnosis and treatment, which comprises the following steps: comparing continuous physiological parameters, baseline data and a short-term risk score with a preset hierarchical diagnosis and treatment referral rule base; an upward referral suggestion is generated and pushed to a doctor end of the superior medical institution, and the upward referral condition includes that the short-term risk score exceeds a dynamic early warning threshold value or the duration that the continuous physiological parameters continuously exceed a preset red area threshold value reaches a preset first duration; or receiving a confirmation instruction based on the upward referral suggestion or the downward referral suggestion when the weight loss percentage is lower than a preset weight loss target value in a preset intervention period, calling a resource matching algorithm to determine a target medical institution, and generating an electronic referral sheet. According to the scheme of the invention, the congestion of the special number source of the three-first hospital is greatly relieved, and the long-trailing follow-up visit resource of the primary medical institution is activated.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology. Background Technology

[0002] With the continuous improvement of living standards, overweight and obesity have become increasingly prevalent globally, posing a significant risk factor for metabolic syndrome, type 2 diabetes, cardiovascular disease, and even certain malignant tumors, seriously threatening public health. Against this backdrop, internet hospitals, with their convenience and accessibility, have gradually become an important platform for weight management, providing users with services such as remote monitoring, nutritional guidance, and exercise intervention through digital means like mobile applications and wearable devices. However, current mainstream internet-based weight management models still face multiple challenges in practical operation, making it difficult to effectively support a systematic and standardized chronic disease management system.

[0003] First, the hierarchical medical system is generally lacking. There is a lack of efficient collaboration between primary healthcare institutions (such as community health service centers) and higher-level specialized hospitals, resulting in a large number of patients with mild symptoms going directly to tertiary hospitals. This leads to an over-congestion of high-quality medical resources, while the service capacity of primary care institutions remains idle for a long time, which violates the hierarchical medical system principle of "primary care at the primary level, two-way referral, separation of acute and chronic care, and coordinated efforts between different levels".

[0004] Secondly, referral decisions heavily rely on doctors' subjective experience or patients' proactive requests for help, lacking a dynamic, quantitative triggering mechanism based on objective data. This leads to delayed judgment of referral timing and even missed optimal intervention windows. For example, when patients present with complex conditions such as pathological obesity, secondary causes, or the need for weight-loss surgery, the system cannot automatically identify and promptly initiate the upward referral process. Conversely, when patients' conditions stabilize after specialist treatment and they enter the rehabilitation and follow-up stage, there is also a lack of an effective mechanism for downward referral to the community for long-term management. Furthermore, the existing referral process largely relies on telephone communication, paper documents, or unstructured information transmission, resulting in fragmented health records for patients across different levels of medical institutions, creating "information silos." Attending physicians struggle to fully grasp the patient's historical weight change trends, the implementation of intervention measures, and the evolution of key physiological indicators, affecting not only the continuity of diagnosis and treatment but also potentially leading to duplicate examinations or treatment deviations.

[0005] The aforementioned problems collectively constrain the quality, efficiency, and sustainability of weight management services, necessitating a new technological solution that integrates real-time data monitoring, intelligent risk assessment, and automated referral triggering to achieve precise triage, efficient collaboration, and full-cycle closed-loop management of obesity and related metabolic diseases within a tiered medical system. Summary of the Invention

[0006] This invention addresses the common technical shortcomings of existing internet-based weight management models, such as the lack of a tiered medical service mechanism, the lack of quantitative basis for referral decisions, fragmented health data, and unstructured referral processes. It provides a two-way referral triggering method for weight management in internet hospitals based on a tiered medical service system. This method constructs an intelligent referral system integrating multi-source real-time physiological monitoring, a dynamic risk prediction model, a rule-driven decision engine, and an electronic health record access control mechanism. This enables precise, orderly, and traceable two-way referrals of obese patients and those with related metabolic diseases between primary healthcare institutions and higher-level specialized hospitals. A two-way referral triggering method for weight management in internet hospitals based on a tiered medical service system includes the following steps: Step S1: Establish electronic health records for patients in an internet hospital platform that includes primary healthcare institutions and higher-level healthcare institutions, collect baseline data including basic physiological indicators, biochemical test data and past medical history, set an initial management level based on the baseline data and assign a primary healthcare institution to the patient. Step S2: Obtain the continuous physiological parameters collected in real time by the wearable device through the Internet of Things interface, and update the continuous physiological parameters to the electronic health record; Step S3: Extract the sliding window statistical features of the continuous physiological parameters within the preset time window to construct a multidimensional time-series feature vector, input the multidimensional time-series feature vector into a pre-trained risk prediction model based on a long short-term memory neural network, and output a short-term risk score; Step S4: Compare the continuous physiological parameters, the baseline data, and the short-term risk score with the preset hierarchical medical referral rule base: When the conditions for upward referral are met, an upward referral suggestion is generated and pushed to the doctor's end of the superior medical institution. The conditions for upward referral include: the short-term risk score exceeds the dynamic warning threshold, or the duration for which the continuous physiological parameters exceed the preset red zone threshold reaches the preset first duration, or the percentage of weight loss within the preset intervention period is lower than the preset weight loss target value. When the conditions for downward referral are met, a downward referral suggestion is generated and pushed to the doctor's terminal of the primary healthcare institution. The conditions for downward referral include: the patient's continuous physiological parameters after treatment at the higher-level healthcare institution are maintained within the preset target range for more than a preset second duration, or the patient receives a system instruction to change the current treatment plan status to the long-term maintenance phase. Step S5: Receive confirmation instructions based on the upward or downward referral suggestions, call the resource matching algorithm to determine the target medical institution, and generate an electronic referral form.

