Smart medical information pushing method and system applied to Internet hospital

Abstract

The invention discloses an intelligent medical information pushing method and system applied to an internet hospital, and the method comprises the steps: obtaining the real-time state data, including physiological monitoring, terminal behaviors and current diagnosis and treatment stage data, of a target patient; calculating the current state stability of the patient through a state evaluation model based on the real-time state data; when to-be-pushed information exists, dynamically determining a recommendation strategy including a pushing opportunity, a channel and a content rendering mode according to the state stability, the information attribute and historical feedback; and generating and sending the adapted information according to the strategy. According to the method, the state of the patient is perceived in real time and the pushing strategy is intelligently decided, so that the conversion from'broadcast 'pushing to'patient center' self-adaptive and personalized pushing is realized, the accuracy, timeliness and user experience of information pushing are remarkably improved, and meanwhile, the safety and reachability of medical emergency information are ensured.

Description

Technical Field

[0001] This invention relates to the field of sampling device technology, and specifically to a smart medical information push method and system applied to internet hospitals. Background Technology

[0002] With the increasing prevalence of internet-based healthcare services, hospital information systems are pushing medical information to patients in more and more scenarios, such as test reports, medication reminders, and follow-up appointment notices. Currently, such information pushes are mostly based on preset fixed schedules, static patient groups, or simple business trigger rules. For example, reports are sent immediately after generation, or medication reminders are given at fixed time intervals.

[0003] However, this push notification method lacks awareness and consideration of the patient's real-time status. Patients may be resting, undergoing stressful treatment, experiencing physical discomfort, or highly focused. Receiving non-urgent, routine information pushes in these situations not only disrupts the experience and reduces user experience but may also lead to important information being overlooked. Conversely, if the system fails to recognize that the patient is in a state where they can receive information, it may delay the delivery of non-urgent but time-sensitive information. This "one-size-fits-all" push notification model is essentially a system-centric "broadcast," making it difficult to achieve accurate and personalized information delivery, and has become a key bottleneck in improving the service quality and patient satisfaction of internet hospitals. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a smart medical information push method and system for internet hospitals, thereby resolving the problems mentioned in the background section.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for pushing smart medical information to internet hospitals, which is applied to a smart medical information push system. The system is connected to a hospital information system, multiple IoT sensing terminals, and patient mobile terminals. The method includes: S110. Obtain real-time status data of the target patient, wherein the real-time status data includes physiological monitoring data, terminal behavior data, and data of the current diagnosis and treatment stage; S120. Based on the real-time status data, calculate the current status stability of the target patient using a preset status assessment model; S130. When there is medical information to be pushed, a recommendation push strategy is dynamically determined based on the current state stability, the attributes of the medical information and historical push feedback data. The recommendation push strategy includes the recommendation push timing, the recommendation push channel and the recommendation content rendering method. S140. Based on the recommended push strategy, generate and send adapted medical information to the target patient's mobile terminal.

[0006] Preferably, the step of calculating the current state stability of the target patient based on the real-time state data using a preset state assessment model includes: S210. The physiological monitoring data is preprocessed to extract a first feature vector, which includes heart rate variability, blood pressure trend slope, and blood oxygen saturation fluctuation value. S220. Analyze the terminal behavior data and extract a second feature vector. The second feature vector includes screen activation frequency, application usage category, and motion trajectory silent duration. S230. Map the current treatment stage data to a treatment tension coefficient; S240. Input the first feature vector, the second feature vector, and the diagnosis and treatment tension coefficient into a preset multimodal fusion neural network model; S250. Output the current state stability score of the target patient through the multimodal fusion neural network model.

[0007] Preferably, the step of dynamically determining the recommendation push strategy based on the current state stability, the attributes of the medical information, and historical push feedback data includes: S310. Analyze the attributes of the medical information to obtain the information urgency level, information type, and expected reading time; S320. Query the historical push feedback data of the target patient under similar information types, and calculate the average reading rate, average confirmation time and negative feedback index. S330. Construct a strategy decision function with the goal of maximizing the expected delivery effect. The variables of the strategy decision function include candidate push timing, candidate push channel, and candidate content rendering method. S340. Input the current state stability score, the information urgency level, the expected reading time, the average reading rate, the average confirmation time, and the negative feedback index as constraints and weight parameters into the strategy decision function; S350. By solving the strategy decision function, determine the recommendation push timing, the recommendation push channel, and the recommendation content rendering method from the candidate strategy set.

[0008] Preferably, the method further includes a step of performing an information environment coupling degree verification before generating the adapted medical information: S410. Obtain the real-time environmental parameters of the target patient's mobile terminal, including ambient noise decibels, ambient light intensity, and network signal quality. S420. Determine the sensitivity requirements of the recommended content to environmental parameters based on the recommended content rendering method; S430. Calculate the coupling score between the current environmental parameters and the sensitivity requirement; S440. If the coupling score is lower than the preset coupling threshold, the recommended content rendering method is optimized according to environmental parameters, or a decision is made on whether to delay the push based on the information urgency level.

[0009] Preferably, generating and sending adapted medical information according to the recommendation push strategy includes: S510. According to the recommended content rendering method, call the corresponding content template skeleton from the preset template library; S520. Fill the key data of the medical information into the content template skeleton to generate the original content; S530. Based on the current state stability score and the expected reading time, perform intelligent summarization or key information highlighting on the original content; S540. Based on the technical specifications of the recommended push channel, the processed content is formatted and compressed. S550, at the recommended push timing, the encapsulated information data packet is sent to the target patient's mobile terminal through the recommended push channel.

[0010] Preferably, the method further includes a closed-loop feedback learning step after the push: S610. Monitor the target patient's interaction with the pushed medical information, and record the actual reading time, confirmation operation, ignore operation or close operation; S620. Calculate the actual effect score of this push based on the interaction behavior; S630. The complete data chain of this push is used as a training sample. The data chain includes the real-time status data before the push, the push strategy adopted, and the actual effect score. S640. Periodically use newly added training sample sets to incrementally learn and optimize the parameters in the preset state evaluation model and policy decision function.

[0011] Preferably, the dynamically determined recommendation push strategy further includes priority processing rules for urgent information: S710. If the urgency level of the information is higher than the preset critical threshold, the recommendation push timing in the recommendation push strategy is immediately set to "execute immediately". S720. At the same time, ignore the portion of the current state stability score that is lower than the preset interference threshold, and force the transmission to be carried out using a redundant push channel combination, wherein the redundant push channel combination includes at least two independent communication channels.

[0012] Preferably, before acquiring the real-time status data of the target patient, an adaptive adjustment step for the patient status perception cycle is further included: S810. Set the basic sensing frequency based on the target patient's disease type, historical status fluctuation pattern, and whether the patient is currently hospitalized. S820. When abnormal fluctuations are detected in the physiological monitoring data or the terminal behavior data indicates a significant increase in patient activity, the sensing frequency is automatically increased. S830. When the current state stability score is continuously higher than the high stability threshold and the duration exceeds the preset window, the sensing frequency is automatically reduced.

