An intelligent reminding method and system for assisting the elderly in taking medicine and health management

Abstract

The application relates to an intelligent reminding method and system for assisting the elderly in taking medicine and health management. The method comprises the following steps: unsupervised learning of basic perception data to generate an initial response and background environmental characteristics, combining a medicine prescription instruction to generate an individualized response delay benchmark line, an environmental acoustic background noise spectrum and a medicine taking time window; adaptive multi-modal reminding scheduling is performed based on multi-dimensional sensor data to generate an optimal strategy combination vector to control a reminding device group; the user action state is judged through a hidden Markov model, secondary reminding and cognitive verification are performed when an anomaly occurs, and a light-seeking guidance instruction is issued when the score is low; after taking medicine, weak network causal synchronization is completed, and data instructions are broadcast to local network nodes. The method can accurately adapt to the medicine taking habits of the elderly, and improve the success rate of medicine taking reminding and the reliability of health management.

Description

Technical Field

[0001] This invention relates to the field of intelligent health management technology, and in particular to an intelligent reminder method and system for assisting the elderly in medication use and health management. Background Technology

[0002] With the global population aging rapidly, the daily management of chronic diseases is becoming increasingly important, and regular medication is key to controlling these conditions. Addressing issues such as memory decline, sensory function impairment, and poor medication adherence among the elderly, smart medication reminders based on the Internet of Things and interactive terminals are becoming an important tool for home-based elderly care. These reminders can monitor medication intake on time, reduce the risk of missed doses, overdosing, or incorrect dosage, and ensure health and safety.

[0003] However, existing reminder technologies have significant shortcomings in home-based aging-in-place scenarios: traditional devices have limited alarm strength and sensory dimensions, making them susceptible to environmental interference or causing discomfort and alarm fatigue in the elderly; medication monitoring relies heavily on simple mechanical triggers, failing to identify behavioral gaps, easily generating false medication records, and not forming a true swallowing loop; the system is highly dependent on centralized cloud networks, and network outages or signal blind spots can easily lead to asynchronous device data and logical conflicts, posing a risk of duplicate medication. Summary of the Invention

[0004] Therefore, there is a need for an intelligent medication reminder method and system for the elderly that is adaptable to the environment, accurately verifies medication behavior, and has network disaster recovery capabilities.

[0005] Firstly, this application provides an intelligent reminder method to assist elderly people in medication use and health management, including:

[0006] S1. Perform unsupervised learning on the acquired basic perception data to generate initial reaction features and background environment features. Based on the acquired medication prescription instructions, perform time-space mapping on the initial reaction features and background environment features, extract the information and environmental parameters corresponding to the time window and the target drug warehouse, and generate individualized reaction delay baseline, ambient acoustic background noise spectrum and medication time window.

[0007] S2. Acquire multi-dimensional sensor data of the current environment, and perform adaptive multi-modal reminder scheduling based on the ambient acoustic background noise spectrum, medication time window and multi-dimensional sensor data. Calculate the reach probability and annoyance cost of auditory mode, tactile mode and visual mode, and construct a utility objective function for solution to generate the optimal strategy combination vector. The optimal strategy combination vector is used to control the reminder device group to remind the user respectively.

[0008] S3. Obtain the user's action signal after being reminded, perform state probability fitting and temporal decoding on the user's action signal through a hidden Markov model to generate a physical temporal state sequence, and perform anomaly judgment on the physical temporal state sequence based on the temporal integrity comparison rule to generate action state judgment result; wherein, the action state judgment result is a medication-taken status confirmation tag or an action abnormal interruption tag.

[0009] S4. If the action status assessment result is an action abnormal interruption label, a user reminder signal is generated and sent to the reminder device group. Based on the action abnormal interruption label and the individualized response delay baseline, the user's voice response signal is cognitively verified to generate a cognitive verification comprehensive score. When the cognitive verification comprehensive score is lower than the set cognitive verification threshold, the spatial positioning distance of the reminder device group is obtained. Based on the spatial positioning distance, a light-guided instruction is generated and sent to the reminder device group. After receiving the medication retrieval signal from the reminder device group, the action abnormal interruption label is updated to a medication-taken status confirmation label.

[0010] S5. Based on the medication status confirmation tag, perform causal synchronization of weak network status, combine with the acquired local vector clock to encapsulate data, generate data instructions including synchronization data packets, and broadcast the data instructions to other node devices in the local area network; wherein, the data instructions are used to update the medication status and vector clock data of other node devices.

[0011] Secondly, this application also provides an intelligent reminder method and system to assist the elderly in medication and health management, for implementing the method described in the first aspect, including:

[0012] The baseline feature extraction module is used to perform unsupervised learning on the acquired basic perception data to generate initial reaction features and background environment features. Based on the acquired medication prescription instructions, the initial reaction features and background environment features are mapped in time and space. Information and environmental parameters corresponding to the time window and the target pharmacy are extracted to generate individualized reaction delay baseline, ambient acoustic background noise spectrum and medication time window.

[0013] The multimodal scheduling module is used to acquire multidimensional sensor data of the current environment, and to perform adaptive multimodal reminder scheduling based on the ambient acoustic background noise spectrum, medication time window and multidimensional sensor data. It calculates the reach probability and annoyance cost of auditory mode, tactile mode and visual mode and constructs a utility objective function for solution to generate the optimal strategy combination vector. The optimal strategy combination vector is used to control the reminder device group to remind the user respectively.

[0014] The action tracking and analysis module is used to acquire user action signals after being reminded, perform state probability fitting and temporal decoding on the user action signals through a hidden Markov model to generate a physical temporal state sequence, and perform anomaly analysis on the physical temporal state sequence based on the temporal integrity comparison rule to generate action state analysis results.

[0015] The cognitive verification guidance module is used to generate a user reminder signal and send it to the reminder device group if the action status judgment result is an action abnormal interruption label. Based on the action abnormal interruption label and the individualized reaction delay baseline, it performs cognitive verification on the feedback user voice response signal and generates a cognitive verification comprehensive score. When the cognitive verification comprehensive score is lower than the set cognitive verification threshold, it obtains the spatial positioning distance of the reminder device group, generates a light-touch guidance instruction based on the spatial positioning distance and sends it to the reminder device group. After receiving the medication retrieval signal from the reminder device group, it updates the action abnormal interruption label to a medication taken status confirmation label.

[0016] The weak network status synchronization module is used to perform causal synchronization of weak network status based on the medication status confirmation tag. It combines the acquired local vector clock to encapsulate data and generate data instructions including synchronization data packets, and broadcasts the data instructions to other node devices in the local area network.

[0017] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform any of the above steps.

[0018] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs any of the above steps.

[0019] The aforementioned intelligent reminder method and system for assisting elderly people in medication and health management, through unsupervised learning and adaptive multimodal scheduling mechanisms, can dynamically output the optimal combination of sound, light, and tactile reminder strategies based on individual user response differences and current environmental characteristics. This effectively improves the reach of reminder signals and reduces environmental disturbance. By decoding action sequences using a Hidden Markov Model and combining it with user voice responses for cognitive verification, accurate judgment of users' actual medication behavior can be achieved, reducing the probability of misjudgments caused by traditional devices relying on single-state monitoring. A phototactic guidance mechanism is introduced to address abnormal action interruptions, combined with causal synchronization processing in weak network conditions. This not only improves the convenience of medication retrieval for individuals with mild cognitive impairment but also enables reliable synchronization of medication status across multiple devices in complex home network environments, thereby comprehensively improving the medication management level and adherence of the elderly population. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an intelligent reminder method for assisting elderly people in medication use and health management, as described in one embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of an implementation environment provided in an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of the structure of an intelligent reminder system for assisting elderly people in medication and health management according to one embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0025] This application provides an intelligent reminder method to assist elderly people in medication use and health management, which can be applied to, for example... Figure 2 In the application environment shown, the management device 101 communicates wirelessly with the reminder device group 102. A data storage system can store the data that the management device 101 and the reminder device group 102 need to process. The data storage system can be integrated into the management device 101 or located in the cloud or on other network servers. The management device 101 can be, but is not limited to, various personal computers, laptops, smartphones, and tablets. The reminder device group 102 can be, but is not limited to, smart bracelets, smart pillboxes, head-mounted devices, etc.

[0026] In one exemplary embodiment, such as Figure 1 As shown, an intelligent reminder method is provided to assist the elderly in medication and health management. This method is applied to... Figure 2 Taking the management device 101 as an example, the following steps may be included:

[0027] S1. Perform unsupervised learning on the acquired basic perception data to generate initial reaction features and background environment features. Based on the acquired medication prescription instructions, perform time-space mapping on the initial reaction features and background environment features, extract the information and environmental parameters corresponding to the time window and target pharmacy, and generate individualized reaction delay baseline, ambient acoustic background noise spectrum and medication time window.