[0007] Step S6: After the referral is completed, the electronic health record and the corresponding diagnosis and treatment intervention record data packet are synchronized to the workstation of the receiving medical institution, and the modification permission of the transferring medical institution on the electronic health record is changed to read-only permission.

[0008] Step S3 extracts sliding window statistical features of the continuous physiological parameters within a preset time window to construct a multidimensional time-series feature vector, including: The linear regression slope of weight data within a preset number of days is calculated, and the result is set as the weight loss trend feature; the standard deviation of continuous heart rate intervals is calculated, and the result is set as the heart rate variability feature; the average difference between the maximum and minimum blood glucose values ​​within a day is calculated, and the result is set as the blood glucose fluctuation amplitude feature; the weight loss trend feature, the heart rate variability feature, and the blood glucose fluctuation amplitude feature are concatenated and combined to obtain the multidimensional time-series feature vector.

[0009] The dynamic early warning threshold in step S4 is set based on the medical history tags in the electronic health record: When the medical history tag contains a record of diabetes, the dynamic warning threshold is set to a first score; when the medical history tag does not contain a record of diabetes, the dynamic warning threshold is set to a second score, wherein the first score is less than the second score.

[0010] The training process of the pre-trained risk prediction model based on a long short-term memory neural network in step S3 includes: Obtain desensitized patient data, extract multidimensional temporal feature sample sequences from the desensitized patient data to construct a training set; extract records of hyperglycemic crisis, hypotension, or weight management plan failure events that occur within a preset time period after the time window corresponding to the multidimensional temporal feature sample sequence from the desensitized patient data, and use the records as the predicted labels of the training set; input the training set and the corresponding predicted labels into a long short-term memory neural network for parameter iterative training.

[0011] Step S5, which involves calling a resource matching algorithm to determine the target medical institution, includes: When performing an upward referral, the patient's permanent residence coordinates are extracted from the baseline data, the urgency score corresponding to the upward referral recommendation is obtained, and the appointment saturation values ​​of the weight management specialties of each superior medical institution within the Internet hospital platform are obtained. The permanent residence coordinates, the urgency score, and the appointment saturation values ​​are input into a preset recommendation algorithm model to filter out the superior medical institution whose appointment saturation value is lower than a preset load threshold and is closest to the permanent residence coordinates as the target medical institution.

[0012] Step S6, which involves synchronizing the electronic health record and corresponding medical intervention record data packet to the workstation of the receiving medical institution, includes: When a downward referral is performed, the patient's outpatient medical records, discharge summary, and current medication regimen at the higher-level medical institution are packaged into a downward referral data package. The downward referral data package and the electronic health record are then pushed synchronously to the general practitioner's terminal at the primary healthcare institution. Upon receiving the periodic follow-up data entered by the general practitioner's terminal, the periodic follow-up data is appended to the electronic health record stored in the cloud, and read-only access permissions are configured on the doctor's terminal at the higher-level medical institution for the periodic follow-up data.

[0013] The upward referral conditions in step S4 also include dietary behavior triggering rules: The daily dietary compliance score calculated within a preset number of consecutive days is lower than a preset compliance lower limit threshold; or the calculated actual energy intake is lower than the basal metabolic rate calculated based on the patient's baseline data for a preset number of consecutive days, and the monitored weight data shows a trend of decreasing at a rate exceeding a preset threshold.

[0014] The upward referral conditions in step S4 also include event-triggered conditions based on natural language processing: The system acquires consultation text data input by the patient through the internet hospital platform or physical examination report text data uploaded by the patient; it calls a pre-trained natural language processing model to perform entity recognition on the consultation text data or the physical examination report text data and extract symptom keywords; when the symptom keywords match the keywords such as chest tightness, joint injury, or sleep apnea in the preset critical illness terminology database, it triggers the generation of the upward referral suggestion and appends the symptom keywords to the referral summary of the upward referral suggestion.