[0013] Secondly, the present invention provides a machine-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for pushing smart medical information to internet hospitals.

[0014] Thirdly, the present invention provides a smart medical information push system for internet hospitals, including a memory and a processor. The memory stores a computer program, and the processor executes the program to implement the above-mentioned smart medical information push method for internet hospitals.

[0015] The beneficial effects of this invention are as follows: By comprehensively sensing the patient's physiological, behavioral, and treatment stages in real time and intelligently assessing their "state stability," this invention can dynamically determine the optimal time for information push. Combining information attributes and user historical preferences, the system of this invention can adaptively select the optimal push channel and content format, thereby ensuring that information is delivered to the user at a suitable time and in an appropriate manner. This effectively solves the problems of push harassment or information omission caused by traditional fixed strategies and improves the user experience.

[0016] This invention, through a closed-loop feedback mechanism, learns from the interaction results of each push notification, continuously optimizing the state assessment model and strategy decision-making algorithm, making the push service increasingly accurate over time. By setting a highest priority rule for emergency information, this invention ensures that critical information is delivered immediately and redundantly, regardless of normal interference, firmly safeguarding the bottom line of medical safety. Simultaneously, the system can dynamically adjust the monitoring frequency according to the patient's status, achieving intelligent scheduling of computing and communication resources and improving overall operational efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The flowchart illustrates the steps of the smart medical information push method provided by this invention.

[0019] Figure 2 The flowchart below shows the specific process of the smart medical information push method provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the patient status assessment and strategy decision-making process provided by the present invention.

[0021] Figure 4 The flowchart of the closed-loop feedback learning mechanism provided by this invention is shown. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example: like Figure 1 As shown in the figure, this application provides a method for pushing smart medical information to an internet hospital. The method is applied to a smart medical information push system, which is connected to a hospital information system, multiple IoT sensing terminals, and patient mobile terminals. The method includes the following steps: S110. Obtain real-time status data of the target patient, wherein the real-time status data includes physiological monitoring data, terminal behavior data, and data of the current diagnosis and treatment stage; S120. Based on the real-time status data, calculate the current status stability of the target patient using a preset status assessment model; S130. When there is medical information to be pushed, a recommendation push strategy is dynamically determined based on the current state stability, the attributes of the medical information and historical push feedback data. The recommendation push strategy includes the recommendation push timing, the recommendation push channel and the recommendation content rendering method. S140. Based on the recommended push strategy, generate and send adapted medical information to the target patient's mobile terminal.

[0024] Specifically, in step S110, acquiring the real-time status data of the target patient is the data foundation for the dynamic adaptive push of this method. The real-time status data is a multi-dimensional collection, which aims to depict the patient's current immediate situation from three levels: physiological, behavioral and environmental.

[0025] In step S110 of this invention, the physiological monitoring data originates from various medical and civilian-grade IoT sensing terminals connected to the system: Specifically, for hospitalized patients, the system securely connects to devices such as bedside monitors and wireless vital signs monitoring patches through the hospital intranet to obtain continuous or periodic sampling data such as electrocardiogram (ECG), non-invasive blood pressure (NIBP), blood oxygen saturation (SpO2), respiration (RESP), and body temperature (TEMP) in real time. For patients receiving treatment at home or in outpatient settings, the system, through an authorized protocol, asynchronously or synchronously collects relevant physiological parameters from wearable devices or home medical devices such as smart bracelets, smartwatches, Bluetooth blood pressure monitors, and blood glucose meters worn by the patient. All collection processes must comply with medical data privacy and security regulations, de-identify patient information, and securely associate only anonymized or pseudo-named device identifiers with patient records.

[0026] In step S110 of this invention, terminal behavior data refers to data collected from the patient's own smart mobile terminal (mainly a smartphone) that can indirectly reflect their current activity state and attention. After obtaining the patient's explicit informed consent and authorization, data is collected under the premise of protecting privacy by integrating a lightweight, low-power behavior-aware SDK into the patient's terminal application (APP) or by utilizing the limited state interface provided by the operating system (which requires user authorization).

[0027] The data collected includes: 1) Device status: such as whether the screen is on (activated / off), locked screen status, and battery level; 2) Application usage: the name or category of the application currently running in the foreground (which can be mapped through app store category tags, such as "social entertainment", "video playback", "e-reading", "work office", "health care"), and the frequency of switching between different applications; 3) Interactive activities: touch screen operation frequency and keyboard input events; 4) Physical activities and location: sitting, walking, running, etc., identified by the terminal's built-in accelerometer and gyroscope, as well as rough location information obtained by combining GPS / Wi-Fi / Bluetooth beacons (such as "bedroom at home", "hospital outpatient building", "outdoor park"). The location information is only used to determine the general environment category and is not used for precise trajectory tracking. After these behavioral data are initially aggregated and de-identified locally, they are then encrypted and transmitted to the system server.

[0028] In step S110 of this invention, the current treatment stage data is standardized information about the patient's current position in the medical process, obtained from a hospital information system (HIS), electronic medical record system (EMR), or internet hospital platform business system. This data is typically stored in structured fields. For example, outpatients may be in states such as "Successfully booked an appointment," "Waiting for appointment and check-in," "Doctor seeing patients," "Tests and examinations pending," "Tests and examinations in progress," "Queuing for medication," or "Discharge after treatment." Inpatients may be in states such as "Newly admitted," "Pre-operative preparation," "During surgery," "Post-operative recovery," "Returning to ward after surgery," "Routine treatment period," "Rehabilitation treatment period," or "Discharge preparation." For patients under chronic disease management, they may be in states such as "Routine follow-up period," "Abnormal indicators pending re-examination," or "Emergency situation reported and processed." The system periodically polls or receives event notifications through standard HL7, FHIR interfaces, or business APIs provided by the hospital to update the patient's current treatment stage.

[0029] In step S120, based on the multi-source heterogeneous real-time state data obtained in step S110, a comprehensive and quantitative current state stability is calculated using a preset state assessment model. This stability is not a single physiological indicator, but a composite score that integrates physiological stability, behavioral accessibility, and environmental interference. Its core purpose is to assess the patient's suitability to "receive and process non-urgent medical information" at the current moment. A higher score indicates that the patient is in a relaxed, idle, and low-interference state, suitable for receiving notifications such as education and routine reports; a lower score indicates that the patient may be experiencing physiological discomfort, busy behavior, or in a high-pressure treatment process, in which case unnecessary disturbances should be avoided, and only urgent information should be accessible.

