Safe medication nursing data checking and reminding system for patients with elderly eye diseases

Abstract

The invention relates to the technical field of medical health, in particular to a senile eye disease patient safety medication nursing data checking and reminding system which obtains intraocular pressure, blood sugar, heart rate and medication records of a patient in real time through a data acquisition module; a reinforcement learning model module is used for analyzing data, and an individualized drug effect-physiological index correlation model is constructed; personalized medication suggestions are generated through a dose self-adaptive adjustment module and prompts are pushed; a drug sensitivity attenuation modeling module is innovatively adopted to predict long-term drug effect change, and a clinical rule constraint module is combined to ensure medication safety; and finally, local real-time decision and cloud collaborative updating are realized through an edge computing deployment module. According to the invention, individualization, dynamics and safety of medication management are realized, and the quality and safety of medication nursing of old eye disease patients are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of medical and health technology, specifically to a system for verifying and reminding elderly patients with eye diseases about safe medication use and nursing data. Background Technology

[0002] As the global population ages, the prevalence of age-related eye diseases, such as glaucoma, cataracts, and macular degeneration, continues to rise, becoming one of the major health problems affecting the quality of life of the elderly. Elderly patients often have multiple chronic diseases and need to take multiple medications for a long time. However, due to their declining physiological functions, cognitive decline, and differences in drug metabolism, medication adherence and safety face serious challenges. Improper medication use may lead to adverse reactions such as uncontrolled intraocular pressure, blood sugar fluctuations, and abnormal heart rate, and may even cause serious complications.

[0003] Currently, there are several medication reminder systems on the market, mainly including simple reminder applications based on timers, physical reminder devices based on electronic pillboxes, and medication management software based on rule engines. These systems typically rely on preset fixed medication plans and remind patients to take their medication on time via mobile push or voice prompts. Some advanced systems also integrate basic physiological data monitoring functions, such as collecting heart rate and blood glucose data through wearable devices and generating alarms based on simple threshold judgments. In addition, some studies have attempted to introduce machine learning models for medication behavior analysis, but these focus more on data recording and static prediction, lacking dynamic adjustment and personalized optimization capabilities.

[0004] Most systems use static rules, which cannot adapt to individual patient differences and dynamic changes in their condition, resulting in rigid reminder strategies that are difficult to cope with complex situations such as drug sensitivity decline and fluctuations in physiological indicators. Secondly, existing systems often lack deep integration with real-time physiological data, making it impossible to build a drug efficacy-physiological indicator correlation model, thus making it difficult to generate accurate dosage adjustment suggestions. Therefore, there is an urgent need for an intelligent medication care system that can integrate multi-source data, dynamically optimize strategies, and ensure safety. Summary of the Invention

[0005] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a data verification and reminder system for safe medication care for elderly patients with eye diseases, which can effectively solve the problems mentioned in the existing technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a system for verifying and reminding elderly patients with eye diseases about safe medication use and nursing data, including:

[0008] The data acquisition module is used to collect patients' physiological indicators and medication history data in real time. The physiological indicators include intraocular pressure, blood glucose and heart rate. The medication history data includes drug name, dosage, medication time and adverse reaction records.

[0009] The reinforcement learning model module constructs an individualized drug efficacy-physiological indicator correlation model based on the physiological indicators and medication history data, and dynamically optimizes the medication reminder strategy through deep reinforcement learning algorithms.

[0010] The dose adaptive adjustment module generates personalized dose adjustment suggestions based on the output of the reinforcement learning model and pushes medication reminder information to patients via mobile devices;

[0011] The drug sensitivity decay modeling module dynamically corrects the drug efficacy prediction threshold by fitting the Weibull decay curve of the long-term drug use effect of patients.

[0012] The clinical rule constraint module is used to embed safety constraints in the action space to prevent the generation of adjustment suggestions for overdosing or underdosing.

[0013] The edge computing deployment module deploys lightweight models on wearable devices to enable local real-time decision-making and collaborates with cloud servers to update model parameters.