[0028] Specifically, medication prescription instructions can be set dosing time data and drug category data. Time-space mapping can be the coordinate transformation calculation of data in the time and space dimensions. Target pharmacy can be the physical location where the drugs are stored. Personalized response delay baseline can be the statistical value of the time difference between the user receiving a signal and taking an action. Ambient acoustic background noise spectrum can be the frequency distribution of background sound.

[0029] Optionally, the management device can separate basic sensing data, action data, and background white noise data, outputting initial reaction characteristics and background environmental characteristics. The management device can extract dosing time data and drug number data from the medication prescription instruction to establish a prescription timeline. The management device can introduce initial reaction characteristics into the prescription timeline and physical coordinate system, performing time-space mapping calculations. The management device can extract the reaction time record and coordinate data of the target drug compartment within a time window. The management device can calculate environmental parameters, extracting the mean and variance of sound pressure level. The management device can establish an environmental acoustic background noise spectrum based on the mean and variance of sound pressure level. The management device can fit the time difference data from receiving the instruction to contacting the drug box, outputting an individualized reaction delay baseline. The management device can perform numerical expansion on the prescription timeline using the individualized reaction delay baseline. The management device can add a numerical range beyond the dosing time base, outputting a medication time window.

[0030] S2. Acquire multi-dimensional sensor data of the current environment, and perform adaptive multi-modal reminder scheduling based on the ambient acoustic background noise spectrum, medication time window and multi-dimensional sensor data. Calculate the reach probability and annoyance cost of auditory modality, tactile modality and visual modality, and construct a utility objective function for solution to generate the optimal strategy combination vector. The optimal strategy combination vector is used to control the reminder device group to remind the user respectively.

[0031] Specifically, multidimensional sensor data can be a dataset of acoustic signals, photometric signals, and motion variance signals. Adaptive multimodal alert scheduling can be a decision-making calculation that adjusts the signal output method based on environmental parameters. Auditory modality can be an alert method that outputs acoustic signals through a speaker. Tactile modality can be an alert method that outputs vibration signals through a motor. Visual modality can be an alert method that outputs light beam signals through a light-emitting diode. The probability of reaching the signal can be the numerical value of the likelihood of the signal being received. The interference cost can be a quantified numerical value of the interference the signal causes to the environment. The utility objective function can be the equation for calculating the probability of reaching the signal and the interference cost parameters. The optimal policy combination vector can be the combination parameters that maximize the utility objective function.

[0032] Optionally, the management device can perform dimensionality reduction calculations on multi-dimensional sensor data to extract illuminance values, sound pressure level values, motion variance values, and spatial distance values ​​from the multi-dimensional sensor data. The management device can input the sound pressure level values ​​and the ambient acoustic background noise spectrum into an activation model to calculate the reach probability corresponding to the auditory modality. The management device can input the sound pressure level values ​​and the ambient acoustic background noise spectrum into a penalty model to calculate the disturbance cost corresponding to the auditory modality. The management device can establish a visual attenuation mapping relationship based on illuminance values ​​and spatial distance values ​​to calculate the reach probability corresponding to the visual modality. The management device extracts the vibration intensity values ​​of the motor. The management device can input the motor vibration intensity values ​​and motion variance values ​​into an activation model to calculate the reach probability corresponding to the tactile modality. The management device can establish a utility objective function, substitute the combination parameters of the auditory, visual, and tactile modalities into the utility objective function, and calculate the corresponding benefit values ​​based on the utility objective function. The management device can select the combination parameters corresponding to the upper limit of the benefit values ​​and output the optimal strategy combination vector.

[0033] S3. Obtain the user's action signal after being reminded, perform state probability fitting and temporal decoding on the user's action signal through a hidden Markov model to generate a physical temporal state sequence, and perform anomaly judgment on the physical temporal state sequence based on the temporal integrity comparison rule to generate action state judgment results.

[0034] Specifically, the action status assessment result is either a medication-confirmed status label or an action abnormality interruption label.

[0035] Specifically, a Hidden Markov Model (HMM) can be a statistical model that includes a transition matrix and an observation matrix. A physical temporal state sequence can be a dataset with state labels arranged chronologically. A temporal completeness comparison rule can be a criterion for evaluating the coverage of a state sequence. A medication-taken state confirmation label can be recorded data indicating the completion of an action loop. An abnormal action interruption label can be recorded data indicating action stagnation.

[0036] Optionally, the management device can align and merge motion parameters with the level signals of the pillbox sensor based on timestamps to construct a multi-source observation sequence. The management device can perform state probability fitting operations on the multi-source observation sequence based on a Hidden Markov Model, calculate the emission probability distribution of multi-source observation data at each hidden state at each time point, and establish a state emission probability matrix. The management device can apply a decoding algorithm to the state emission probability matrix to perform temporal decoding calculations, find the hidden state transition path corresponding to the maximum posterior probability, and generate a physical temporal state sequence. The management device can substitute the physical temporal state sequence into the temporal integrity comparison rules to perform matching operations. Under the condition that the physical temporal state sequence covers a set baseline state closed loop and the dwell time of each state does not exceed a set threshold boundary, the management device can perform anomaly analysis and generate an action state analysis result including a medication-taken state confirmation tag. Under the condition that the physical temporal state sequence does not cover the baseline state closed loop or the dwell time exceeds a set threshold boundary, the management device can perform anomaly analysis and generate an action state analysis result including an action anomaly interruption tag.

[0037] S4. If the action status assessment result is an action abnormality interruption label, a user reminder signal is generated and sent to the reminder device group. Based on the action abnormality interruption label and the individualized response delay baseline, the user's voice response signal is cognitively verified to generate a cognitive verification comprehensive score. When the cognitive verification comprehensive score is lower than the set cognitive verification threshold, the spatial positioning distance of the reminder device group is obtained. Based on the spatial positioning distance, a light-guided instruction is generated and sent to the reminder device group. After receiving the medication retrieval signal from the reminder device group, the action abnormality interruption label is updated to a medication-taken status confirmation label.

[0038] Specifically, cognitive verification can be a computational mechanism that quantitatively evaluates the semantics of speech and text, as well as the difference in reaction time. The cognitive verification overall score can be a quantitative evaluation value reflecting the response state to voice questions and answers. The cognitive verification threshold can be the boundary value that divides the cognitive verification overall score interval. The spatial positioning distance can be the physical spatial distance between the user and the group of reminder devices. The light-guiding command can be an optical control message that controls the light-emitting component to execute brightness output.

[0039] Optionally, the management device can generate a user reminder signal when the action state assessment result is equivalent to an action abnormality interruption label. The management device can send the user reminder signal to the reminder device group to initiate voice question-and-answer interaction. The management device can receive the user's voice response signal and perform cognitive verification calculations, converting the user's voice response signal into a hidden layer representation based on a natural language processing model. The management device can substitute this hidden layer representation and the expected target text feature vector into a cross-attention function to calculate the semantic matching quantity. The management device can calculate the time difference variable of the acquired user's voice response signal, and perform an exponential decay operation on this time difference variable in conjunction with an individualized response delay baseline to calculate a penalty factor. The management device can perform a weighted calculation based on the semantic matching quantity and the penalty factor to generate a comprehensive cognitive verification score. The management device can compare the cognitive verification threshold with the comprehensive cognitive verification score. If the comprehensive cognitive verification score is lower than the set cognitive verification threshold, the management device reads the signal strength indication parameter of the reminder device group and calculates the spatial positioning distance based on a spatial filtering algorithm. The management device can extract the negative correlation parameter of the spatial positioning distance to calculate the target luminous intensity and generate a light-guided instruction based on the luminous intensity parameter. The management device can send light-guided commands to the reminder device group to drive the light-emitting components. Upon receiving a medication dispensing signal from the reminder device group, the management device can update the abnormal interruption tag to a medication-taken confirmation tag.

[0040] S5. Based on the medication status confirmation tag, perform causal synchronization of weak network status, combine with the acquired local vector clock to encapsulate data, generate data instructions including synchronization data packets, and broadcast the data instructions to other node devices in the local area network.

[0041] Specifically, data commands are used to update the medication status and vector clock data of other node devices.

[0042] Specifically, weak network causal synchronization can be a mechanism for maintaining the logical consistency of distributed device states under network disconnection conditions. A local vector clock can be an array variable recording the sequence of events occurring within the device. Synchronization data packets can be communication payloads containing state identifiers and time-series arrays. Node devices can be a collection of physical hardware within a local area network configured with processing processors and communication hardware.

[0043] Optionally, upon obtaining a medication status confirmation tag, the management device can read the local vector clock array variable from its local memory, perform numerical increment calculations on its own device event components within the local vector clock, modify and output the current stage's local vector clock data. The management device can perform byte alignment and concatenation processing on the extracted medication status confirmation tag data and the current stage's local vector clock data, perform data encapsulation operations, and generate data instructions containing synchronization data packets. The management device can calculate its own hardware's remaining power parameters, establish an energy regulation function, and calculate the communication broadcast frequency. Based on the communication broadcast frequency, the management device can drive the underlying communication hardware to emit radio frequency signals, broadcasting the data instructions, including synchronization data packets, to other node devices in the local area network environment.