[0030] The preset state assessment model is a machine learning model trained on a large amount of labeled data. The training data comes from historical anonymous patient data, which is manually labeled with the "information reception appropriateness" level by clinical experts and user experience experts based on the patient's current situation (combined with complete physiological, behavioral, and treatment stage records).

[0031] The model can employ gradient boosting decision trees (such as XGBoost, LightGBM), deep neural networks (such as multilayer perceptrons MLP), or more complex multimodal fusion models. The input to the model is a state data vector after feature engineering, and the output is a continuous score value between 0 and 100, or divided into five discrete levels: "very low", "low", "medium", "high" and "very high". The model periodically uses newly generated feedback data for incremental learning to continuously optimize its evaluation accuracy.

[0032] In step S130, when the system's message receiving module receives a piece of medical information that needs to be pushed to the target patient from various hospital business systems (such as LIS, PACS, pharmacy system, and follow-up system), it will not immediately trigger the sending action, but will start a dynamic strategy decision engine. The core task of this engine is to calculate an optimal combination of push parameters for this specific information, at this specific moment, for this specific patient, i.e., a recommended push strategy.

[0033] In this invention, the decision-making process comprehensively considers three main categories of inputs: (1) Patient's immediate situation: that is, the current state stability calculated in step S120, which is the primary factor in determining "when to push" and "whether to push"; (2) Information attributes: Analyze the information to be pushed and obtain its structured attributes, including: information urgency level (e.g., divided into L1 life-threatening, L2 requiring emergency treatment, L3 routine importance, L4 general notice, L5 health education), information type (e.g., test report, imaging report, medication reminder, follow-up appointment, fee notice, satisfaction survey, health knowledge article), information content complexity / expected reading time (which can be estimated by the average reading time of similar information in the past, or calculated based on text length and number of charts); (3) Patient historical preferences and habits: From the patient's historical push log database, query the patient's historical feedback data on similar types of information, perform statistical analysis, and obtain: average reading rate (the proportion of clicks to view after push), average confirmation / processing time (the time from push to the patient completing key operations, such as "confirm receipt" or "schedule the next step"), negative feedback index (such as the proportion of directly ignoring and clearing notifications, the number of times "do not remind this type of message again" is set in the APP, etc.).

[0034] The decision engine encapsulates decision logic based on rules, utility functions, or reinforcement learning algorithms. It evaluates the "expected utility" of each candidate strategy within a strategy search space consisting of dimensions such as push timing (immediate, delayed to a preset period, delayed until state stability is improved), push channel (APP in-app messages, APP push notifications, SMS, email, telephone voice), and content rendering method (plain text summary, image and text details, voice broadcast, short video). The utility function is designed to reward pushing non-urgent information when the state is stable and penalize the inconvenience caused by pushing non-urgent information when the state is stable. It rewards channels and timings with high historical read rates and fast confirmations for the patient, while also constraining the priority and rapid delivery of urgent information. Ultimately, the engine outputs the strategy with the highest expected utility as the recommendation push strategy for this time.

[0035] In step S140, the system's message generation and sending module performs "adaptive processing" and "precise delivery" of the original medical information according to the recommended push strategy determined in step S130.

[0036] First, based on the recommended content rendering method in the strategy, a matching content skeleton is selected from the template library, and key data (such as patient name, project name, result value, time and location) is filled in to generate basic content. Then, the content may undergo secondary intelligent processing based on the patient's current stability score and expected reading time. For example, when the stability score is low or the expected reading time is short, a summary text containing the most critical conclusions and action suggestions is automatically generated, and outliers are highlighted. When the stability score is high, a complete and detailed report is pushed. Next, according to the technical requirements of the recommendation push channel (such as the character limit of SMS, the image specifications of APP notifications, and the parameters of speech synthesis), the content is formatted, encapsulated, and compressed to generate the final information data packet. Finally, the scheduler calls the corresponding channel service interface at the recommendation push time specified by the strategy to send out the information data packet. For strategies that are not sent immediately, the scheduler will continuously monitor changes in the patient's stability. Once the trigger condition is met (such as the status score rising above a threshold, or entering a preset recommendation period), the sending will be executed immediately.

[0037] Through the above steps S110 to S140, the present invention constructs a closed loop of "perception-evaluation-decision-execution", realizing the transformation from broadcast push from "system center" to conversational and adaptive push from "patient center".

[0038] Figure 2 This is a schematic diagram of the patient status assessment and strategy decision-making process provided in the embodiments of this application. The diagram intuitively shows the complete data flow and logical relationship from multi-source data collection, feature extraction and fusion, status assessment model calculation, to strategy decision-making by combining information attributes and historical feedback, and finally outputting the recommended strategy.

[0039] In one embodiment of this example, calculating the current state stability of the target patient based on the real-time state data using a preset state assessment model includes the following steps: S210. Preprocess the physiological monitoring data and extract the first feature vector; S220. Analyze the terminal behavior data and extract the second feature vector; S230. Map the current treatment stage data to a treatment tension coefficient; S240. Input the first feature vector, the second feature vector, and the diagnosis and treatment tension coefficient into a preset multimodal fusion neural network model; S250. Output the current state stability score of the target patient through the multimodal fusion neural network model.

[0040] In step S210, preprocessing and feature extraction of the physiological monitoring data are crucial, as raw physiological signals often contain noise, artifacts, and missing values. First, data cleaning is performed: obvious outliers and noise are removed using methods such as sliding window midpoint filtering and wavelet transform denoising; for short-term missing data, linear interpolation or prediction based on adjacent sequences is used for filling; for long-term missing data, that time period is marked as invalid. After cleaning, features reflecting the "stability" and "stationarity" of the data are extracted for different parameters, forming the first feature vector. For example:

[0041] For ECG or pulse wave signals: calculate time-domain metrics (such as SDNN, standard deviation of normal sinus intervals) and frequency-domain metrics (such as the LF / HF ratio) of heart rate variability (HRV). Decreased HRV is often associated with stress, fatigue, or certain pathological conditions, suggesting potentially lower stability.

[0042] For blood pressure sequences: Calculate the trend slope of recent (e.g., past 30 minutes) systolic blood pressure (SBP) and diastolic blood pressure (DBP). A significantly rising or falling trend slope may indicate that the patient's physiological state is changing rapidly and is not stable. Also calculate its variability (e.g., standard deviation).

[0043] For blood oxygen saturation (SpO2) sequences: calculate their fluctuation values ​​(e.g., moving standard deviation) and detect whether there are instances where they are consistently below a specific threshold (e.g., 95%). Abnormal fluctuations or persistent decreases in SpO2 are important indicators of unstable conditions;

[0044] For respiratory rate: calculate its regularity (e.g., coefficient of variation); For body temperature: calculate the degree to which it deviates from the individual's baseline value.

[0045] After these features are extracted, they are standardized (Z-score normalization) to form the first feature vector V_physio.