[0015] In step S2, when acquiring continuous physiological parameters collected in real time by the wearable device, the process also includes abnormal filtering and triggering of heart rate data. The raw heart rate data is filtered using a Kalman filter algorithm to remove motion artifacts and noise. Based on the baseline data, a target heart rate upper limit for the patient is calculated. If the filtered heart rate data continuously exceeds the target heart rate upper limit for a preset warning duration, a warning command is sent to the wearable device, and the heart rate exceedance event is recorded and added to the electronic health record as input data for the risk prediction model.

[0016] The method also includes a step of processing dietary behavior data based on computer vision recognition: The system receives meal images uploaded by the patient and uses a convolutional neural network model to identify food types and estimate food portions. Based on a built-in food composition database, it decomposes the identified food types and portions to calculate the actual energy intake and macronutrient intake for a single meal. The actual energy intake and macronutrient intake are compared with the target values ​​of an individualized dietary plan generated based on the baseline data to calculate a daily dietary adherence score.

[0017] Compared with existing technologies, the two-way referral triggering method for weight management in internet hospitals based on hierarchical medical services provided by this invention has the following significant technological advancements and beneficial effects: Current weight management referral systems heavily rely on extreme values ​​of single vital signs (such as occasional hypertension or hyperglycemia) or the doctor's subjective experience, often resulting in passive referrals only after a substantial deterioration of the patient's condition, exhibiting a significant time lag. This invention creatively extracts sliding-window statistical features such as the slope of linear regression of weight, heart rate variability, and blood glucose fluctuation amplitude within a preset time window, eliminating the random errors of single measurements; these features are then input into a pre-trained long short-term memory neural network. This technique can accurately capture the long-term correlation patterns and subtle deterioration trends of a patient's metabolic state, outputting an objective and quantifiable short-term risk score. This buys valuable clinical time for emergency treatment and specialist intervention.

[0018] After generating referral suggestions, existing technologies often require patients or primary care physicians to find available appointments themselves, which can easily lead to resource congestion at a single well-known tertiary hospital. This invention proactively invokes a pre-set recommendation algorithm model, comprehensively constraining calculations based on the patient's residential location coordinates, the urgency score of their condition calculated using a risk model, and the "appointment saturation values" of various higher-level medical institutions' specialties. This mechanism can automatically filter out medical institutions already operating beyond their capacity and intelligently recommend and lock in the nearest target specialty with available appointments. This not only significantly reduces the time cost and physical distance involved in patient referrals but also achieves intelligent peak-shaving and valley-filling of specialty medical resources at the regional medical consortium network level.

[0019] This invention innovatively introduces computer vision (CV) and natural language processing (NLP) technologies. It uses convolutional neural networks to estimate the actual energy of meal images, transforming the hidden "extreme dieting" into a rigid numerical indicator that triggers mandatory referral. NLP entity recognition technology scans textual consultations and medical examination reports in real time; once it matches critical illness semantic features such as "chest tightness" or "sleep apnea," it immediately triggers the highest-level green channel alarm. Furthermore, the system's underlying Kalman filter algorithm accurately filters out motion artifacts and noise from heart rate, ensuring the accuracy of the warnings. This multimodal, three-dimensional defense greatly compensates for monitoring blind spots in out-of-hospital, offline scenarios, guaranteeing the absolute safety of patients. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of the two-way referral triggering method for weight management in internet hospitals based on hierarchical diagnosis and treatment according to the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0022] System Architecture This invention provides a two-way referral triggering method for weight management in internet hospitals based on hierarchical medical services, which operates on a multi-terminal collaborative internet medical cloud platform. The physical layer and logical architecture of this platform mainly include: The cloud server is equipped with a distributed electronic health record (EHR) database, a referral rule decision engine, a risk prediction model based on long short-term memory neural network (LSTM), a natural language processing (NLP) model, a computer vision (CV) model, and a medical resource recommendation and matching algorithm module. Doctor workstations are divided into general practitioner terminals deployed in primary healthcare institutions (such as community health service centers and township health centers) and specialist physician terminals deployed in higher-level healthcare institutions (such as endocrinology, clinical nutrition, and bariatric and metabolic surgery departments of tertiary hospitals). Patient terminals include smart terminals (such as smartphones) with the Internet Hospital mobile client installed, as well as wearable devices (such as smart bracelets, Bluetooth body fat scales, continuous dynamic blood glucose meters (CGM), smart blood pressure monitors, etc.) that are connected via IoT standard protocols.