[0044] This embodiment provides an intelligent reminder method for assisting elderly people in medication and health management. Through unsupervised learning and an adaptive multimodal scheduling mechanism, it dynamically outputs the optimal combination of sound, light, and tactile reminder strategies based on individual user response differences and current environmental characteristics. This effectively improves the reach of reminder signals and reduces environmental disturbance. By decoding action sequences using a Hidden Markov Model and combining this with user voice responses for cognitive verification, it achieves accurate judgment of the user's actual medication behavior, reducing the probability of misjudgment caused by traditional devices relying on single-state monitoring. A phototactic guidance mechanism is introduced to address abnormal action interruptions, combined with causal synchronization processing in weak network conditions. This not only improves the convenience of medication retrieval for individuals with mild cognitive impairment but also enables reliable synchronization of medication status across multiple devices in complex home network environments, thereby comprehensively improving the medication management level and adherence of the elderly population.

[0045] In one embodiment, S1 may include:

[0046] S11. Perform unsupervised feature clustering on the acquired basic perception data to separate daily action lag data and environmental background white noise, and generate initial response features and background environmental features.

[0047] Specifically, unsupervised feature clustering can be a data processing operation that classifies samples into categories without introducing labeled data. Daily movement lag data can be the time difference data between when a user receives an interaction signal and generates a physical response. Environmental background white noise can be a sound signal with a uniform power spectral density distribution within a space. Initial response features can be data variables reflecting user movement lag. Background environmental features can be extracted data variables reflecting the spatial sound distribution.

[0048] Optionally, the management device can perform unsupervised feature clustering on the basic sensing data, and apply a clustering model to construct a feature space for the data samples. Within the feature space, the management device can calculate the distance measure for the corresponding dimension of the basic sensing data, and divide the basic sensing data into clusters based on the distance measure, separating daily motion lag data from environmental background white noise. The management device can construct a feature matrix based on the separated daily motion lag data and output initial response features. The management device can extract state parameters based on the separated environmental background white noise and output background environmental features.

[0049] S12. Perform frequency domain analysis and distribution probability extraction on the initial response characteristics and background environmental characteristics, calculate the mean and variance of the background sound pressure level, and generate the ambient acoustic background noise spectrum.

[0050] Specifically, probability distribution extraction can be an operation that calculates the frequency of data occurrences within a numerical range. Ambient background sound pressure level can be a physical quantity parameter reflecting the intensity of ambient sound.

[0051] Optionally, the management device can perform frequency domain analysis on the background environmental characteristics, applying a Fourier transform algorithm to convert the time-domain signal into a frequency-domain signal. The management device can perform filtering operations on the signal data within the frequency domain to extract the ambient background sound pressure level. The management device can perform probability distribution extraction processing on the ambient background sound pressure level, statistically analyzing the frequency of occurrence of the ambient background sound pressure level within a given interval. The management device can perform statistical derivation based on the probability distribution results to calculate the mean of the ambient background sound pressure level. The management device can perform a sum-of-squares operation on the mean and ambient background sound pressure level to calculate the variance of the ambient background sound pressure level. The management device can combine the extracted mean and variance variables to generate the ambient acoustic background noise spectrum.

[0052] S13. Obtain medication prescription instructions, extract the theoretical medication time and drug type for each day of the medication prescription instructions, and generate the initial prescription timeline.

[0053] Specifically, the initial prescription timeline can be a data structure that arranges medication times and drug classification identifiers in chronological order.

[0054] Optionally, the management device can separate the field content within the medication prescription instruction and extract the daily theoretical dosing time and drug type label from the parsed field content. The management device can perform data alignment operations on the extracted theoretical dosing time and drug type to establish a correspondence between time nodes and drug labels. The management device can use timestamps as a benchmark to perform ascending order calculations on the established correspondence of theoretical dosing times and drug types, construct a data structure with node information, and generate an initial prescription timeline.

[0055] S14. Based on the initial prescription timeline, perform spatiotemporal mapping calibration on the initial reaction characteristics, fit the individual-differentiated lag time from hearing the instruction to actually touching the smart pillbox, and generate an individualized reaction delay baseline.

[0056] Specifically, spatiotemporal mapping calibration processing can be a computational operation that aligns motion feature data with temporal and spatial coordinates and eliminates errors. Individual-specific hysteresis time can be the numerical value of the time span between the subject receiving instruction information and triggering a physical action.

[0057] Optionally, the management device can use the time parameters within the initial prescription timeline as a benchmark to calculate the deviation between the action recording time in the initial reaction characteristics and the nodes of the initial prescription timeline. The management device can extract the timestamps triggered by the wearer hearing the instruction and the timestamps triggered by the actual contact with the smart pillbox, and perform a subtraction operation on the timestamps to calculate a time difference sample. The management device can apply a regression model to perform a fitting operation on the time difference sample to calculate and output the wearer's individualized lag time. The management device can extract the mean of the individualized lag time data to generate an individualized reaction delay baseline.

[0058] S15. The initial prescription timeline is redundantly expanded by combining the individualized response delay baseline. Based on the theoretical medication time, a safe medication interval that allows for individual hysteresis fluctuations is defined, and a medication time window is generated.

[0059] Specifically, redundancy expansion processing can be a computational operation that adds a fault tolerance span to the data range. Individual hysteresis fluctuations can be the numerical variation in user action reaction time that deviates from the baseline parameter.

[0060] Optionally, the management device can perform redundancy expansion processing on the initial prescription timeline based on individualized response delay baseline data. The management device can calculate the data variance of individual hysteresis fluctuations based on the magnitude of the individualized response delay baseline. The management device can convert the data variance into a time offset, add the time offset to the value above the theoretical medication time, and subtract the time offset from the value below the theoretical medication time to calculate the upper and lower limits of the safe medication interval. The management device can merge the upper and lower limits of the safe medication interval to generate a medication time window.

[0061] This embodiment provides an intelligent reminder method for assisting elderly people in medication and health management. By performing unsupervised feature clustering and frequency domain analysis on basic sensor data, it is possible to separate and extract environmental background noise and user action features, thereby improving the device's noise immunity to environmental interference. Redundant expansion of the initial prescription timeline based on an individualized response delay baseline and the delineation of a safe medication interval can accommodate the slowed movements caused by declining physical function in the elderly. This reduces the frequency of false alarms and improves the adaptive fault tolerance and interactive experience of the medication reminder mechanism.

[0062] In one embodiment, S2 may include:

[0063] S21. Based on the time comparison mechanism, when entering the medication time window, the multi-dimensional sensor data is subjected to feature dimensionality reduction processing, redundant interference terms are removed and the illumination intensity, current sound pressure level, motion variance and spatial distance parameters are extracted to generate a real-time environmental feature matrix.

[0064] Specifically, the time comparison mechanism can be a logical rule that checks the current system time parameters against the boundaries of a preset time interval. Redundant interference items can be useless feature variables in the dataset that are irrelevant to the judgment logic. The current sound pressure level can be a numerical value reflecting the real-time sound intensity in the environment. The spatial distance parameter can be the spatial straight-line length of the target object relative to the sensing device. The real-time environmental feature matrix can be a matrix structure that combines the extracted environmental and spatial data.

[0065] Optionally, the management device can perform comparative calculations on system time data and medication time windows based on a time comparison mechanism. The management device can perform feature dimensionality reduction on multidimensional sensor data when the comparison results indicate that the medication time window has been entered, and calculate the data feature contribution parameter. The management device can remove redundant interference terms contained within the multidimensional sensor data based on the feature contribution parameter. The management device can extract illumination intensity, current sound pressure level, motion variance, and spatial distance parameters from the dimensionality-reduced dataset. The management device can combine illumination intensity, current sound pressure level, motion variance, and spatial distance parameters to establish a data structure and generate a real-time environmental feature matrix.

[0066] S22. Based on the ambient acoustic background noise spectrum, the current sound pressure level of the real-time environmental feature matrix and the ambient acoustic background noise spectrum are substituted into the Sigmoid activation function and the quadratic term penalty function for calculation to generate the auditory modality reach probability and the disturbance cost.

[0067] For example, the formula for calculating the reach probability of an auditory modality can be:

[0068]

[0069] in, For auditory modality reach probability, Indicates hearing; Indicates the acoustic activation slope; This is the current sound pressure level; The ambient acoustic background noise spectrum; Indicates the hearing compensation factor; This represents the environmental masking effect constant.

[0070] For example, the formula for calculating the cost of annoyance can be:

[0071]

[0072] in, For the cost of annoyance, Indicates hearing; This represents the disturbance cost coefficient.