[0046] In step S220, the terminal behavior data is analyzed to quantify the patient's "behavioral busyness" and "cognitive accessibility." The extracted second feature vector V_behavior includes: Screen activation frequency: The number of times the screen is turned on from off within a unit of time (e.g., the past 15 minutes). High activation frequency may indicate anxiety, frequent task processing, or idly scrolling through the screen, suggesting a lack of attention.

[0047] Application Usage Category: Identifies the categories of currently foreground applications and recently switched applications. Using a predefined mapping table, application package names are mapped to broad categories such as "Leisure & Entertainment," "Social Communication," "Work & Study," "Transportation & Navigation," "Health & Fitness," and "System Tools." Continued use of "Leisure & Entertainment" applications may indicate idle time, while frequent switching between "Work & Study" and "Social Communication" applications may indicate busy time.

[0048] Motion trajectory silence duration: Based on accelerometer and location information, this determines the continuous duration of a patient's stationary state (e.g., sitting or lying down). Longer silence durations may correspond to rest, sleep, or focused states (e.g., reading), during which the tolerance for disturbance may be lower; while short silences and high activity levels may correspond to periods of daily activity.

[0049] Interaction density: The number of touch screen clicks or swipes per unit of time; These behavioral characteristics have also been standardized.

[0050] In step S230, the structured data of the current treatment stage is converted into a scalar value—the treatment stress coefficient K_treatment—through a predefined treatment stress mapping table, with a value range of, for example, [0.0, 1.0]. This coefficient, developed with the participation of medical experts, reflects the psychological stress, environmental noise, and degree of attentional demands commonly associated with different treatment stages. For example, "under surgery" is mapped to 0.95, "emergency resuscitation" to 1.0, "under doctor's consultation" to 0.7, "waiting for examination" to 0.6, "routine intravenous therapy" to 0.4, and "home rehabilitation" to 0.2. This coefficient serves as a strong prior information input model.

[0051] In steps S240 and S250, a pre-defined multimodal fusion neural network model is responsible for deep fusion and joint inference of features from different modalities. This model can adopt a multimodal fusion architecture based on an attention mechanism. The input layer receives V_physio, V_behavior, and K_treatment, respectively. First, V_physio and V_behavior are abstracted and transformed in dimensionality within the modality through fully connected sub-networks to obtain high-order feature representations. K_treatment can be embedded as conditional information; then, a cross-modal attention module is introduced to allow the physiological feature sub-network and the behavioral feature sub-network to "attention" each other and learn the correlation between them (for example, high heart rate and frequent screen switching may both point to an anxiety state); finally, the fused features of all modalities are concatenated and then combined for decision-making through several fully connected layers. The output layer uses the Sigmoid activation function to map the output to the [0, 1] interval, and then multiplies it by 100 to obtain a current state stability score (State_Score) of 0-100.

[0052] The model uses mean squared error (MSE) or cross-entropy loss function and is trained with "fitness scores" labeled by historical experts as supervision signals. In this way, the model can learn complex nonlinear relationships such as "when HRV is high, entertainment applications are used, or the patient is in the home recovery stage, the state stability score should be higher".

[0053] In one embodiment of this example, dynamically determining the recommendation push strategy based on the current state stability, the attributes of the medical information, and historical push feedback data includes the following steps: S310. Analyze the attributes of the medical information to obtain the information urgency level, information type, and expected reading time; S320. Query the historical push feedback data of the target patient under similar information types, and calculate the average reading rate, average confirmation time and negative feedback index. S330. Construct a strategy decision function with the goal of maximizing the expected delivery effect; S340. Input the current state stability score, the information urgency level, the expected reading time, the average reading rate, the average confirmation time, and the negative feedback index as constraints and weight parameters into the strategy decision function; S350. By solving the strategy decision function, determine the recommendation push timing, the recommendation push channel, and the recommendation content rendering method from the candidate strategy set.

[0054] In step S310, parsing the attributes of medical information is a prerequisite for strategic decision-making. The urgency level of information is usually automatically determined by the business system that generates the information according to clinical rules, or manually set by medical staff. For example, critical values ​​for tests are automatically labeled as L1, abnormal results exceeding the reference range but not critical are labeled as L2, reports within the normal range are labeled as L3, medication reminders are labeled as L2 or L3 (depending on the importance of the medication), and health articles are labeled as L5. The information type is obtained from the message template ID or content category field, and the expected reading time T_estimate can be set based on empirical values ​​for the information type, or estimated by using natural language processing technology to perform readability analysis and length assessment of the text content.

[0055] In step S320, querying historical push notification data aims to capture the patient's personal habits and preferences. From the patient's personal push notification history log, historical records with the same or similar information type as the current message (e.g., all "test report" push notifications from the past 30 days) are filtered out. The following indicators are calculated:

[0056] Average read rate R_read: The number of times the message was clicked (i.e., accessed from the notification bar to the application details page) in the history / the total number of push notifications; Average Confirmation Time T_confirm_avg: Calculates the median or average time interval between successful push and the patient performing a key confirmation action (such as clicking the "Read" or "Understood" button) across all read messages. For message types without a confirmation action, this field can be left blank or replaced with the reading time.

[0057] Negative Feedback Index F_neg: This is a comprehensive indicator that can be calculated as: F_neg = α * (number of times ignored / number of push notifications) + β * (number of times messages were set to Do Not Disturb), where α and β are weighting coefficients. It quantifies the degree of a patient's aversion to this type of information.

[0058] In step S330, a strategy decision function Expected_Effect = f(State_Score, Urgency, T_estimate, R_read, T_confirm_avg, F_neg; t, c, r) is constructed. Here, State_Score, etc., are known input parameters, and t (push timing), c (push channel), and r (content rendering method) are decision variables. Function f calculates the expected effect value of adopting strategy (t, c, r) under given parameters. The design of function f reflects the business logic and optimization goals. A simplified exemplary design is as follows:

[0059] Expected_Effect = Base_Utility(t, c, r) - Disturbance_Cost(State_Score, Urgency) + Preference_Bonus(c, r, R_read, T_confirm_avg, F_neg) -Delay_Penalty(Urgency, t) Base_Utility: Measures the technical cost and basic effect of the strategy (t,c,r). For example, telephone channels have high costs but high delivery rates, thus giving them higher base utility; graphic details are more user-friendly than plain text summaries, so their base utility is slightly higher.

[0060] Disturbance Cost: This is the core penalty. It's designed as a function of State_Score and Urgency. When both State_Score and Urgency are low (patient condition unstable) and Urgency is low (information not urgent), the disturbance cost is very high, severely reducing the expected effectiveness. As State_Score or Urgency increases, the disturbance cost decreases. This ensures that non-urgent information is not sent when the patient is in a poor condition.