[0014] Furthermore, the data acquisition module is implemented through the following method:

[0015] Using a non-contact tonometer or implantable sensor, intraocular pressure data is collected once per hour;

[0016] Blood glucose levels were collected every 15 minutes using a continuous glucose monitor.

[0017] Wearable devices are used to collect heart rate variability data in real time, and an outlier detection algorithm is used to filter noise. The outlier detection algorithm is as follows: The principle or isolated forest algorithm.

[0018] Furthermore, the reinforcement learning model module includes:

[0019] State space definition unit: used to transform physiological indicators and medication records into standardized state vectors. ,in This is the current intraocular pressure value. This is the current blood glucose level. This is the current heart rate value. The most recent dosage. This refers to the time interval since the last medication.

[0020] Action space definition unit: used to define the set of optional actions. ,in This indicates maintaining the original dose and interval. This indicates a reduction in dosage and a longer interval. This indicates that the dosage was increased and the interval shortened, and the action selection met clinical safety constraints.

[0021] The policy network unit, based on a deep neural network architecture, receives a state vector. As input, the output is action selection, wherein the action selection is... , or The action selection logic relies on the prediction results of the individualized pharmacodynamic-physiological index correlation model;

[0022] Reward function design unit: Dynamically generates reward signals based on the actual efficacy and safety indicators of the drug. The quantization formula for the reward signal is:

[0023] ;

[0024] in, To score the effectiveness of the targeted treatment, For the ideal intraocular pressure threshold, For standard dose, For safety rating, , and These are the weighting coefficients.

[0025] Furthermore, the prediction result of the individualized drug efficacy-physiological index association model is the drug efficacy prediction value, and the specific working steps include:

[0026] Multi-source heterogeneous data were extracted from historical patient databases, including time-series data of intraocular pressure, blood glucose, and heart rate, as well as data on drug dosage and dosing intervals. Short-term change rates of each physiological indicator based on a 24-hour rolling window and long-term trend characteristics based on a 30-day moving average were calculated.

[0027] Based on the aforementioned multi-source heterogeneous data, a multi-dimensional feature vector is constructed, including the current intraocular pressure value, current blood glucose value, current heart rate value, most recent medication dose, time interval since the last medication, short-term rate of change vector, and long-term trend feature vector;

[0028] The multidimensional feature vectors are standardized by using the Z-score standardization method to eliminate dimensional differences, and the multidimensional feature vectors are reduced in dimensionality by principal component analysis.

[0029] A dual-channel deep neural network model is used for feature fusion. The first channel uses a 1D convolutional layer to extract local temporal features, while the second channel uses a Transformer encoder to capture long-range dependencies. The outputs of the two channels are weighted and fused through an attention mechanism to obtain the fused feature vector, which is then input into a fully connected layer and outputs the predicted drug efficacy value. The calculation formula is as follows:

[0030] ;

[0031] in, This is the predicted value for drug efficacy. It is the sigmoid activation function. The weights of the fully connected layer, These are the bias parameters for the fully connected layer. The fused feature vector;

[0032] Based on real-time physiological data and medication feedback, the Adam optimizer is used to update the model weights every 24 hours, and a sliding window mechanism is used to suppress noise interference.

[0033] Furthermore, the specific implementation of the policy network unit includes the following steps:

[0034] Construct a policy network model based on a deep neural network, with the input layer receiving a state vector. It includes five dimensions of features: current intraocular pressure, blood glucose level, heart rate, most recent medication dose, and time interval since last medication.

[0035] The hidden layers employ a two-layer fully connected structure. The first layer contains 64 neurons using the ReLU activation function, and the second layer contains 3 neurons using the Softmax activation function. The output layer generates the action selection probability distribution. , and ;

[0036] Define the policy gradient loss function, and its calculation formula is as follows:

[0037] ;

[0038] in, For the parameter vector of the policy network, Let be the probability of choosing action a in state s. For the dominant function, To estimate the expected value of a state-action-reward sequence;

[0039] The Adam optimizer is used for parameter updates, with a learning rate of α=0.001 and a batch size of 32. Gradient updates are performed after every 10 state-action-reward sequences are received.