[0073] Specifically, the auditory modal reach probability can be an estimate of the probability that a sound signal is perceived and received by the receiver. The acoustic activation slope can be a parameter controlling the steepness of the activation function curve. The hearing compensation factor can be a numerical term correcting for individual hearing loss errors. The environmental masking effect constant can be a constant characterizing the degree to which background sound masks the target sound. The interference cost coefficient can be a multiplier factor adjusting the proportion of interference calculation results.

[0074] Optionally, the management device can read preset acoustic activation slopes, hearing compensation factors, and environmental masking effect constants. The management device can substitute the current sound pressure level and the ambient acoustic background noise spectrum into a Sigmoid activation function model to calculate the auditory modality reach probability. The management device can also read preset disturbance cost coefficients. The management device can substitute the current sound pressure level and the ambient acoustic background noise spectrum into a quadratic penalty function model to calculate the squared difference between the two and multiply it by the disturbance cost coefficient to generate the disturbance cost.

[0075] S23. Perform visual attenuation mapping on the illumination intensity and spatial distance parameters in the real-time environment feature matrix, evaluate the effective line-of-sight transmittance of light under the spatial distance parameters, and generate the visual modal reach probability.

[0076] Specifically, visual attenuation mapping processing can be an algorithmic operation that calculates the degree of optical signal intensity reduction based on physical distance. Effective line-of-sight transmittance can be the percentage of light that remains at the target location after spatial attenuation. Visual modal reach probability can be an estimate of the probability that the light signal is captured by the target's perception.

[0077] Optionally, the management device can extract illuminance and spatial distance parameters from the real-time environmental feature matrix. The management device can input these parameters as independent variables into the physical attenuation model and perform visual attenuation mapping on them. Based on an exponential attenuation algorithm, the management device can calculate the energy loss of the light signal transmitted through the spatial medium. Based on the calculated energy loss and spatial distance parameters, the management device can evaluate the effective line-of-sight transmittance of light under the given spatial distance parameters. Based on the effective line-of-sight transmittance combined with ambient background illuminance data, the management device can perform a normalized mapping operation to calculate and output the visual modality reach probability.

[0078] S24. Obtain the driving vibration intensity of the device, substitute the driving vibration intensity and the motion variance in the real-time environmental feature matrix into the Sigmoid activation function for calculation, and generate the tactile modal reach probability.

[0079] For example, the formula for calculating the probability of tactile modality reach can be:

[0080]

[0081] in, For tactile modal reach probability, Indicates the sense of touch; Indicates the tactile activation slope; To drive vibration intensity; For the variance of motion; This represents the constant of the surface nerve masking effect.

[0082] Specifically, the driving vibration intensity can be a data parameter controlling the amount of mechanical energy output by the hardware vibration motor. The tactile modal reach probability can be the probability value that the vibration mechanical signal is perceived by the target skin nerves. The tactile activation slope can be a variable parameter that adjusts the steepness of the tactile function response curve. The surface nerve masking effect constant can be a fixed constant characterizing the interference effect of motion on vibration perception.

[0083] Optionally, the management device can substitute the driving vibration intensity and the motion variance in the real-time environmental feature matrix into the Sigmoid activation function for calculation. The difference between the driving vibration intensity and the motion variance is calculated, and the surface neural masking effect constant is introduced to perform an exponential operation on the activation function, obtaining the output probability value mapped to the zero-to-one interval. The management device can then generate the tactile modality reach probability through the above calculation.

[0084] S25. Construct a global utility objective function by integrating the reach probability of the auditory modality, the cost of disturbance, the reach probability of the tactile modality, and the reach probability of the visual modality. Calculate the expected return value corresponding to each candidate modality combination and take the candidate modality combination with the largest expected return value as the optimal policy combination vector.

[0085] For example, the expression for the global utility objective function can be:

[0086]

[0087] in, Let the expected value of the objective function be... Candidate modality combinations are any combination of auditory, tactile, and visual modalities. It is a single sensory modality; For modality The probability of reaching; This represents the weighting parameters for medical drugs; For the cost of annoyance; This comes at the cost of energy consumption; This is the annoyance penalty coefficient; This is the wear and tear penalty coefficient.

[0088] Specifically, the global utility objective function can be a mathematical equation that comprehensively evaluates the reach probability and various cost weights. The candidate modality combination can be a sequence of different sensory alert methods. The expected return value can be the comprehensive evaluation score calculated by the objective function. The optimal strategy combination vector can be a strategy parameter matrix that maximizes the expected return value. The medical drug weight parameter can be a numerical variable reflecting the importance level of the drug. The energy consumption cost can be a measure of the power loss caused by the output signal. The annoyance penalty coefficient can be a constant used to adjust the annoyance cost weight. The energy consumption penalty coefficient can be a constant used to adjust the energy consumption cost weight.

[0089] Optionally, the management device can construct a global utility objective function by comprehensively considering the auditory modality reach probability, annoyance cost, tactile modality reach probability, and visual modality reach probability. The management device can establish a set of candidate modality combinations composed of different modality arrangements. The management device can substitute various probability values ​​and cost parameters into the global utility objective function to calculate the expected return value corresponding to each candidate modality combination. The management device can perform numerical sorting and maximum value search comparison algorithms on the calculated expected return value set to extract the modality combination parameter matrix corresponding to the maximum expected return value. The management device can assign and output the candidate modality combination with the largest expected return value as the optimal policy combination vector.

[0090] This embodiment provides an intelligent reminder method for assisting elderly people in medication and health management. By extracting multi-dimensional sensor data based on a time comparison mechanism and generating a real-time environmental feature matrix, fine-grained quantification of environmental conditions can be achieved. Independent activation and penalty functions are constructed based on sound, light, and touch multi-dimensional sensory channels to calculate reach probability and annoyance cost, allowing for the evaluation of the intervention efficacy of each modality in specific environments. A global utility objective function is constructed based on these probability and cost data, and its maximum value is solved, enabling adaptive adjustment and dynamic switching of the multi-modal reminder strategy. This can increase the probability of successful signal reception by users in complex home environments, reduce psychological annoyance to the elderly and surrounding family members, and improve the environmental inclusiveness and interactive experience of the overall medication management system.

[0091] In one embodiment, S3 may include:

[0092] S31. Perform multi-dimensional sequence fusion on the user action signal, align the kinematic signal collected by the wearable device with the opening and closing level of the smart medicine box sensor in the reminder device group according to the timestamp, and generate a multi-source observation sequence.

[0093] Specifically, user action signals can be a dataset recording user interaction behavior characteristics. Wearable devices can be hardware attached to the user's body surface to collect physiological and motion parameters. Kinematic signals can be a data array containing spatial acceleration and angular velocity parameters. Opening / closing levels can be high and low voltage signals characterizing the physical state of the smart medicine box's door opening and closing. Multi-source observation sequences can be composite data matrices that fuse data features from multiple sensors on the same time axis.

[0094] Optionally, the management device can read the kinematic signals transmitted back by the wearable device via a network interface. The management device can also read the opening / closing level data uploaded by the smart pillbox sensor in the reminder device group via a communication channel. The management device can extract the timestamp parameters of the kinematic signals and the opening / closing level data. Using the timestamp as a benchmark, the management device can splice and reassemble kinematic signals and opening / closing level data at the same time point to establish a matching correspondence between the kinematic signals and opening / closing level data at the same timestamp, outputting a multi-source observation sequence with a unified timeline and multi-dimensional sensor characteristics.

[0095] S32. Based on the Hidden Markov Model, the state probability of the multi-source observation sequence is fitted. By calculating the emission probability of the hidden state corresponding to the physical observation data at different times, the state emission probability matrix is ​​generated.

[0096] Specifically, state probability fitting can be a computational operation that maps actual observation data to a potential state interval based on mathematical model parameters. Physical observation data can be a set of sensor values ​​collected at corresponding time points. Latent states can be internal variables within the model that are not directly observable but affect the output. Emission probability can be the conditional probability value for generating specific physical observation data under a specific latent state. The state emission probability matrix can be a two-dimensional matrix structure containing emission probability data corresponding to all time and latent state nodes.

[0097] Optionally, the management device can extract physical observation data from the data nodes corresponding to each time point in the multi-source observation sequence. The management device can then substitute the physical observation data into a Hidden Markov Model (HMM) and, based on the HMM's probability density function, calculate the emission probability of the physical observation data in each hidden state at different times. The management device can iterate through all time points and all hidden states, recording the calculated emission probabilities. Finally, the management device can arrange the extracted emission probabilities according to time and state dimensions, establishing a two-dimensional numerical array to generate a state emission probability matrix.

[0098] S33. Perform time-series decoding on the state emission probability matrix, and recursively search for the optimal path with the maximum global posterior probability under state transition constraints to generate a physical time-series state sequence.

[0099] Specifically, temporal decoding can be a computational operation that derives possible hidden state sequences based on observed data and probabilistic models. The Viterbi algorithm is a decoding algorithm based on dynamic programming to solve for state sequences in a hidden Markov model. The hidden state transition path can be a sequence of nodes that connect independent hidden states in chronological order.