[0061] Preference_Bonus: A positive reward is given if the candidate channel c and rendering mode r are of the type that the patient has historically received good feedback (high R_read, short T_confirm_avg, low F_neg).

[0062] Delay_Penalty: For messages with a high urgency level (Urgency is L1 / L2), any non-immediate sending strategy (t is not "immediate") will be penalized with a penalty proportional to the urgency level and the delay time, forcing the engine to choose immediate sending for urgent messages.

[0063] In step S340, the State_Score calculated in step S120, the Urgency and T_estimate parsed in step S310, and the R_read, T_confirm_avg, and F_neg calculated in step S320 are used as specific numerical parameters and substituted into the corresponding variables of the policy decision function f. Meanwhile, the weight coefficients within the function (such as the slope of the disturbance cost function and the magnitude of the preference reward) are pre-set based on business experience or optimized through offline reinforcement learning.

[0064] In step S350, solving the strategy decision function involves searching for the strategy (t, c, r*) that maximizes the Expected_Effect value from a finite, predefined set of candidate strategies. The set of candidate strategies is discrete: the push timing t may include options such as {immediately, delayed to [18:00-20:00], delayed until status score > 70, delayed until 9:00 the next day}; the push channel c may include {strong APP notification, weak APP notification, SMS, email}; and the rendering method r may include {text summary, image and text details, voice summary}. The search method can be a simple grid search (traversing all combinations to calculate the effect value) or a more efficient heuristic algorithm (such as fixing t=immediately for highly urgent information and then searching c and r). Finally, the strategy (t*, c, r) that maximizes the Expected_Effect is the recommended push strategy for this push.

[0065] In one embodiment of this invention, the method further includes a step of performing an information environment coupling degree verification before generating the adapted medical information: S410. Obtain the real-time environmental parameters of the target patient's mobile terminal, including ambient noise decibels, ambient light intensity, and network signal quality. S420. Determine the sensitivity requirements of the recommended content to environmental parameters based on the recommended content rendering method; S430. Calculate the coupling score between the current environmental parameters and the sensitivity requirement; S440. If the coupling score is lower than the preset coupling threshold, the recommended content rendering method is optimized according to environmental parameters, or a decision is made on whether to delay the push based on the information urgency level.

[0066] Step S410 aims to obtain the physical environment conditions of the information receiving end, which is an important factor affecting the effective reception of information. Environmental parameters are obtained through sensors on the patient's mobile terminal or by inferring from the context. Ambient noise decibel level: Short-term sampling is performed using the terminal microphone (with user authorization and the APP in the foreground or a specific background mode) to calculate the average sound pressure level, or to infer typical noise levels based on time and geographical location (such as "hospital", "subway station", "library"). Ambient light intensity: obtained through the terminal's ambient light sensor, or roughly determined based on time (night) and location (indoor / outdoor); Network signal quality: Obtain the current Wi-Fi or cellular network (4G / 5G) connection status and signal strength indication (such as RSSI, RSRP). Weak networks may cause slow loading of images and text or voice interruptions.

[0067] In step S420, different content rendering methods have different environmental requirements, and the system maintains a rendering method-environment sensitivity matrix. For example:

[0068] Voice broadcasting / reading aloud: Highly sensitive to ambient noise. In noisy environments (e.g., >65 decibels), voice audibility is poor, resulting in a subpar user experience. Its requirements for lighting and network connectivity are generally moderate.

[0069] Detailed text / images / videos: Require a certain level of light intensity. Reading small text or watching videos in dim environments (e.g., <10 lux) can easily cause eye strain. Additionally, it requires high-quality network signal; large images or videos may fail to load or stutter under weak network conditions.

[0070] Plain text summary: It has the lowest sensitivity to the environment, mainly depends on screen brightness, and is highly adaptable.

[0071] In step S430, the coupling score (Coupling_Score) is calculated. An ideal range [IdealMin_i, IdealMax_i] and a tolerance range [TolMin_i, TolMax_i] are set for each environmental parameter. For example, for "voice broadcasting," the ideal range for environmental noise is [30, 50] dB, and the tolerance range is [20, 65] dB.

[0072] The coupling score is calculated using the following formula: Coupling_Score = Σ (w_i * s_i) in: i represents the i-th environmental parameter (such as noise, light, network); w_i is the weighting coefficient of this parameter, satisfying Σ w_i = 1.

[0073] s_i is the sub-score of this parameter, and it is calculated as follows: s_i = max( 0, 1 - |Env_i - Mid_i| / R_i ) Env_i is the current measured value of the i-th environment parameter.

[0074] Mid_i is the center value of the ideal range, and it is calculated as Mid_i = (IdealMax_i + IdealMin_i) / 2.

[0075] R_i is half the width of the tolerance range, and is calculated as R_i = (TolMax_i - TolMin_i) / 2.

[0076] This formula calculates the normalized distance of the current value from the ideal center value and converts it into a score within the interval [0, 1]. The closer the value is to the center value, the higher the score; if it exceeds the tolerance range, the score is 0.

[0077] Step S440: Perform verification and strategy adjustment. Set a coupling threshold, such as 0.6. If the calculated Coupling_Score is lower than this threshold, it means that the current environment is not conducive to rendering information according to the original strategy.

[0078] Optimize rendering methods: The system automatically downgrades or switches rendering methods based on environmental parameters. For example, if the original strategy is "voice broadcast" but the ambient noise is high, it will automatically switch to "text pop-up + strong vibration". If the original strategy is "detailed text and images" but the lighting is dim and the network is poor, it will downgrade to "plain text core summary".

[0079] Delayed Push Decision: If the optimized rendering method is still unsatisfactory (e.g., voice is required but the environment remains noisy), or if there is limited room for optimization, the system will make further decisions based on the urgency level of the information. If the information is urgent (L1 / L2), it will still be sent immediately according to the optimized method to ensure information reachability is prioritized. If the information is not urgent (L3 and below), a delayed push decision may be made: The system records this pending task and starts environmental monitoring. When it detects that environmental parameters have improved (e.g., reduced noise, improved network) and the patient's condition is relatively stable, it will then execute the sending. This avoids ineffective communication and negative experiences caused by pushing non-urgent information in poor environments.

[0080] In one embodiment of this example, generating and sending adapted medical information according to the recommendation push strategy includes the following steps: S510. According to the recommended content rendering method, call the corresponding content template skeleton from the preset template library; S520. Fill the key data of the medical information into the content template skeleton to generate the original content; S530. Based on the current state stability score and the expected reading time, perform intelligent summarization or key information highlighting on the original content; S540. Based on the technical specifications of the recommended push channel, the processed content is formatted and compressed. S550, at the recommended push timing, the encapsulated information data packet is sent to the target patient's mobile terminal through the recommended push channel.