[0040] During the training of the policy network model, an experience replay buffer mechanism is used to store the most recent 1000 state-action-reward sequences, and 64 samples are randomly sampled for training each time an update is performed.

[0041] During the inference phase, based on the current state vector The action selection probability distribution is calculated through a policy network, and the action with the highest probability is selected as the final decision. .

[0042] Furthermore, the specific implementation of the drug sensitivity attenuation modeling module includes the following steps:

[0043] Collect patients' long-term medication records and corresponding efficacy data, and extract the correlation between drug use time series and efficacy indicators;

[0044] Based on time series data, the maximum likelihood estimation method is used to fit the Weibull distribution model to determine the shape parameter k and scale parameter of sensitivity decay. Establish a mathematical expression for how drug efficacy changes over time:

[0045] ;

[0046] in, This is the sensitivity attenuation coefficient;

[0047] Sensitivity attenuation coefficient Predicted efficacy values ​​output by the personalized pharmacodynamic-physiological index association model Multiply the results to generate a revised dynamic efficacy prediction threshold, which is used as the decision input for the reinforcement learning model.

[0048] Furthermore, the specific steps for embedding security constraints include:

[0049] Transform clinical medication guidelines into mathematical constraints, including minimum / maximum dose limits, lower limits for dosing intervals, contraindications for specific drug combinations, and individualized side effect thresholds;

[0050] In the set of optional actions In the process, actions that violate safety rules are eliminated based on the real-time patient status;

[0051] The design constrains the reward function, applies negative weights to high-risk actions, reduces the policy priority, and finally performs secondary verification through the clinical rule constraint module. If an action violates the safety constraints, it is forcibly replaced with the default safe action.

[0052] Furthermore, the edge computing deployment module includes:

[0053] The local device is used to complete physiological data collection, preliminary feature extraction, and emergency decision calculation.

[0054] The cloud server is used to store long-term medication records and perform global training and policy updates for the model.

[0055] The technical solution provided by this invention has the following advantages compared with the known prior art:

[0056] This invention, by introducing a deep reinforcement learning model, achieves a leap from static rule reminders to dynamic personalized decision-making; the system can continuously learn and optimize medication strategies based on the patient's real-time physiological data and medication history, and automatically generate dosage adjustment suggestions that fit individual needs, thereby significantly improving the effectiveness and safety of treatment.

[0057] This invention constructs a dual safety mechanism consisting of drug sensitivity decay prediction and clinical rule constraints. By modeling the long-term drug efficacy decay trend, it predicts the risk of insufficient dosage and embeds it into the decision-making process of strict constraints of clinical drug use guidelines. This effectively eliminates the risks of overdose and hidden treatment failure, and ensures the safety and reliability of the system output. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0059] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0061] The present invention will be further described below with reference to embodiments.

[0062] Example:

[0063] Reference Figure 1 A data verification and reminder system for safe medication use and nursing care in elderly patients with eye diseases, including:

[0064] The data acquisition module is used to collect patients' physiological indicators and medication history data in real time. The physiological indicators include intraocular pressure, blood glucose and heart rate, and the medication history data includes drug name, dosage, medication time and adverse reaction records.

[0065] The reinforcement learning model module constructs an individualized drug efficacy-physiological indicator correlation model based on physiological indicators and medication history data, and dynamically optimizes the medication reminder strategy through deep reinforcement learning algorithms;

[0066] The dose adaptive adjustment module generates personalized dose adjustment suggestions based on the output of the reinforcement learning model and pushes medication reminder information to patients via mobile devices;

[0067] The drug sensitivity decay modeling module dynamically corrects the drug efficacy prediction threshold by fitting the Weibull decay curve of the long-term drug use effect of patients.