[0100] Optionally, the management device can perform temporal decoding on the state emission probability matrix, apply the Viterbi algorithm to perform dynamic programming calculations, and recursively calculate the local maximum posterior probability for each hidden state at each time point by combining the initial state distribution parameters and the state transition probability matrix parameters. At the end of the recursive calculation, the management device can search for an extreme value among the local maximum posterior probabilities at the final time. The management device can obtain the state node corresponding to this extreme value, perform a reverse backtracking operation, and extract the hidden state node corresponding to each time point. The management device can connect the extracted nodes to obtain the hidden state transition path corresponding to the maximum posterior probability. The management device can perform label conversion based on the obtained hidden state transition path to generate a physical temporal state sequence.

[0101] S34. Based on the preset complete medication retrieval and swallowing time sequence, compare the temporal coverage completeness of the physical temporal state sequence, calculate the node coverage ratio of the physical temporal state sequence to the complete medication retrieval and swallowing time sequence, and generate a state completeness score.

[0102] Specifically, the complete drug retrieval and swallowing sequence can be used to represent the standard physical state transition process from the start of drug retrieval to the final completion of drug swallowing.

[0103] Specifically, temporal coverage completeness comparison can be an operation that verifies the degree to which the actual state sequence contains nodes of the standard state sequence. Node coverage ratio can be the proportion of the number of actual hit standard action nodes to the total number of nodes. State completeness score can be a quantitative score parameter reflecting the degree of execution of the action chain.

[0104] Optionally, the management device can parse the standard physical state transition process nodes from the start of medication retrieval to the final completion of medication swallowing contained in the complete medication retrieval and swallowing sequence. The management device can extract the actual state node labels from the physical time-series state sequence. Based on the preset complete medication retrieval and swallowing sequence, the management device can perform a time-series coverage completeness comparison operation on the physical time-series state sequence, checking and counting the total number of nodes that the physical time-series state sequence hits in the complete medication retrieval and swallowing sequence. The management device can calculate the node coverage ratio of the physical time-series state sequence to the complete medication retrieval and swallowing sequence by dividing the total number of hit nodes by the total number of nodes set in the complete medication retrieval and swallowing sequence. Based on the node coverage ratio, the management device can perform data format conversion operations to generate a state completeness score.

[0105] S35. Perform threshold classification and judgment on the status integrity score. When the status integrity score reaches the preset integrity threshold, the closed loop is determined to be complete and a medication confirmation label is generated. When the status integrity score is lower than the preset status qualification line, the action chain is determined to be broken and an action abnormal interruption label is generated.

[0106] Specifically, threshold classification and judgment can be a logical operation that divides data samples into intervals based on boundary values. The integrity threshold can be a numerical parameter that sets a standard for an action sequence to meet. The state qualification line can be a critical parameter for distinguishing between normal and abnormal actions.

[0107] Optionally, the management device can compare the status integrity score with a preset integrity threshold. If the status integrity score reaches the preset integrity threshold, the management device can trigger a first judgment branch, determining that the closed loop is complete. Based on the logical result of the closed loop being complete, the management device can perform a tag assignment operation to generate a medication-taken status confirmation tag. Alternatively, the management device can compare the status integrity score with a preset status pass line. If the status integrity score is lower than the preset status pass line, the management device can trigger a second judgment branch, determining that the action chain is broken. Based on the logical result of the broken action chain, the management device can perform a tag assignment operation to generate an action abnormality interruption tag.

[0108] This embodiment provides an intelligent reminder method for assisting elderly people in medication and health management. By fusing multi-dimensional sequences of the kinematic signals from wearable devices and the opening and closing levels of the smart pillbox, a multi-source observation sequence recording medication-taking actions can be constructed. Based on a Hidden Markov Model, state probability fitting and temporal decoding can identify the physiological temporal state sequence of medication. Matching the coverage and dwell time of the state sequence based on temporal integrity comparison rules enables logical judgment of the actual medication-taking behavior chain. This solves the behavioral deception problem caused by traditional devices relying solely on the pillbox's open state for judgment, improves the system's capture rate of abnormal situations such as "taking out medication but not swallowing" in patients with mild cognitive impairment, and enhances the reliability and data accuracy of the medication safety monitoring mechanism.

[0109] In one embodiment, S4 may include:

[0110] S41. If the action status assessment result is an action abnormality interruption label, a user reminder signal is generated and sent to the reminder device group. Acoustic feature extraction and speech recognition processing are performed on the feedback user voice response signal to filter out background noise and obtain a filtered speech signal. The filtered speech signal is converted into a speech text sequence and input into a natural language coding model for word embedding. The speech text sequence is segmented into multiple basic words, each basic word is converted into a high-dimensional dense vector, and a word embedding feature sequence is generated by concatenating multiple high-dimensional dense vectors. The word embedding feature sequence is encoded and mapped to generate a user semantic hidden layer representation.

[0111] Specifically, a natural language coding model can be a mathematical model that processes text data and extracts deep features. Basic lexical units can be independent lexical units that constitute a text sequence. High-dimensional dense vectors can be matrix arrays with multiple numerical dimensions and non-sparse features. Word embedding feature sequences can be data matrices composed of combined lexical vectors. User semantic hidden layer representations can be deep feature parameters that reflect the core meaning of speech.

[0112] Optionally, the management device can generate a user reminder signal and send it to the reminder device group when the action status assessment result is an action abnormality interruption label. The management device can receive the user's voice response signal returned by the reminder device group and perform acoustic feature extraction and speech recognition processing on the user's voice response signal. The management device can perform a filtering algorithm on the user's voice response signal to remove background noise and obtain a filtered voice signal. The management device can convert the filtered voice signal into a speech-text sequence and input the speech-text sequence into a natural language coding model to segment the speech-text sequence into multiple basic units. The management device can apply a vocabulary lookup table to convert each basic unit into a high-dimensional dense vector. The management device can perform a concatenation operation on the high-dimensional dense vector along the feature dimension to generate a word embedding feature sequence concatenated from multiple high-dimensional dense vectors. The management device can encode and map the word embedding feature sequence and generate a user semantic hidden layer representation through matrix calculation.

[0113] S42. Based on the user's semantic hidden layer representation and the individualized response delay baseline, perform cross-attention verification and delay penalty calculation, calculate the matching degree of the current response context semantics, and apply a decay penalty to the response time that exceeds the baseline to generate a cognitive verification comprehensive score.

[0114] For example, the formula for calculating the cognitive verification composite score can be:

[0115]

[0116] in, The overall score for cognitive verification; For semantic matching functions; For user semantic hidden layer representation; Label the target with a semantic hidden layer; This is the actual reaction delay; As a baseline for individualized response delay; This refers to the delay decay rate; This is the penalty coefficient; Extract the penalty factor for the "confused" category.

[0117] Specifically, cross-attention verification can be a mathematical operation comparing the correlation between two feature sequences. Delay penalty calculation can be a mathematical operation deducting points from a timeout variable. The semantic matching function can be a mathematical equation calculating feature distance. The target annotation semantic hidden layer can be feature parameters of a pre-defined answer. The actual response delay can be the time difference recorded in this interaction. The delay decay rate can be a numerical parameter controlling the deduction weights.

[0118] Optionally, the management device can perform cross-attention verification and delay penalty calculation based on the user's semantic hidden layer representation and the individualized response delay baseline. The management device can substitute the user's semantic hidden layer representation and the target labeled semantic hidden layer into a semantic matching function to calculate the matching degree of the current response context semantics. The management device can obtain the actual response delay of this interaction, perform an algebraic comparison between the actual response delay and the individualized response delay baseline, and obtain the excessive response time. The management device can substitute the excessive response time, delay decay rate, and penalty coefficient into an exponential function to perform decay penalty calculation on the response time. The management device can match the user's text with a preset dictionary, extract the corresponding confusion category tags, and extract penalty factors. The management device can combine the matching degree, decay penalty result, and confusion category tags to extract penalty factors, perform algebraic addition and subtraction operations to solve, and generate a comprehensive cognitive verification score.

[0119] S43. When the cognitive verification comprehensive score is lower than the set cognitive verification threshold, obtain the spatial positioning distance of the reminder device group, calculate the RGB brightness value required to meet the glare-free intuitive guidance according to the inverse relationship of the spatial positioning distance, and generate the target luminous intensity.

[0120] Specifically, glare-free intuitive guidance can be a state of position guidance through optical control.

[0121] Optionally, the management device can perform a numerical comparison calculation between the cognitive verification comprehensive score and the cognitive verification threshold. Under the logical condition that the cognitive verification comprehensive score is lower than the set cognitive verification threshold, the positioning module is activated. The management device can obtain the spatial positioning distance of the alerting device group through a radio frequency ranging protocol. The management device can establish an optical brightness mapping model. Based on the inverse relationship of spatial positioning distance, the management device can input the spatial positioning distance into the optical brightness mapping model to perform nonlinear calculations and calculate the RGB brightness values ​​required to meet the glare-free intuitive guidance. The management device can encapsulate the RGB brightness values ​​into a data format to generate the target luminous intensity.