[0081] In step S510, the system maintains a multimodal content template library. Each template is bound to one or more content rendering methods and defines the structure, layout, style, and dynamic data placeholders of the information. For example, the "Laboratory Report Graphic Details Template" includes a patient information area, a report summary area, a list of abnormal indicators (with charts), a normal indicator folding area, a doctor's suggestion area, and an operation button area. The "Medication Reminder Voice Template" defines the structure of the voice script and variable insertion points.

[0082] Step S520 is the content generation stage. The system parses the raw medical information (usually a structured JSON message or database record) and extracts key data fields (such as patient name, test name, result value, unit, reference range, submission time, etc.). Then, this data is filled into the corresponding placeholders in the template skeleton called in step S510 to generate a complete "raw content" that conforms to the template format. This process may involve unit conversion, numerical formatting, time localization, etc.

[0083] Step S530 is the content adaptation optimization stage. Even if a certain rendering method is selected, the level of detail and emphasis of the content presentation still need to be fine-tuned according to the receiving context. This is mainly based on two factors: the patient's current state stability score (State_Score) and the expected reading time of the information (T_estimate).

[0084] Intelligent summarization: When the State_Score is low (e.g., <40) or the T_estimate is short, the system triggers the summary generation module. This module may be based on rules (e.g., extracting all "abnormal" indicators) or use a natural language processing model (e.g., a text summarization model) to extract the most core conclusions, action items, and key data from the generated original content, generating a concise summary version. For example, a summary of a complete blood routine report could be: "Your blood routine report is out: White blood cell count is high, suggesting a possible infection; a follow-up visit is recommended. Other indicators are normal."

[0085] Key information highlighting: For content that maintains its original length, the system will analyze the content and highlight key information (such as outliers, critical values, and action deadlines) visually or aurally. In text and image content, this may involve changing the font color, bolding, or adding warning icons; in audio content, it may involve emphasizing through changes in speech rate and tone or inserting prompts.

[0086] Step S540 is the channel adaptation and encapsulation stage. Different push channels have specific technical limitations and requirements. The system calls the corresponding channel adapter based on the type of the recommended push channel.

[0087] App push notifications: Content needs to be converted into a platform-specific notification format (such as NotificationCompat for Android and UNNotificationContent for iOS), and the title, body, icon, click action, and extended content (for displaying details after expansion) need to be set. Large images or rich media content may also need to be uploaded to a CDN to obtain the URL.

[0088] SMS: The text content needs to be compressed to a specified number of characters (e.g., 70 characters / message), long links need to be shortened, and the format requirements of the carrier need to be followed; Email: Requires generating an HTML-formatted email body, setting a plain text option, handling attachments, and adhering to email subject guidelines; Speech synthesis: This requires submitting text content to a speech synthesis (TTS) service to generate an audio file or obtain an audio stream URL, and may insert prosodic markers based on highlight information.

[0089] After encapsulation, a standardized information data packet is generated, which includes the target device identifier, channel instructions, and content payload that has been adapted and possibly compressed.

[0090] Step S550 is the final sending stage, where the system's message scheduler is responsible for executing the sending at the recommended push timing determined by the strategy decision. For "immediate" sending, the scheduler immediately places the task into the sending queue after the strategy decision is completed. For delayed sending (such as "after 18:00" or "when the status score is >70"), the scheduler will create a scheduled task or subscribe to patient status events. When the timing is met, the scheduler retrieves the corresponding information data packet from the queue, calls the API of the corresponding channel service provider (such as JPush, Alibaba Cloud SMS, SendGrid email, or telephone voice API), and sends the information data packet out. The sending process is usually asynchronous, and the sending result (success, failure, device not registered, etc.) is recorded in the log for subsequent delivery rate analysis and strategy optimization.

[0091] In one embodiment of this invention, the method further includes a closed-loop feedback learning step after the push: S610. Monitor the target patient's interaction with the pushed medical information, and record the actual reading time, confirmation operation, ignore operation or close operation; S620. Calculate the actual effect score of this push based on the interaction behavior; S630. The complete data chain of this push is used as a training sample. The data chain includes the real-time status data before the push, the push strategy adopted, and the actual effect score. S640. Periodically use newly added training sample sets to incrementally learn and optimize the parameters in the preset state evaluation model and policy decision function.

[0092] Step S610 involves feedback data collection. After the information is successfully delivered to the patient's terminal, the system continuously monitors the patient's subsequent interactions using technologies such as in-application embeddings, push service callbacks, SMS status reports, and email open tracking. Key indicators include: Whether it was read / clicked: For notification messages, record whether the user clicked through the notification bar to enter the application details page; Actual reading time: Within the application details page, the time spent on the page is recorded through front-end tracking (excluding inactive time), which serves as an approximation of the actual reading time; Confirmation Action: Record whether the user has performed a preset key action, such as clicking the "Read", "Understood", "Book Now", "Pay" buttons; Ignore action: Record whether the user cleared the notification directly without clicking it; Turn off / Do Not Disturb: Record whether the user has set "Do Not Remind Again" or "Do Not Disturb" for this type of message; These interaction events are associated with push notification IDs and include timestamps.

[0093] In step S620, based on the collected interaction behaviors, the actual_Effect_Score of this push is calculated. This score can be calculated using a weighted evaluation model, as shown in the example below: Actual_Effect_Score = α * I_read + β * (T_estimate / max(T_actual,1)) + γ * I_confirm - δ * I_ignore - ε * I_block.

[0094] in: I_read is an indicator function; it returns 1 if the user clicks to read the page, and 0 otherwise. T_estimate is the estimated reading time, and T_actual is the actual reading time. The ratio reflects the sufficiency of the reading (it is best to assume that the reading time is close to or exceeds the estimated time). I_confirm is an indicator function; it is 1 if there is a critical confirmation operation. I_ignore is an indicator function; it is 1 if ignored. I_block is an indicator function; it returns 1 (severe negative) if it causes the user to set Do Not Disturb. α, β, γ, δ, ε are the weighting coefficients of each item, which can be adjusted according to business objectives (e.g., whether to focus more on read rate or confirmation rate). The score is a continuous value of 0-1 or 0-100, which quantifies the overall effect of this push.

[0095] Step S630 is sample construction, which encapsulates a complete push process into a sample data set that can be used for machine learning training. This sample contains: Input features (X): Real-time status data (or its extracted feature vector) collected at the moment before the push and used for decision-making, as well as medical information attributes; Actions taken (A): The actual push strategy implemented (push timing, channel, rendering method); The reward (R) obtained: the calculated Actual_Effect_Score.

[0096] This data chain (X, A, R) is stored in a dedicated experience replay buffer or sample database.

[0097] Step S640 is model optimization, where the system periodically (e.g., daily or weekly) initiates model update tasks.