[0068] The clinical rule constraint module is used to embed safety constraints in the action space to prevent the generation of adjustment suggestions for overdosing or underdosing.

[0069] The edge computing deployment module deploys lightweight models on wearable devices to enable local real-time decision-making and collaborates with cloud servers to update model parameters.

[0070] Specifically, the system dynamically adjusts medication strategies through the aforementioned modules to reduce human error; through multi-module collaboration, it ensures that medication recommendations are both personalized and meet clinical safety standards; and edge computing deployment improves system response speed and guarantees real-time decision-making capabilities.

[0071] Furthermore, the data acquisition module is implemented through the following methods:

[0072] Using a non-contact tonometer or implantable sensor, intraocular pressure data is collected once per hour;

[0073] Blood glucose levels were collected every 15 minutes using a continuous glucose monitor.

[0074] Wearable devices are used to collect heart rate variability data in real time, and an outlier detection algorithm is used to filter noise. The outlier detection algorithm is as follows: The principle or isolated forest algorithm.

[0075] Specifically, the data acquisition module provides high-frequency, high-precision monitoring of physiological indicators, ensuring data reliability and real-time performance, reducing data acquisition errors, and improving the quality of model input; it reduces the burden on patients and enhances system availability through automated acquisition; and the outlier detection algorithm effectively removes noise and avoids erroneous decisions.

[0076] Furthermore, the reinforcement learning model module includes:

[0077] State space definition unit: used to transform physiological indicators and medication records into standardized state vectors. ,in This is the current intraocular pressure value. This is the current blood glucose level. This is the current heart rate value. The most recent dosage. This refers to the time interval since the last medication.

[0078] Action space definition unit: used to define the set of optional actions. ,in This indicates maintaining the original dose and interval. This indicates a reduction in dosage and a longer interval. This indicates that the dosage was increased and the interval shortened, and the action selection met clinical safety constraints.

[0079] The policy network unit, based on a deep neural network architecture, receives a state vector. As input, the output is the action selection, and the action selection is... , or The action selection logic relies on the prediction results of the individualized pharmacodynamic-physiological index correlation model;

[0080] Reward function design unit: Dynamically generates reward signals based on the actual efficacy and safety indicators of the drug. The quantization formula for the reward signal is:

[0081] ;

[0082] in, To score the effectiveness of the targeted treatment, For the ideal intraocular pressure threshold, For standard dose, For safety rating, , and These are the weighting coefficients.

[0083] Specifically, the target treatment outcome score and security rating The calculation steps are as follows:

[0084] Targeted treatment outcome score Calculation:

[0085] Let the predicted efficacy value be The formula for scoring the target treatment effect is:

[0086] ;

[0087] Security rating Calculation:

[0088] Define the current physiological index value as (intraocular pressure), (blood sugar), (Heart rate), the corresponding threshold range is:

[0089] Safe range of intraocular pressure: ;

[0090] Safe blood sugar range: ;

[0091] Heart rate safe zone: ;

[0092] The formula for calculating the safety of each indicator is as follows:

[0093] ;

[0094] ;

[0095] ;

[0096] in, For intraocular pressure safety, For blood sugar safety, For heart rate safety;

[0097] The safety scoring formula is:

[0098] .

[0099] Furthermore, the prediction results of the individualized drug efficacy-physiological index association model are the predicted drug efficacy values. The specific steps include:

[0100] Multi-source heterogeneous data were extracted from historical patient databases, including time-series data of intraocular pressure, blood glucose, and heart rate, as well as data on drug dosage and dosing intervals. Short-term change rates of each physiological indicator based on a 24-hour rolling window and long-term trend characteristics based on a 30-day moving average were calculated.

[0101] A multidimensional feature vector is constructed based on multi-source heterogeneous data, including current intraocular pressure, current blood glucose, current heart rate, most recent medication dose, time interval since last medication, short-term rate of change vector, and long-term trend feature vector;

[0102] The multidimensional feature vectors are standardized by using the Z-score standardization method to eliminate dimensional differences, and the multidimensional feature vectors are reduced in dimensionality by principal component analysis.