[0122] S44. Generate a phototaxis guidance instruction based on the target luminescence intensity and send the phototaxis guidance instruction to the reminder device group. After receiving the medication retrieval signal from the reminder device group, update the action abnormality interruption label to a medication taken status confirmation label.

[0123] Optionally, the management device can generate phototaxis guidance instructions based on the target luminescence intensity. The management device can call the communication port to send the phototaxis guidance instructions to the reminder device group. The management device can start an event listening thread to wait for data feedback from the hardware nodes and receive the medication retrieval signal from the reminder device group. Upon receiving the medication retrieval signal, the management device can trigger the state machine's transition operation, query the current identifier data in the memory database, execute an overwrite command on the underlying memory, and update the action exception interruption tag to a medication-taken status confirmation tag.

[0124] Specifically, the phototropic guidance command is used to instruct and remind the device group to perform the following steps:

[0125] S441. Based on the light-guided instruction, drive the light-emitting components of the reminder device group to emit light according to the target light intensity.

[0126] Optionally, the alerting device group can receive phototropic guidance command data packets via a communication interface. The alerting device group can extract the target luminous intensity parameters encapsulated within the phototropic guidance command, input these parameters into its internal microprocessor unit, and generate a pulse-width modulation (PWM) signal matching the target luminous intensity. The alerting device group can then transmit the PWM signal to its internal power management module, driving the light-emitting component based on the PWM signal. The alerting device group can adjust the voltage and current parameters across the power supply pins of the light-emitting component to control its luminous operation according to the target luminous intensity. The alerting device group can maintain the optical energy output state of the light-emitting component to form a spatial optical guidance path.

[0127] S442. After the built-in sensor of the reminder device group detects the actual drug dispensing action, a drug dispensing signal is generated.

[0128] Optionally, the alerting device group can continuously detect the physical spatial parameters around the medicine box compartment using built-in sensors, and output an electrical signal reflecting the physical spatial parameters to the microprocessor. The alerting device group can perform analog-to-digital conversion and benchmark threshold comparison operations on the amplitude of the electrical signal. If the comparison result shows a level transition in the physical spatial parameters, it is determined that an actual medicine retrieval action has been captured. The alerting device group can obtain the system timestamp parameter corresponding to the captured actual medicine retrieval action. The alerting device group can read the internally stored device physical address code and action trigger event identifier code, and perform byte alignment and concatenation processing on the device physical address code, action trigger event identifier code, and system timestamp parameter to generate a medicine retrieval signal.

[0129] This embodiment provides an intelligent reminder method for assisting elderly people in medication and health management. By extracting features and verifying cross-attention from the user's voice response signal, combined with calculation of reaction time decay penalty, a non-intrusive implicit assessment of the elderly person's true cognitive state can be achieved. When the overall cognitive verification score is low, a light-oriented guidance command is dynamically generated based on the spatial positioning distance, driving the device to emit light. This allows the user to complete the medication retrieval action loop through basic human visual instincts, without the burden of logical thinking. This can solve the stagnation problem that is common in people with mild cognitive impairment, and can improve the system's intervention success rate and overall medication adherence when users experience cognitive gaps.

[0130] In one embodiment, S5 may include:

[0131] S51. Obtain the local vector clock, and accumulate the values ​​of the local node event components in the local vector clock according to the medication status confirmation tag to generate the updated local vector clock.

[0132] Specifically, a local vector clock can be an array of data structures that records the order in which events occur on each node in a distributed system. Local node event components can be numerical elements in the local vector clock array that represent the event counts for the corresponding device.

[0133] Optionally, the management device can trigger the auto-increment logic mechanism of the local status counter upon reading a medication-taken status confirmation tag. The management device can parse the local vector clock data structure array and locate the specific memory offset address of the local node event component within the array. The management device can extract the historical event count value corresponding to this specific memory offset address and perform a fixed-step increment operation on the extracted historical event count value of the local node event component. The management device can then write the output of the increment operation back to the memory space corresponding to the local vector clock data structure array, integrate the modified data structure array elements, and generate an updated local vector clock.

[0134] S52. Perform weak network causal synchronization processing based on the medication status confirmation tag, establish a causal mapping relationship between the medication status confirmation tag and the updated local vector clock, and encapsulate the updated local vector clock and medication status confirmation tag with the established causal mapping relationship to generate a synchronization data packet.

[0135] Specifically, weak network causal synchronization processing can be an operational mechanism to maintain the logical consistency of data from distributed devices under unstable network connectivity conditions. Causal mapping association can be a corresponding matching relationship established between state data and time-series parameters.

[0136] Optionally, the management device can extract medication status confirmation tag data from a specified address range in system memory, and thereby initiate a weak network causal synchronization mechanism for distributed nodes on the local area network. The management device can read the updated local vector clock array from the cache queue, establish a causal mapping association between tags and clocks through memory pointer allocation, and bind their addresses using reference operations to form a key-value pair structure. The management device can then pass the associated clock and tag data to the underlying communication protocol stack for data encapsulation. Finally, the management device can perform byte serialization according to the application layer payload format, align and concatenate the output binary stream, and generate synchronization data packets for network distribution.

[0137] S53. Construct a network communication protocol for the synchronization data packet, configure the target broadcast address and instruction message header, generate a data instruction including the synchronization data packet, and broadcast the data instruction to other node devices in the local area network.

[0138] Specifically, the command header can be a control information field that includes the protocol version, data length, and checksum. The data command can be a complete communication message that encapsulates the control information and the data payload.

[0139] Optionally, the management device can retrieve synchronization data packets from the transmission register queue of the communication controller, read the system network configuration file, and obtain the target broadcast address of the local area network. The management device can calculate the length of the synchronization data packets and generate a cyclic redundancy check (CRC) code using a hash algorithm, combining the protocol version number, length, and check code to generate a command message header. The management device can configure the broadcast address and message header into the message control field, concatenate the command message header and the synchronization data packets, and generate a complete data command message. The management device can then activate the underlying radio frequency communication hardware interface to broadcast data commands to other node devices within the local area network, transmitting radio frequency signals through the radio frequency antenna at a set transmission power to complete cross-node data push.

[0140] This embodiment provides an intelligent reminder method for assisting elderly people in medication and health management. By accumulating the values ​​of local node event components within the local vector clock and establishing a causal mapping relationship between medication status confirmation tags and the vector clock, logical sequencing of medication events among distributed nodes can be achieved. A network communication protocol is constructed based on a weak network causal synchronization processing mechanism, and data instructions including synchronization data packets are broadcast to other node devices within the local area network, improving the reliability of data synchronization in environments with signal obstruction or network fluctuations.

[0141] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0142] Based on the same inventive concept, this application also provides a system for implementing the aforementioned intelligent reminder method for assisting elderly people in medication and health management. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of an intelligent reminder system for assisting elderly people in medication and health management provided below can be found in the limitations of the intelligent reminder method for assisting elderly people in medication and health management described above, and will not be repeated here.

[0143] In one exemplary embodiment, such as Figure 3 As shown, an intelligent reminder system 200 is provided to assist elderly people in medication and health management, and is used to implement the methods in the above-described method embodiments. The device may include:

[0144] The baseline feature extraction module 201 is used to perform unsupervised learning on the acquired basic perception data to generate initial reaction features and background environment features. Based on the acquired medication prescription instructions, the initial reaction features and background environment features are time-space mapped to extract the information and environmental parameters corresponding to the time window and the target drug warehouse, and individualized reaction delay baseline, ambient acoustic background noise spectrum and medication time window are generated.

[0145] The multimodal scheduling module 202 is used to acquire multi-dimensional sensor data of the current environment, and to perform adaptive multimodal reminder scheduling based on the ambient acoustic background noise spectrum, medication time window and multi-dimensional sensor data. It calculates the reach probability and annoyance cost of auditory mode, tactile mode and visual mode and constructs a utility objective function for solution to generate the optimal strategy combination vector. The optimal strategy combination vector is used to control the reminder device group to remind the user respectively.

[0146] The action tracking and analysis module 203 is used to acquire user action signals after being reminded, perform state probability fitting and temporal decoding on the user action signals through a hidden Markov model to generate a physical temporal state sequence, and perform anomaly analysis on the physical temporal state sequence based on the temporal integrity comparison rule to generate action state analysis results.