[0023] The automated triggering of the main process for two-way referral based on hierarchical medical treatment, such as... Figure 1 ,include: Step S1: Establish health records and initial classification When a patient registers for the first time on the internet hospital platform, the system creates a unique electronic health record for them. The platform collects the patient's baseline data through standardized online questionnaires and interfaces with third-party medical examination institutions' HIS systems. This baseline data specifically includes: basic physiological indicators (height, weight, initial BMI, waist-to-hip ratio, etc.), biochemical test data (fasting blood glucose, glycated hemoglobin, lipid profile, blood pressure, etc.), past medical history (whether there are structured tags such as diabetes, hypertension, polycystic ovary syndrome, cardiovascular disease, etc.), and the patient's permanent residence coordinates (latitude and longitude).

[0024] The system has a built-in hierarchical management model that automatically sets the initial management level and assigns the primary care institution based on the baseline data. For example, if a patient's BMI is between 24 and 28 and there is no history of significant complications, the system sets them to the "green zone (routine primary care management)" and automatically assigns them to the nearest primary care institution as their primary care institution based on their place of residence. If the patient's BMI is ≥32.5 or they have severe metabolic complications, they are marked as "red zone (intensive specialist management)" and are directly assigned to the weight loss specialist department of a higher-level medical institution for their primary care.

[0025] Step S2: Real-time data acquisition and heart rate anomaly filtering processing via IoT The platform provides a standardized Internet of Things (IoT) interface, allowing patients to bind wearable devices via Bluetooth or Wi-Fi. These wearable devices collect continuous physiological parameters from patients in real time and update them frequently to their electronic health records in the cloud.

[0026] During the data collection process, for physiological parameters that are highly susceptible to interference from the environment and patient body movement (such as heart rate), the system implements an anomaly filtering and triggering mechanism: First, after receiving the raw heart rate data uploaded by the wearable device, the system calls the built-in Kalman filter algorithm to perform state estimation and smoothing, accurately filtering out motion artifact noise data caused by violent arm swinging, poor sensor contact, etc., and restoring the true continuous heart rate curve.

[0027] Secondly, the system calculates the absolute safe target heart rate upper limit for exercise based on the patient's baseline data (such as age and resting heart rate) using the Karvonen formula. During routine dynamic monitoring, if the filtered real heart rate data continuously exceeds the target heart rate upper limit for a preset warning duration (e.g., 3 consecutive minutes), the system will immediately send a vibration or beep warning command to the wearable device, forcing the patient to stop high-intensity exercise. At the same time, the system adds this "heart rate exceeding the limit event" as an independent high-risk record to the electronic health record, serving as input data for the risk prediction model in subsequent steps.

[0028] Step S3: Multidimensional Temporal Feature Extraction and LSTM Risk Model Prediction The system does not directly rely on single-point extreme values ​​for disease diagnosis. Instead, it extracts sliding window statistical features of continuous physiological parameters based on a preset time window (e.g., the past 7 or 14 days) to eliminate random errors. The specific calculation logic includes: (1) Weight loss trend characteristics: Extract the daily weight value sequence within the time window, use the least squares method to fit a univariate linear regression equation, and set its slope as the weight loss trend characteristics to accurately quantify whether the weight is in a steady decline, a plateau period, or a rapid rebound state.

[0029] (2) Heart rate variability: The continuous heart rate interval (RR interval) sequence transmitted from the heart rate sensor was extracted, and its standard deviation (SDNN) was calculated and set as the heart rate variability feature. This feature can objectively reflect the patient's cardiovascular autonomic nervous system's ability to regulate metabolic stress and fatigue state.

[0030] (3) Blood glucose fluctuation amplitude characteristics: Extract continuous dynamic blood glucose monitoring (CGM) data, calculate the difference between the maximum and minimum blood glucose values ​​within the day, and obtain the average value (MAGE) of multiple days within the window period, which is set as the blood glucose fluctuation amplitude characteristics.

[0031] The system concatenates the aforementioned weight loss trend features, heart rate variability features, and blood glucose fluctuation amplitude features to obtain a multidimensional time-series feature vector. This vector is then input into a pre-deployed cloud-based risk prediction model based on a Long Short-Term Memory (LSTM) neural network. This model utilizes its ability to remember long-term dependencies on time-series data for forward propagation inference, outputting a "short-term risk score" between 0 and 1, which characterizes the probability of a patient experiencing metabolic deterioration or developing serious complications within a specific future timeframe (e.g., 72 hours).