[0103] A dual-channel deep neural network model is used for feature fusion. The first channel uses a 1D convolutional layer to extract local temporal features, while the second channel uses a Transformer encoder to capture long-range dependencies. The outputs of the two channels are weighted and fused through an attention mechanism to obtain the fused feature vector, which is then input into a fully connected layer and outputs the predicted drug efficacy value. The calculation formula is as follows:

[0104] ;

[0105] in, This is the predicted value for drug efficacy. It is the sigmoid activation function. The weights of the fully connected layer, These are the bias parameters for the fully connected layer. The fused feature vector;

[0106] Based on real-time physiological data and medication feedback, the Adam optimizer is used to update the model weights every 24 hours, and a sliding window mechanism is used to suppress noise interference.

[0107] Specifically, deep learning models accurately predict drug efficacy and capture short-term and long-term trends in physiological indicators, improving the accuracy and robustness of drug efficacy prediction. At the same time, dual-channel neural networks integrate temporal and long-term dependency features to enhance the model's expressive power, and regular model updates ensure adaptation to dynamic changes in patients.

[0108] Furthermore, the specific implementation of the policy network unit includes the following steps:

[0109] Construct a policy network model based on a deep neural network, with the input layer receiving a state vector. It includes five dimensions of features: current intraocular pressure, blood glucose level, heart rate, most recent medication dose, and time interval since last medication.

[0110] The hidden layers employ a two-layer fully connected structure. The first layer contains 64 neurons using the ReLU activation function, and the second layer contains 3 neurons using the Softmax activation function. The output layer generates the action selection probability distribution. , and ;

[0111] Define the policy gradient loss function, and its calculation formula is as follows:

[0112] ;

[0113] in, For the parameter vector of the policy network, Let be the probability of choosing action a in state s. For the dominant function, To estimate the expected value of a state-action-reward sequence;

[0114] The Adam optimizer is used for parameter updates, with a learning rate of α=0.001 and a batch size of 32. Gradient updates are performed after every 10 state-action-reward sequences are received.

[0115] During the training of the policy network model, an experience replay buffer mechanism is adopted to store the most recent 1000 state-action-reward sequences, and 64 samples are randomly sampled for training each time an update is performed;

[0116] During the inference phase, based on the current state vector The action selection probability distribution is calculated through a policy network, and the action with the highest probability is selected as the final decision. .

[0117] Furthermore, the specific implementation of the drug sensitivity decay modeling module includes the following steps:

[0118] Collect patients' long-term medication records and corresponding efficacy data, and extract the correlation between drug use time series and efficacy indicators;

[0119] Based on time series data, the maximum likelihood estimation method is used to fit the Weibull distribution model to determine the shape parameter k and scale parameter of sensitivity decay. Establish a mathematical expression for how drug efficacy changes over time:

[0120] ;

[0121] in, This is the sensitivity attenuation coefficient;

[0122] Sensitivity attenuation coefficient Predicted efficacy values ​​output by the personalized pharmacodynamic-physiological index association model Multiply the results to generate a revised dynamic efficacy prediction threshold, which is used as the decision input for the reinforcement learning model.

[0123] Furthermore, the specific steps for embedding security constraints include:

[0124] Clinical medication guidelines are transformed into mathematical constraints, including minimum / maximum dose limits, lower limits for dosing intervals, contraindications to specific drug combinations, and individualized side effect thresholds, ensuring that the system output meets medical safety standards.

[0125] In the set of optional actions In the process, actions that violate safety rules are eliminated based on the real-time patient status;

[0126] Specifically, actions that violate safety rules include excluding [the product / service] when the current dose is already below the minimum safe dose. (Reduce dosage); exclude when predicted side effect risk exceeds preset threshold. (Increase dosage);

[0127] The design incorporates a constraint reward function that applies negative weights to high-risk actions, reducing the policy priority. Finally, a secondary validation is performed using a clinical rule constraint module. If an action violates safety constraints, it is forcibly replaced with a default safe action. The constraint reward function guides the model to prioritize safe actions, reducing the risk of side effects.