[0098] Optimizing the state assessment model: New samples provide indirect feedback that "under specific state data X, the push ultimately achieved an effect R". Although R is not a direct label for the state assessment, it can be indirectly used to adjust the state assessment model through logic such as "under state X, pushing non-urgent information received a low score R, which may mean that the state assessment score State_Score was too high at that time". More commonly, (X, R) is used as a weak supervision signal, and the state assessment model is fine-tuned by using inverse reinforcement learning or joint training with a policy model, so that the stability of the assessed state can better predict the effect of subsequent pushes.

[0099] Optimizing the policy decision function: This is a typical reinforcement learning scenario. The accumulated samples (X, A, R) can be directly used to optimize the parameters in the policy decision function (such as the weights of each term in function f and the shape of the cost function). Offline reinforcement learning algorithms such as policy gradient methods or Q-learning (discretizing the policy) can be used to learn from historical experience. The goal is to find policy parameters that maximize the long-term cumulative effect (i.e., the expected value of R). Through continuous learning, the policy decision engine can predict the effects of different policies more and more accurately, thereby making better decisions.

[0100] This closed-loop feedback learning mechanism transforms the entire system from static, rule-based to evolutionary and data-driven. The system learns from every interaction with the patient, continuously optimizing its "perception" and "decision-making" capabilities, thereby providing increasingly personalized and effective push services over time.

[0101] In one embodiment of this invention, the dynamically determined recommendation push strategy further includes priority processing rules for urgent information: S710. If the urgency level of the information is higher than the preset critical threshold, the recommendation push timing in the recommendation push strategy is immediately set to "execute immediately". S720. At the same time, ignore the portion of the current state stability score that is lower than the preset interference threshold, and force the transmission to be carried out using a redundant push channel combination, wherein the redundant push channel combination includes at least two independent communication channels.

[0102] Step S710 defines the highest priority principle for the timeliness of emergency information. The preset critical thresholds typically correspond to information urgency levels L1 (life-threatening) and L2 (requiring emergency treatment). Once the decision engine determines that the information urgency level has reached or exceeded this threshold, regardless of the current state stability score calculated by the state assessment model or the "optimal" timing calculated by the strategy decision function, the system will override the push timing field in the original strategy and force it to be set to "immediate execution." This means that the message sending task will be given the highest priority, inserted at the front of the sending queue, and enter the sending process with almost no delay. This rule ensures the bottom line of medical safety and avoids delaying the delivery of critical information in pursuit of "non-intrusion."

[0103] Step S720 defines the highest principle of accessibility for emergency information, which includes two layers of protection: Ignoring Status Interference: Normally, a low status stability score inhibits the delivery of non-urgent information. However, for urgent information, the system sets a threshold for interference (e.g., a status score below 20 might indicate the patient is in deep sleep or experiencing severe discomfort). When the status score falls below this threshold, the system ignores the negative impact of this low score on the push notification decision, meaning it no longer delays or blocks push notifications. This is because the risk of being "disturbed" by urgent information in this situation is far lower than the potential medical risks of missing information. The system records this coverage but ensures that push notifications are executed.

[0104] Redundant Channel Transmission: To ensure absolute information delivery and avoid single-channel failures (such as app process termination, SMS gateway congestion, or poor signal), the system forcibly activates a redundant push channel combination for emergency information. This is not an "OR" relationship, but an "AND" relationship; a typical combination might include: "APP strong notification (supporting penetration do-not-disturb mode) + SMS + automatic outbound call". The APP strong notification utilizes the operating system's highest priority channel; SMS serves as a near-universal alternative; the automatic outbound call directly broadcasts key content through voice synthesis and may require key confirmation. The system will send the same message in parallel or rapidly sequentially (e.g., immediately activating the next channel if the previous one fails) through all these independent channels. Simultaneously, the system may add a note to the message, such as "Emergency medical notification, sent via multiple methods, please check."

[0105] These two priority rules exist in the system as hard-coded rules or the highest-weight strategy, and their priority overrides any adaptive learning algorithm. They set an insurmountable security boundary for the entire intelligent push system, ensuring that communication in medical emergencies is absolutely reliable and timely while pursuing personalized experience and reducing harassment.

[0106] In one embodiment of this invention, before acquiring the real-time status data of the target patient, an adaptive adjustment step for the patient status perception cycle is further included: S810. Set the basic sensing frequency based on the target patient's disease type, historical status fluctuation pattern, and whether the patient is currently hospitalized. S820. When abnormal fluctuations are detected in the physiological monitoring data or the terminal behavior data indicates a significant increase in patient activity, the sensing frequency is automatically increased. S830. When the current state stability score is continuously higher than the high stability threshold and the duration exceeds the preset window, the sensing frequency is automatically reduced to save system resources.

[0107] Step S810 initializes the sensing frequency. To balance the real-time performance of status sensing with the system's resource consumption (power, data traffic, computing power) (including terminals and servers), the system sets differentiated baseline sensing frequencies for different patients. The decision criteria include: Disease type: For patients with severe, critical or unstable conditions (such as ICU patients or heart failure patients), high-frequency monitoring is required, and the baseline frequency may be set to once every 1-5 minutes; for patients with stable chronic diseases (such as well-controlled hypertension), the frequency can be set to once every 30-60 minutes; for healthy follow-up populations, the frequency can be set to several times a day.

[0108] Historical fluctuation patterns: Analyze the patient's historical status data. If their physiological parameters and behavioral patterns are usually regular and have small fluctuations, the baseline frequency can be appropriately reduced; if the historical fluctuations are large, the frequency should be maintained or increased.

[0109] Hospitalization Status: Hospitalized patients are typically under close medical monitoring in a controllable environment, allowing for a higher baseline sensing frequency. For patients at home, however, considerations must be given to terminal battery consumption and network usage, resulting in a relatively lower frequency.

[0110] The base frequency is set via configuration file or patient profile attributes.

[0111] Step S820 involves increasing the frequency of anomaly triggering. While operating at the base frequency, the system monitors the data stream in real time. If an anomaly signal that may indicate a sudden change in state is detected, an additional state data acquisition and evaluation is immediately triggered, and the sensing frequency may be temporarily increased for several subsequent iterations. Triggering conditions include:

[0112] Abnormal fluctuations in physiological monitoring data: For example, a sudden increase or decrease in heart rate exceeding a set percentage threshold within a short period of time; blood pressure exceeding the safe range twice consecutively; blood oxygen saturation continuously decreasing and falling below the threshold; Significant changes in terminal behavior data: for example, a sudden change from a long period of stillness to high-frequency movement (which may indicate a fall and struggle or an emergency trip); the screen being frequently lit up late at night (which may indicate insomnia or discomfort).

[0113] When these events occur, the system enters a "high alert mode," increasing the detection frequency to a higher level (such as every 30 seconds or every minute) to more closely track changes in status and provide the latest status assessment for possible emergency information pushes.