[0032] Step S4: Automated Determination Mechanism for Two-Way Referral Conditions The rule-based decision engine performs real-time parallel comparisons of continuous physiological parameters, baseline data, and the aforementioned short-term risk scores with the referral rule base. (A) Determination of upward referral conditions (if any one condition is met, a push to the superior doctor's end will be triggered): ① Model-triggered early warning: Short-term risk score exceeds the dynamic early warning threshold. To achieve precision medicine, this dynamic early warning threshold is not a fixed constant, but rather adaptively set based on past medical history tags in the electronic health record. When a patient has vulnerable metabolic tags such as "history of diabetes," the system sets its threshold to the first score (e.g., 0.6); when there is no relevant medical history record, it is set to the second score (e.g., 0.75). The first score is lower than the second score, achieving a more acute risk detection for high-risk groups.

[0033] ② Continuous red zone trigger: The physiological parameters continuously exceed the set absolute red zone threshold (such as systolic blood pressure continuously >160mmHg or fasting blood glucose continuously >11.1mmol / L) for a preset first duration (such as 3 consecutive days).

[0034] ③ Intervention failure trigger: At the end of the preset intervention cycle (such as a 12-week lifestyle intervention) set by the primary care physician, the system automatically calculates the actual weight loss percentage. If it is lower than the preset weight loss target (such as 5%), it proves that the conservative intervention method at the primary care level is ineffective.

[0035] (B) Determination of Downward Referral Conditions (If any one condition is met, the referral will be sent to the primary care physician): ① Physiological stability trigger: After receiving specialized treatment (such as bariatric surgery or GLP-1 receptor agonist therapy) at a higher-level medical institution, the patient's continuous physiological parameters (such as blood glucose, weight, blood lipids, etc.) remain within the preset safety target range for more than the preset second duration (such as 4 consecutive weeks).

[0036] ② System command trigger: Receives a system command issued proactively by a senior specialist doctor on the workstation, changing the current treatment plan status from "intensive intervention phase" to "long-term maintenance phase" or "postoperative recovery period".

[0037] Step S5: Automatic matching of medical resources and generation of electronic referral forms Once the system generates an upward or downward referral suggestion and the doctor confirms acceptance, the system uses a resource matching algorithm to accurately locate the target medical institution.

[0038] Taking the upward referral scenario as an example: the system extracts the patient's permanent residence coordinates from the file and obtains the urgency score of the condition calculated by the risk model (the higher the short-term risk score, the higher the corresponding urgency score); at the same time, the system obtains the current "appointment saturation value" (such as the current percentage of people waiting in line) of the weight management specialty of each higher-level medical institution in real time through the medical consortium data bus.

[0039] The above three types of parameters are input into the preset recommendation algorithm model. The model first performs hard constraint elimination: medical institutions with appointment saturation values ​​exceeding the preset load threshold (e.g., >90%) are screened out to avoid severe overcrowding and long waiting times for patients. Then, in the remaining list of available institutions, the physical straight-line distance to the patient's residence coordinates is calculated (or the commute time is calculated by calling the map API), the nearest superior medical institution is selected as the target medical institution, and appointments are automatically locked for the patient, generating a standardized electronic referral form containing the patient's basic information, the core reason for referral, and the target department.

[0040] Step S6: Closed-loop file transfer and data access control After the referral is completed, the system implements a "data follows the person" and a strict permission handover process at the underlying level, breaking down information silos.

[0041] When a patient is referred downwards, the system automatically retrieves information from the higher-level hospital's HIS system, packaging the patient's outpatient medical records, discharge summaries (including bariatric surgery records), and the latest medication regimen at the higher-level medical institution into a "downward referral data package." This data package, along with the patient's original electronic health record, is simultaneously pushed to the general practitioner's workstation at the receiving institution (primary healthcare facility), enabling primary care physicians to have a comprehensive understanding of the patient's past history of severe illness treatment.

[0042] Subsequently, regular follow-up data generated by patients in the community and at home (such as re-examination vital signs and daily monitoring data) will be incrementally added to the electronic health record stored in the cloud by general practitioners or the patients themselves. To ensure data tamper-proofing and clear accountability, the system is configured with a role-based access control (RBAC) policy, forcibly changing the modification permission of the referrer (the specialist at the higher-level medical institution) to "read-only permission" for the electronic health record. Higher-level experts can use read-only permission to view the long-term efficacy curve of the patient after being transferred to the primary care level at any time, but they do not have the right to modify the current primary care medical orders before initiating a new upward referral, thus forming a tightly structured and clearly defined closed loop for medical data management.

[0043] Offline training process of Long Short-Term Memory Neural Network (LSTM) This embodiment details the process of building and training parameters for the risk prediction model, which is typically performed on an offline server before the system goes live or during a major version update.

[0044] Data preparation and feature engineering: Obtain the historically accumulated de-identified patient database (with names, ID cards, and other private information removed) from the internet hospital platform. Using a time-series sliding window mechanism, extract multi-dimensional temporal features (including the aforementioned weight loss trends, SDNN, MAGE, etc.) within a preset time window from the de-identified patient data, constructing a multi-dimensional temporal feature sample sequence as the input (X) to the model training set.