[0128] Specifically, high-risk actions include overdosing or underdosing;

[0129] Furthermore, the edge computing deployment module includes:

[0130] The local device is used to complete physiological data collection, preliminary feature extraction, and emergency decision calculation.

[0131] The cloud server is used to store long-term medication records and perform global training and policy updates for the model.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A system for verifying and reminding elderly patients with eye diseases regarding safe medication use and nursing data, characterized in that: include: The data acquisition module is used to collect patients' physiological indicators and medication history data in real time. The physiological indicators include intraocular pressure, blood glucose and heart rate. The medication history data includes drug name, dosage, medication time and adverse reaction records. The reinforcement learning model module constructs an individualized drug efficacy-physiological indicator correlation model based on the physiological indicators and medication history data, and dynamically optimizes the medication reminder strategy through deep reinforcement learning algorithms. The dose adaptive adjustment module generates personalized dose adjustment suggestions based on the output of the reinforcement learning model and pushes medication reminder information to patients via mobile devices; The drug sensitivity decay modeling module dynamically corrects the drug efficacy prediction threshold by fitting the Weibull decay curve of the long-term drug use effect of patients. The clinical rule constraint module is used to embed safety constraints in the action space to prevent the generation of adjustment suggestions for overdosing or underdosing. The edge computing deployment module deploys lightweight models on wearable devices to enable local real-time decision-making and collaborates with cloud servers to update model parameters.

2. The data verification and reminder system for safe medication use in elderly patients with eye diseases according to claim 1, characterized in that, The data acquisition module is implemented through the following method: Using a non-contact tonometer or implantable sensor, intraocular pressure data is collected once per hour; Blood glucose levels were collected every 15 minutes using a continuous glucose monitor. Wearable devices are used to collect heart rate variability data in real time, and an outlier detection algorithm is used to filter noise. The outlier detection algorithm is as follows: The principle or isolated forest algorithm.

3. The data verification and reminder system for safe medication use in elderly patients with eye diseases according to claim 1, characterized in that, The reinforcement learning model module includes: State space definition unit: used to transform physiological indicators and medication records into standardized state vectors. ,in This is the current intraocular pressure value. This is the current blood glucose level. This is the current heart rate value. The most recent dosage. This refers to the time interval since the last medication. Action space definition unit: used to define the set of optional actions. ,in This indicates maintaining the original dose and interval. This indicates a reduction in dosage and a longer interval. This indicates that the dosage was increased and the interval shortened, and the action selection met clinical safety constraints. The policy network unit, based on a deep neural network architecture, receives a state vector. As input, the output is action selection, wherein the action selection is... , or The action selection logic relies on the prediction results of the individualized pharmacodynamic-physiological index correlation model; Reward function design unit: Dynamically generates reward signals based on the actual efficacy and safety indicators of the drug. The quantization formula for the reward signal is: ； in, To score the effectiveness of the targeted treatment, For the ideal intraocular pressure threshold, For standard dose, For safety rating, , and These are the weighting coefficients.