[0147] The cognitive verification guidance module 204 is used to generate a user reminder signal and send it to the reminder device group if the action status judgment result is an action abnormal interruption label. Based on the action abnormal interruption label and the individualized response delay baseline, it performs cognitive verification on the feedback user voice response signal and generates a cognitive verification comprehensive score. When the cognitive verification comprehensive score is lower than the set cognitive verification threshold, it obtains the spatial positioning distance of the reminder device group, generates a light-touch guidance instruction based on the spatial positioning distance and sends it to the reminder device group. After receiving the medication retrieval signal from the reminder device group, it updates the action abnormal interruption label to a medication taken status confirmation label.

[0148] The weak network status synchronization module 205 is used to perform causal synchronization of weak network status based on the medication status confirmation tag, encapsulate data in combination with the acquired local vector clock, generate data instructions including synchronization data packets, and broadcast the data instructions to other node devices in the local area network.

[0149] In one embodiment, the baseline feature extraction module 201 can also be used for:

[0150] Unsupervised feature clustering is performed on the acquired basic perception data to separate daily action lag data and environmental background white noise, generating initial response features and background environmental features.

[0151] Frequency domain analysis and probability distribution extraction are performed on the initial response characteristics and background environmental characteristics to calculate the mean and variance of the background sound pressure level and generate the ambient acoustic background noise spectrum.

[0152] Obtain medication prescription instructions, extract the theoretical medication time and drug type for each day of the medication prescription instructions, and generate an initial prescription timeline.

[0153] Based on the initial prescription timeline, the initial reaction characteristics are calibrated by spatiotemporal mapping to fit the individual-differentiated lag time from hearing the instruction to actually touching the smart pillbox, and generate an individualized reaction delay baseline.

[0154] The initial prescription timeline is redundantly expanded by combining an individualized response delay baseline. Based on the theoretical medication time, a safe medication interval that allows for individual hysteresis fluctuations is defined, thus generating a medication time window.

[0155] In one embodiment, the multimodal scheduling module 202 can also be used for:

[0156] Based on the time comparison mechanism, when entering the medication time window, the multi-dimensional sensor data is subjected to feature dimensionality reduction processing, redundant interference terms are removed and the parameters of light intensity, current sound pressure level, motion variance and spatial distance are extracted to generate a real-time environmental feature matrix.

[0157] Based on the ambient acoustic background noise spectrum, the current sound pressure level of the real-time environmental feature matrix and the ambient acoustic background noise spectrum are substituted into the Sigmoid activation function and the quadratic penalty function for calculation to generate the auditory modality reach probability and the disturbance cost.

[0158] Visual attenuation mapping is performed on the illumination and spatial distance parameters in the real-time environment feature matrix to evaluate the effective line-of-sight transmittance of light under the spatial distance parameters and generate the visual modal reach probability.

[0159] The driving vibration intensity of the device is obtained, and the driving vibration intensity and the motion variance in the real-time environmental feature matrix are substituted into the Sigmoid activation function for calculation to generate the tactile modal reach probability.

[0160] A global utility objective function is constructed by integrating the reach probability of the auditory modality, the cost of disturbance, the reach probability of the tactile modality, and the reach probability of the visual modality. The expected revenue value corresponding to each candidate modality combination is calculated, and the candidate modality combination with the largest expected revenue value is taken as the optimal policy combination vector.

[0161] In one embodiment, the motion tracking and analysis module 203 can also be used for:

[0162] Multidimensional sequence fusion of user action signals is performed, and the kinematic signals collected by wearable devices are aligned with the opening and closing levels of the smart pillbox sensor in the reminder device group according to the timestamp to generate a multi-source observation sequence.

[0163] The state probability is fitted to the multi-source observation sequence based on the hidden Markov model. By calculating the emission probability of the hidden state corresponding to the physical observation data at different times, the state emission probability matrix is ​​generated.

[0164] The state emission probability matrix is ​​temporally decoded, and the optimal path with the maximum global posterior probability is recursively searched under state transition constraints to generate a physical temporal state sequence.

[0165] Based on the preset complete medication retrieval and swallowing time sequence, the temporal coverage integrity of the physical temporal state sequence is compared, the node coverage ratio of the physical temporal state sequence to the complete medication retrieval and swallowing time sequence is calculated, and a state integrity score is generated.

[0166] Threshold classification and judgment are performed on the status integrity score. When the status integrity score reaches the preset integrity threshold, the closed loop is determined to be complete and a medication-taken status confirmation label is generated. When the status integrity score is lower than the preset status qualification line, the action chain is determined to be broken and an action abnormal interruption label is generated.

[0167] In one embodiment, the cognitive verification guidance module 204 can also be used for:

[0168] If the action status assessment result is an action abnormality interruption label, a user reminder signal is generated and sent to the reminder device group. The user's voice response signal is then processed for acoustic feature extraction and speech recognition. Background noise in the user's voice response signal is filtered out to obtain a filtered voice signal. The filtered voice signal is converted into a voice-text sequence and input into a natural language coding model for word embedding. The voice-text sequence is segmented into multiple basic words, and each basic word is converted into a high-dimensional dense vector. A word embedding feature sequence is generated by concatenating multiple high-dimensional dense vectors. The word embedding feature sequence is then encoded and mapped to generate a user semantic hidden layer representation.

[0169] Based on the user's semantic hidden layer representation and individualized response delay baseline, cross-attention verification and delay penalty calculation are performed. The matching degree of the current response context semantics is calculated and the response time exceeding the baseline is attenuated and penalized to generate a cognitive verification comprehensive score.

[0170] When the cognitive verification score is lower than the set cognitive verification threshold, the spatial positioning distance of the reminder device group is obtained, and the RGB brightness value required to meet the glare-free intuitive guidance is calculated based on the inverse relationship of the spatial positioning distance, and the target luminous intensity is generated.

[0171] Based on the target luminous intensity, a phototaxis guidance instruction is generated and sent to the reminder device group. After receiving the medication retrieval signal from the reminder device group, the action abnormality interruption label is updated to a medication-taken status confirmation label.

[0172] In one embodiment, the weak network status synchronization module 205 can also be used for:

[0173] Obtain the local vector clock, and accumulate the values ​​of the local node event components within the local vector clock based on the medication status confirmation tag to generate an updated local vector clock.

[0174] Weak network causal synchronization is performed based on the medication status confirmation tag. A causal mapping relationship is established between the medication status confirmation tag and the updated local vector clock. The updated local vector clock and the medication status confirmation tag with the established causal mapping relationship are encapsulated to generate a synchronization data packet.

[0175] The system constructs a network communication protocol for the synchronization data packet, configures the target broadcast address and instruction header, generates a data instruction including the synchronization data packet, and broadcasts the data instruction to other node devices in the local area network.

[0176] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the intelligent reminder method for assisting elderly people in medication and health management as described above.

[0177] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0178] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0179] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A smart reminder method to assist elderly people in medication and health management, characterized in that, The method includes: S1. Perform unsupervised learning on the acquired basic perception data to generate initial reaction features and background environment features. Based on the acquired medication prescription instructions, perform time-space mapping on the initial reaction features and background environment features, extract the information and environmental parameters corresponding to the time window and the target drug warehouse, and generate an individualized reaction delay baseline, environmental acoustic background noise spectrum and medication time window. S2. Acquire multi-dimensional sensor data of the current environment, and perform adaptive multimodal reminder scheduling based on the ambient acoustic background noise spectrum, the medication time window and the multi-dimensional sensor data. Calculate the reach probability and annoyance cost of auditory mode, tactile mode and visual mode, and construct a utility objective function for solution to generate the optimal strategy combination vector. The optimal strategy combination vector is used to control the reminder device group to remind the user respectively. S3. Obtain the user's action signal after being reminded, perform state probability fitting and temporal decoding on the user's action signal using a hidden Markov model to generate a physical temporal state sequence, and perform anomaly judgment on the physical temporal state sequence based on the temporal integrity comparison rule to generate an action state judgment result; wherein, the action state judgment result is a medication-taken status confirmation tag or an action abnormal interruption tag. S4. If the action status assessment result is an action abnormality interruption tag, a user reminder signal is generated and sent to the reminder device group. Based on the action abnormality interruption tag and the individualized reaction delay baseline, the user voice response signal is cognitively verified to generate a cognitive verification comprehensive score. When the cognitive verification comprehensive score is lower than the set cognitive verification threshold, the spatial positioning distance of the reminder device group is obtained. Based on the spatial positioning distance, a light-guided instruction is generated and sent to the reminder device group. After receiving the medication retrieval signal from the reminder device group, the action abnormality interruption tag is updated to a medication-taken status confirmation tag. S5. Based on the medication status confirmation tag, perform weak network status causal synchronization, combine with the acquired local vector clock to encapsulate data, generate a data instruction including a synchronization data packet, and broadcast the data instruction to other node devices in the local area network; wherein, the data instruction is used to update the medication status and vector clock data of other node devices.