[0045] Labeling: On the same timeline, within a preset time period following the aforementioned sample window (e.g., the next 72 hours), the electronic medical records of the desensitized patient are searched to determine whether there are any records of "hyperglycemic crisis (e.g., ketoacidosis)," "hypotensive syncope (shock)," or emergency or critical care visits explicitly marked by a doctor as "failure of weight management plan (e.g., inducing a serious cardiovascular event)." If any of these records exist, they are treated as positive samples with a prediction label of 1; if none occur, they are treated as negative samples with a prediction label of 0.

[0046] Model training and iteration: The constructed training set (temporal feature matrix) and corresponding predicted label vectors are input into an LSTM neural network containing multiple hidden layers. The cross-entropy loss function is used to calculate the error between the predicted values ​​and the true labels, and the Adam optimization algorithm is used for backpropagation of parameters and iterative training of weights. To prevent overfitting, a Dropout mechanism is introduced during training. Once the model's accuracy, recall, and AUC on the independent validation set reach the clinically set acceptable standards, training is terminated, the model's network weights are fixed, and the model is encapsulated as a microservice API and deployed to the cloud for real-time invocation in step S3.

[0047] A two-way referral triggering system for weight management in an internet hospital based on hierarchical medical services includes: Data collection unit: Establishes electronic health records for patients in the Internet hospital platform that includes primary healthcare institutions and higher-level healthcare institutions, collects baseline data including basic physiological indicators, biochemical test data and past medical history, sets an initial management level based on the baseline data and assigns the patient's first-visit medical institution; Electronic health record unit: acquires continuous physiological parameters collected in real time by wearable devices through an Internet of Things interface, and updates the continuous physiological parameters to the electronic health record; Risk prediction unit: Extracts sliding window statistical features of the continuous physiological parameters within a preset time window to construct a multidimensional time-series feature vector, inputs the multidimensional time-series feature vector into a pre-trained risk prediction model based on a long short-term memory neural network, and outputs a short-term risk score; Comparison Unit: Compares the continuous physiological parameters, the baseline data, and the short-term risk score with a preset triage and referral rule base. When the conditions for upward referral are met, an upward referral suggestion is generated and pushed to the doctor's end of the superior medical institution. The conditions for upward referral include: the short-term risk score exceeds the dynamic warning threshold, or the duration for which the continuous physiological parameters exceed the preset red zone threshold reaches the preset first duration, or the percentage of weight loss within the preset intervention period is lower than the preset weight loss target value. When the conditions for downward referral are met, a downward referral suggestion is generated and pushed to the doctor's terminal of the primary healthcare institution. The conditions for downward referral include: the patient's continuous physiological parameters after treatment at the higher-level healthcare institution are maintained within the preset target range for more than a preset second duration, or the patient receives a system instruction to change the current treatment plan status to the long-term maintenance phase. Referral form generation unit: Receives confirmation instructions based on the upward or downward referral suggestions, calls a resource matching algorithm to determine the target medical institution, and generates an electronic referral form.

[0048] In this invention, supplementary triggering rules are introduced based on computer vision and natural language processing technologies. Weight management is highly dependent on patients' lifestyle interventions, and conventional numerical physiological parameters often cannot reflect the hidden risks in patients' diet adherence and subjective symptoms in a timely manner. This embodiment makes up for this deficiency by using multimodal artificial intelligence technology.

[0049] 1. Computer vision-based dietary intervention and triggering of referrals for extreme behaviors Before each meal, patients take and upload images of their food using a mobile app. The cloud system then calls upon a deployed convolutional neural network (CNN) image recognition model (such as YOLO or Mask R-CNN architecture) to perform object detection and instance segmentation on the food images. It automatically identifies the types of food in the images (such as rice, oats, broccoli, and pork belly) and combines this information with depth information or reference objects (such as the edge of a standard plate) to perform 3D reconstruction, thereby estimating the accurate portion (volume or weight) of each food item.

[0050] Subsequently, the system queries the built-in national standard food composition database, performs nutritional breakdown on the identified food types and portions, and calculates the actual energy intake (kcal) and precise nutrient intake (carbohydrates, protein, and fat) for that meal. The system compares and verifies the total actual intake for the entire day with the target values ​​of the individualized dietary plan tailored for the patient based on baseline data, and uses a weighted scoring algorithm to calculate the patient's "Daily Dietary Compliance Score" (out of 100 points) for that day.

[0051] Intervention of dietary behavior trigger rules: This score is directly connected to the rule engine in step S4.