4. The data verification and reminder system for safe medication use in elderly patients with eye diseases according to claim 1, characterized in that, The prediction results of the individualized drug efficacy-physiological index correlation model are the predicted drug efficacy values. The specific working steps include: Multi-source heterogeneous data were extracted from historical patient databases, including time-series data of intraocular pressure, blood glucose, and heart rate, as well as data on drug dosage and dosing intervals. Short-term change rates of each physiological indicator based on a 24-hour rolling window and long-term trend characteristics based on a 30-day moving average were calculated. Based on the aforementioned multi-source heterogeneous data, a multi-dimensional feature vector is constructed, including the current intraocular pressure value, current blood glucose value, current heart rate value, most recent medication dose, time interval since the last medication, short-term rate of change vector, and long-term trend feature vector; The multidimensional feature vectors are standardized by using the Z-score standardization method to eliminate dimensional differences, and the multidimensional feature vectors are reduced in dimensionality by principal component analysis. A dual-channel deep neural network model is used for feature fusion. The first channel uses a 1D convolutional layer to extract local temporal features, while the second channel uses a Transformer encoder to capture long-range dependencies. The outputs of the two channels are weighted and fused through an attention mechanism to obtain the fused feature vector, which is then input into a fully connected layer and outputs the predicted drug efficacy value. The calculation formula is as follows: ； in, This is the predicted value for drug efficacy. It is the sigmoid activation function. The weights of the fully connected layer, These are the bias parameters for the fully connected layer. The fused feature vector; Based on real-time physiological data and medication feedback, the Adam optimizer is used to update the model weights every 24 hours, and a sliding window mechanism is used to suppress noise interference.

5. The data verification and reminder system for safe medication use in elderly patients with eye diseases according to claim 3, characterized in that, The specific implementation of the policy network unit includes the following steps: Construct a policy network model based on a deep neural network, with the input layer receiving a state vector. It includes five dimensions of features: current intraocular pressure, blood glucose level, heart rate, most recent medication dose, and time interval since last medication. The hidden layers employ a two-layer fully connected structure. The first layer contains 64 neurons using the ReLU activation function, and the second layer contains 3 neurons using the Softmax activation function. The output layer generates the action selection probability distribution. , and ; Define the policy gradient loss function, and its calculation formula is as follows: ； in, For the parameter vector of the policy network, Let be the probability of choosing action a in state s. For the dominant function, To estimate the expected value of a state-action-reward sequence; The Adam optimizer is used for parameter updates, with a learning rate of α=0.001 and a batch size of 32. Gradient updates are performed after every 10 state-action-reward sequences are received. During the training of the policy network model, an experience replay buffer mechanism is used to store the most recent 1000 state-action-reward sequences, and 64 samples are randomly sampled for training each time an update is performed. During the inference phase, based on the current state vector The action selection probability distribution is calculated through a policy network, and the action with the highest probability is selected as the final decision. .

6. The data verification and reminder system for safe medication use in elderly patients with eye diseases according to claim 1, characterized in that, Specific implementation of the drug sensitivity attenuation modeling module Includes the following steps: Collect patients' long-term medication records and corresponding efficacy data, and extract the correlation between drug use time series and efficacy indicators; Based on time series data, the maximum likelihood estimation method is used to fit the Weibull distribution model to determine the shape parameter k and scale parameter of sensitivity decay. Establish a mathematical expression for how drug efficacy changes over time: ； in, This is the sensitivity attenuation coefficient; Sensitivity attenuation coefficient Predicted efficacy values ​​output by the personalized pharmacodynamic-physiological index association model Multiply the results to generate a revised dynamic efficacy prediction threshold, which is used as the decision input for the reinforcement learning model.

7. The data verification and reminder system for safe medication use in elderly patients with eye diseases according to claim 1, characterized in that, The specific steps for embedding security constraints include: Transform clinical medication guidelines into mathematical constraints, including minimum / maximum dose limits, lower limits for dosing intervals, contraindications for specific drug combinations, and individualized side effect thresholds; In the set of optional actions In the process, actions that violate safety rules are eliminated based on the real-time patient status; The design constrains the reward function, applies negative weights to high-risk actions, reduces the policy priority, and finally performs secondary verification through the clinical rule constraint module. If an action violates the safety constraints, it is forcibly replaced with the default safe action.

8. The data verification and reminder system for safe medication use in elderly patients with eye diseases according to claim 1, characterized in that, The edge computing deployment module includes: The local device is used to complete physiological data collection, preliminary feature extraction, and emergency decision calculation. The cloud server is used to store long-term medication records and perform global training and policy updates for the model.

Images