[0114] Step S830 involves frequency reduction in a stable state. When the system detects that the patient is in a continuous and stable good state, the sensing frequency can be intelligently reduced to save resources. The judgment condition is: the calculated current state stability score is higher than a set high stability threshold (e.g., 85 points) for N consecutive times (e.g., 10 consecutive times) and the duration exceeds a preset time window (e.g., 2 consecutive hours). At this time, the system can determine that the patient is in a low-risk, high-stability stage, and the system will automatically reduce its sensing frequency by one level (e.g., from every 5 minutes to every 15 minutes). After the reduction, the system continues to monitor, and if the state fluctuates again, the frequency will be increased again according to the rules of step S820. This dynamic adjustment mechanism allows system resources to be concentrated on the patients and times that require the most attention, realizing intelligent and efficient resource utilization.

[0115] This application also provides a machine-readable storage medium storing instructions that cause a machine to execute the above-described smart medical information push method applied to internet hospitals.

[0116] This application also provides a smart medical information push system for internet hospitals, including: a memory configured to store instructions; and a processor configured to call the instructions from the memory and to implement the above-described smart medical information push method for internet hospitals when executing the instructions.

[0117] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for pushing smart medical information to internet hospitals, characterized in that, An application in a smart healthcare information push system, wherein the system is connected to a hospital information system, multiple IoT sensing terminals, and patient mobile terminals, the method includes: S110. Obtain real-time status data of the target patient, wherein the real-time status data includes physiological monitoring data, terminal behavior data, and data of the current diagnosis and treatment stage; S120. Based on the real-time status data, calculate the current status stability of the target patient using a preset status assessment model; S130. When there is medical information to be pushed, a recommendation push strategy is dynamically determined based on the current state stability, the attributes of the medical information and historical push feedback data. The recommendation push strategy includes the recommendation push timing, the recommendation push channel and the recommendation content rendering method. S140. Based on the recommended push strategy, generate and send adapted medical information to the target patient's mobile terminal.

2. The smart medical information push method applied to internet hospitals as described in claim 1, characterized in that, The calculation of the current state stability of the target patient based on the real-time state data and using a preset state assessment model includes: S210. The physiological monitoring data is preprocessed to extract a first feature vector, which includes heart rate variability, blood pressure trend slope, and blood oxygen saturation fluctuation value. S220. Analyze the terminal behavior data and extract a second feature vector. The second feature vector includes screen activation frequency, application usage category, and motion trajectory silent duration. S230. Map the current treatment stage data to a treatment tension coefficient; S240. Input the first feature vector, the second feature vector, and the diagnosis and treatment tension coefficient into a preset multimodal fusion neural network model; S250. Output the current state stability score of the target patient through the multimodal fusion neural network model.

3. The smart medical information push method applied to internet hospitals as described in claim 1, characterized in that, The step of dynamically determining the recommendation push strategy based on the current state stability, the attributes of the medical information, and historical push feedback data includes: S310. Analyze the attributes of the medical information to obtain the information urgency level, information type, and expected reading time; S320. Query the historical push feedback data of the target patient under similar information types, and calculate the average reading rate, average confirmation time and negative feedback index. S330. Construct a strategy decision function with the goal of maximizing the expected delivery effect. The variables of the strategy decision function include candidate push timing, candidate push channel, and candidate content rendering method. S340. Input the current state stability score, the information urgency level, the expected reading time, the average reading rate, the average confirmation time, and the negative feedback index as constraints and weight parameters into the strategy decision function; S350. By solving the strategy decision function, determine the recommendation push timing, the recommendation push channel, and the recommendation content rendering method from the candidate strategy set.

4. The smart medical information push method applied to internet hospitals as described in claim 1, characterized in that, The method further includes a step of performing information environment coupling degree verification before generating the adapted medical information: S410. Obtain the real-time environmental parameters of the target patient's mobile terminal, including ambient noise decibels, ambient light intensity, and network signal quality. S420. Determine the sensitivity requirements of the recommended content to environmental parameters based on the recommended content rendering method; S430. Calculate the coupling score between the current environmental parameters and the sensitivity requirement; S440. If the coupling score is lower than the preset coupling threshold, the recommended content rendering method is optimized according to environmental parameters, or a decision is made on whether to delay the push based on the information urgency level.

5. The smart medical information push method applied to internet hospitals as described in claim 1, characterized in that, The step of generating and sending adapted medical information according to the recommendation push strategy includes: S510. According to the recommended content rendering method, call the corresponding content template skeleton from the preset template library; S520. Fill the key data of the medical information into the content template skeleton to generate the original content; S530. Based on the current state stability score and the expected reading time, perform intelligent summarization or key information highlighting on the original content; S540. Based on the technical specifications of the recommended push channel, the processed content is formatted and compressed. S550, at the recommended push timing, the encapsulated information data packet is sent to the target patient's mobile terminal through the recommended push channel.

6. The method for pushing smart medical information to an internet hospital as described in claim 1, characterized in that, The method also includes a closed-loop feedback learning step after the push: S610. Monitor the target patient's interaction with the pushed medical information, and record the actual reading time, confirmation operation, ignore operation or close operation; S620. Calculate the actual effect score of this push based on the interaction behavior; S630. The complete data chain of this push is used as a training sample. The data chain includes the real-time status data before the push, the push strategy adopted, and the actual effect score. S640. Periodically use newly added training sample sets to incrementally learn and optimize the parameters in the preset state evaluation model and policy decision function.

7. The method for pushing smart medical information to an internet hospital as described in claim 3, characterized in that, The dynamically determined recommendation push strategy also includes priority processing rules for urgent information: S710. If the urgency level of the information is higher than the preset critical threshold, the recommendation push timing in the recommendation push strategy is immediately set to "execute immediately". S720. At the same time, ignore the portion of the current state stability score that is lower than the preset interference threshold, and force the transmission to be carried out using a redundant push channel combination, wherein the redundant push channel combination includes at least two independent communication channels.

8. The method for pushing smart medical information to an internet hospital as described in claim 1, characterized in that, Before acquiring the real-time status data of the target patient, an adaptive adjustment step for the patient status perception cycle is also included: S810. Set the basic sensing frequency based on the target patient's disease type, historical status fluctuation pattern, and whether the patient is currently hospitalized. S820. When abnormal fluctuations are detected in the physiological monitoring data or the terminal behavior data indicates a significant increase in patient activity, the sensing frequency is automatically increased. S830. When the current state stability score is continuously higher than the high stability threshold and the duration exceeds the preset window, the sensing frequency is automatically reduced.

9. A machine-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the smart medical information push method for Internet hospitals as described in any one of claims 1 to 8.

10. A smart medical information push system for internet hospitals, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the smart medical information push method for Internet hospitals as described in any one of claims 1 to 8.

Images