[0052] ① Long-term non-compliance trigger: If the calculated daily dietary compliance score is lower than the preset compliance threshold (e.g., 60 points) for a preset number of consecutive days (e.g., 7 days), it indicates that simple lifestyle education at the primary care level has lost its effectiveness, triggering the generation of upward referral recommendations, suggesting that the patient be transferred to the clinical nutrition department of a higher-level hospital for medical nutrition therapy (MNT) or to be prescribed special medical foods.

[0053] ② Extreme Dieting Warning Trigger: If the calculated actual energy intake is significantly lower than the preset number of days (e.g., 3 days), or even lower than the absolute value of the basal metabolic rate (BMR) calculated based on the patient's height and weight, and at the same time, the weight data shows a trend of decreasing at a rate exceeding the preset limit (e.g., a precipitous drop, >3kg per week), the system determines that the patient is in an extreme state of excessive dieting, which is highly likely to induce fatal electrolyte imbalance, hypoglycemic coma, or develop into anorexia nervosa. At this time, the system immediately intercepts this dangerous behavior, automatically generates a referral recommendation, and mandates intervention from a higher-level endocrinology or psychiatry department.

[0054] 2. Emergency symptom interception trigger based on natural language processing In routine management, patients typically input their subjective symptoms of discomfort through the text and image consultation module of the internet hospital platform, or upload photos of their medical examination reports obtained from third-party medical examination institutions (the reports are converted into text using OCR technology).

[0055] The system backend is configured with an asynchronous Natural Language Processing (NLP) scanning service process. This process calls a pre-trained medical vertical domain text analysis model (such as a BERT-based Named Entity Recognition (NER) model) to continuously perform semantic analysis and entity extraction on the aforementioned consultation text data or physical examination report text data, accurately extracting the "symptom keywords".

[0056] The system performs vector space similarity matching between extracted symptom keywords and the built-in "critical and severe illness lexicon." When a match is successfully made with critical illness warning signals such as "chest tightness," "angina pectoris," "severe knee pain," or "nighttime awakening due to shortness of breath (a typical sign of sleep apnea)," the system assigns the highest priority to these events because these complications are sudden and potentially fatal, far exceeding the routine follow-up and treatment capabilities of primary care physicians. At this point, the rule engine directly bypasses the conventional time and model scoring limitations of step S4, instantly triggering the generation of upward referral suggestions. Simultaneously, the system prominently appends these symptom keywords that trigger critical alarms to the upward referral suggestion's referral summary, sending a strong pop-up reminder to the receiving senior physician, thus opening a life-saving green channel for the rescue of high-risk patients.

[0057] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for triggering two-way referrals for weight management in internet hospitals based on hierarchical medical services, characterized in that, Includes the following steps: Step S1: Establish electronic health records for patients in an internet hospital platform that includes primary healthcare institutions and higher-level healthcare institutions, collect baseline data including basic physiological indicators, biochemical test data and past medical history, set an initial management level based on the baseline data and assign a primary healthcare institution to the patient. Step S2: Obtain the continuous physiological parameters collected in real time by the wearable device through the Internet of Things interface, and update the continuous physiological parameters to the electronic health record; Step S3: Extract the sliding window statistical features of the continuous physiological parameters within the preset time window to construct a multidimensional time-series feature vector, input the multidimensional time-series feature vector into a pre-trained risk prediction model based on a long short-term memory neural network, and output a short-term risk score; Step S4: Compare the continuous physiological parameters, the baseline data, and the short-term risk score with the preset hierarchical medical referral rule base: When the conditions for upward referral are met, an upward referral suggestion is generated and pushed to the doctor's end of the superior medical institution. The conditions for upward referral include: the short-term risk score exceeds the dynamic warning threshold, or the duration for which the continuous physiological parameters exceed the preset red zone threshold reaches the preset first duration, or the percentage of weight loss within the preset intervention period is lower than the preset weight loss target value. When the conditions for downward referral are met, a downward referral suggestion is generated and pushed to the doctor's terminal of the primary healthcare institution. The conditions for downward referral include: the patient's continuous physiological parameters after treatment at the higher-level healthcare institution are maintained within the preset target range for more than a preset second duration, or the patient receives a system instruction to change the current treatment plan status to the long-term maintenance phase. Step S5: Receive confirmation instructions based on the upward or downward referral suggestions, call the resource matching algorithm to determine the target medical institution, and generate an electronic referral form.

2. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 1, characterized in that, Step S3 extracts sliding window statistical features of the continuous physiological parameters within a preset time window to construct a multidimensional time-series feature vector, including: Calculate the linear regression slope of weight data within a preset number of days in the past, and set the calculation result as a weight loss trend feature; Calculate the standard deviation of consecutive heart rate intervals and set the calculation results as a characteristic of heart rate variability; Calculate the average difference between the maximum and minimum blood glucose values ​​within a day, and set the calculation result as the blood glucose fluctuation amplitude characteristic; The weight loss trend feature, the heart rate variability feature, and the blood glucose fluctuation amplitude feature are concatenated and combined to obtain the multidimensional time-series feature vector.

3. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 1, characterized in that, The dynamic early warning threshold in step S4 is set based on the medical history tags in the electronic health record: When the medical history tag includes a history of diabetes, the dynamic warning threshold is set to the first score; When the medical history tag does not contain a record of diabetes, the dynamic warning threshold is set to a second score, wherein the first score is less than the second score.

4. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 1, characterized in that, The training process of the pre-trained risk prediction model based on a long short-term memory neural network in step S3 includes: Obtain desensitized patient data, and extract multidimensional temporal feature sample sequences from the desensitized patient data to construct a training set; Extract records of hyperglycemic crisis, hypotension, or weight management plan failure events that occurred within a preset time period after the time window corresponding to the multidimensional temporal feature sample sequence from the desensitized patient data, and use the records as the predicted labels of the training set; The training set and the corresponding predicted labels are input into the long short-term memory neural network for iterative parameter training.

5. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 1, characterized in that, Step S5, which involves calling a resource matching algorithm to determine the target medical institution, includes: When performing upward referral, the patient's permanent residence coordinates are extracted from the baseline data, the urgency score of the condition corresponding to the upward referral suggestion is obtained, and the saturation value of appointment slots for weight management specialties of each superior medical institution within the Internet hospital platform is obtained. The location coordinates of the permanent residence, the urgency score of the illness, and the appointment saturation value are input into a preset recommendation algorithm model to filter out the superior medical institution whose appointment saturation value is lower than a preset load threshold and which is closest to the location coordinates of the permanent residence as the target medical institution.

6. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 1, characterized in that, The method further includes synchronizing the electronic health record and corresponding medical intervention record data package to the workstation of the receiving medical institution, specifically including: When a downward referral is performed, the patient's outpatient medical records, discharge summary, and current medication regimen at the higher-level medical institution are packaged into a downward referral data package. The referral data package and the electronic health record are simultaneously pushed to the general practitioner's terminal at the primary healthcare institution; Upon receiving the regular follow-up data entered by the general practitioner, the regular follow-up data is appended to the electronic health record stored in the cloud, and the doctor's terminal of the superior medical institution is configured with read-only access permission for the regular follow-up data.

7. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 1, characterized in that, The method also includes a step of processing dietary behavior data based on computer vision recognition: Receive food images uploaded by the patient, and call a convolutional neural network model to identify food types and estimate food portions in the food images; Based on the built-in food composition database, the identified food types and quantities are broken down to calculate the actual energy intake and macronutrient intake of a single meal. The daily dietary adherence score is calculated by comparing the actual energy intake and the macronutrient intake with the target values ​​of the individualized dietary plan generated based on the baseline data.

8. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 7, characterized in that, The upward referral conditions in step S4 also include dietary behavior triggering rules: The daily dietary compliance score calculated within a preset number of consecutive days is lower than a preset compliance lower limit threshold; or The calculated actual energy intake is lower than the basal metabolic rate calculated based on the patient's baseline data for a continuous preset number of days, and the monitored weight data shows a trend of decreasing at a rate exceeding the preset limit.

9. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 1, characterized in that, The upward referral conditions in step S4 also include event-triggered conditions based on natural language processing: Obtain consultation text data entered by the patient through the Internet hospital platform or physical examination report text data uploaded by the patient; A pre-trained natural language processing model is invoked to perform entity recognition on the consultation text data or the physical examination report text data, and symptom keywords are extracted. When the symptom keywords match the keywords such as chest tightness, joint injury, or sleep apnea in the preset critical illness terminology database, the upward referral suggestion is generated, and the symptom keywords are appended to the referral summary of the upward referral suggestion.

10. The method for triggering two-way referral for weight management in internet hospitals based on hierarchical medical services according to claim 1, characterized in that, In step S2, when acquiring the continuous physiological parameters collected in real time by the wearable device, the process also includes abnormal filtering and triggering of heart rate data. The raw heart rate data is filtered using the Kalman filter algorithm to remove motion artifact noise data from the raw heart rate data; The upper limit threshold of the patient's target heart rate during exercise is calculated and generated based on the baseline data. When the filtered heart rate data continuously exceeds the target heart rate threshold for exercise for a preset warning duration, a warning command is sent to the wearable device, and the heart rate exceeding the limit event is added to the electronic health record as input data for the risk prediction model.

Images