2. The method according to claim 1, characterized in that, S1 includes: S11. Perform unsupervised feature clustering on the acquired basic perception data to separate daily action lag data and environmental background white noise, and generate initial reaction features and background environmental features. S12. Perform frequency domain analysis and distribution probability extraction processing on the initial response characteristics and the background environment characteristics, calculate the mean and variance of the background sound pressure level, and generate the ambient acoustic background noise spectrum. S13. Obtain medication prescription instructions, extract the theoretical medication time and drug type for each day of the medication prescription instructions, and generate an initial prescription timeline; S14. Based on the initial prescription timeline, perform spatiotemporal mapping calibration on the initial reaction characteristics, fit the individual-differentiated lag time from hearing the instruction to actually touching the smart pillbox, and generate the individualized reaction delay baseline. S15. The initial prescription time axis is redundantly expanded by combining the individualized response delay baseline, and a safe medication interval that allows for individual hysteresis fluctuations is defined based on the theoretical medication time, thereby generating the medication time window.

3. The method according to claim 1, characterized in that, The S2 includes: S21. Based on the time comparison mechanism, when entering the medication time window, the multidimensional sensor data is subjected to feature dimensionality reduction processing, redundant interference terms are removed and the illumination intensity, current sound pressure level, motion variance and spatial distance parameters are extracted to generate a real-time environmental feature matrix. S22. Based on the ambient acoustic background noise spectrum, the current sound pressure level of the real-time environmental feature matrix and the ambient acoustic background noise spectrum are substituted into the Sigmoid activation function and the quadratic penalty function for calculation to generate the auditory modal reach probability and the annoyance cost; wherein, the formula for calculating the auditory modal reach probability is: in, The probability of reaching the auditory modality. Indicates hearing; Indicates the acoustic activation slope; The current sound pressure level; The ambient acoustic background noise spectrum; Indicates the hearing compensation factor; Represents the environmental masking effect constant; The formula for calculating the cost of the disturbance is as follows: in, For the cost of the aforementioned annoyance, Indicates hearing; The disturbance cost coefficient; S23. Perform visual attenuation mapping on the illumination intensity and spatial distance parameters in the real-time environment feature matrix, evaluate the effective line-of-sight transmittance of light under the spatial distance parameters, and generate the visual modal reach probability. S24. Obtain the driving vibration intensity of the device, and substitute the driving vibration intensity and the motion variance in the real-time environmental feature matrix into the Sigmoid activation function for calculation to generate the tactile modal reach probability; wherein, the calculation formula for the tactile modal reach probability is: in, The probability of reaching the tactile modality. Indicates the sense of touch; Indicates the tactile activation slope; The driving vibration intensity; Let Variance be the motion variance; This represents the constant of the surface nerve masking effect; S25. Construct a global utility objective function by combining the auditory modality reach probability, the annoyance cost, the tactile modality reach probability, and the visual modality reach probability; calculate the expected return value corresponding to each candidate modality combination; and select the candidate modality combination with the largest expected return value as the optimal strategy combination vector; wherein, the expression of the global utility objective function is: in, Let the expected value of the objective function be... The candidate modality combination is any combination of auditory, tactile, and visual modalities. It is a single sensory modality; For modality The probability of reaching; This represents the weighting parameters for medical drugs; For the cost of the aforementioned annoyance; This comes at the cost of energy consumption; This is the annoyance penalty coefficient; This is the wear and tear penalty coefficient.

4. The method according to claim 1, characterized in that, The S3 includes: S31. Perform multi-dimensional sequence fusion on the user action signal, align the kinematic signal collected by the wearable device with the opening and closing level of the smart medicine box sensor in the reminder device group according to the timestamp, and generate a multi-source observation sequence; S32. Based on the hidden Markov model, the state probability of the multi-source observation sequence is fitted, and the emission probability matrix is ​​generated by calculating the emission probability of the hidden state corresponding to the physical observation data at different times. S33. Perform time-series decoding on the state emission probability matrix, and recursively search for the optimal path with the global maximum posterior probability under state transition constraints to generate the physical time-series state sequence. S34. Based on the preset complete drug retrieval and swallowing sequence, the physical time sequence state sequence is compared for time sequence coverage integrity. The node coverage ratio of the physical time sequence state sequence to the complete drug retrieval and swallowing sequence is calculated, and a state integrity score is generated. The complete drug retrieval and swallowing sequence is used to represent the standard physical state transition process from the start of drug retrieval to the final completion of drug swallowing. S35. Perform threshold classification and judgment on the state integrity score. When the state integrity score reaches the preset integrity threshold, determine that the closed loop is complete and generate the medication taken status confirmation label. When the state integrity score is lower than the preset state qualification line, determine that the action chain is broken and generate the action abnormal interruption label.

5. The method according to claim 1, characterized in that, The S4 includes: S41. If the action state judgment result is an action abnormal interruption label, a user reminder signal is generated and sent to the reminder device group. Acoustic feature extraction and speech recognition processing are performed on the feedback user voice response signal to filter out background noise and obtain a filtered voice signal. The filtered voice signal is converted into a voice text sequence and input into a natural language coding model for word embedding. The voice text sequence is segmented into multiple basic words and each basic word is converted into a high-dimensional dense vector. A word embedding feature sequence is generated by concatenating multiple high-dimensional dense vectors. The word embedding feature sequence is encoded and mapped to generate a user semantic hidden layer representation. S42. Based on the user semantic hidden layer representation and the individualized response delay baseline, perform cross-attention verification and delay penalty calculation, calculate the matching degree of the current response context semantics, and apply a decay penalty to the response time exceeding the baseline to generate the cognitive verification comprehensive score; wherein, the calculation formula of the cognitive verification comprehensive score is: in, The overall score for the cognitive verification; For semantic matching functions; This is the user semantic hidden layer representation; Label the target with a semantic hidden layer; This is the actual reaction delay; This serves as the baseline for the individualized response delay; This refers to the delay decay rate; This is the penalty coefficient; Extracting penalty factors for the "confused" category; S43. When the cognitive verification comprehensive score is lower than the set cognitive verification threshold, obtain the spatial positioning distance of the reminder device group, calculate the RGB brightness value required to meet the glare-free intuitive guidance according to the inverse relationship of the spatial positioning distance, and generate the target luminous intensity. S44. Generate a phototaxis guidance instruction based on the target luminous intensity and send the phototaxis guidance instruction to the reminder device group. After receiving the medication retrieval signal from the reminder device group, update the action abnormal interruption label to a medication status confirmation label. The light-guided instruction is used to instruct the alerting device group to perform the following steps: S441. Based on the light-guiding instruction, drive the light-emitting components of the reminder device group to emit light according to the target light intensity; S442. After the built-in sensor of the reminder device group detects the actual drug-taking action, the drug-taking signal is generated.

6. The method according to claim 1, characterized in that, The S5 includes: S51. Obtain the local vector clock, and accumulate the values ​​of the local node event components in the local vector clock according to the medication status confirmation tag to generate an updated local vector clock. S52. Based on the medication status confirmation tag, perform weak network causal synchronization processing, establish a causal mapping association between the medication status confirmation tag and the updated local vector clock, and encapsulate the updated local vector clock and the medication status confirmation tag with the established causal mapping association to generate the synchronization data packet. S53. Construct a network communication protocol for the synchronization data packet, configure the target broadcast address and instruction message header, generate the data instruction including the synchronization data packet, and broadcast the data instruction to the other node devices in the local area network.

7. A smart reminder method and system for assisting elderly people in medication use and health management, used to implement the method according to any one of claims 1 to 6, characterized in that, The device includes: The baseline feature extraction module is used to perform unsupervised learning on the acquired basic perception data to generate initial reaction features and background environment features. Based on the acquired medication prescription instructions, the initial reaction features and background environment features are time-space mapped to extract information and environmental parameters corresponding to the time window and the target pharmacy, and individualized reaction delay baseline, ambient acoustic background noise spectrum and medication time window are generated. The multimodal scheduling module is used to acquire multidimensional sensor data of the current environment, perform adaptive multimodal reminder scheduling based on the ambient acoustic background noise spectrum, the medication time window and the multidimensional sensor data, calculate the reach probability and annoyance cost of auditory mode, tactile mode and visual mode and construct a utility objective function for solution, and generate the optimal strategy combination vector. The optimal strategy combination vector is used to control the reminder device group to remind the user respectively. The action tracking and analysis module is used to acquire user action signals after being reminded, perform state probability fitting and temporal decoding on the user action signals through a hidden Markov model to generate a physical temporal state sequence, and perform anomaly analysis on the physical temporal state sequence based on the temporal integrity comparison rule to generate action state analysis results. The cognitive verification guidance module is used to generate a user reminder signal and send it to the reminder device group if the action status judgment result is an action abnormality interruption label. Based on the action abnormality interruption label and the individualized reaction delay baseline, the module performs cognitive verification on the feedback user voice response signal and generates a cognitive verification comprehensive score. When the cognitive verification comprehensive score is lower than the set cognitive verification threshold, the module obtains the spatial positioning distance of the reminder device group, generates a light-touch guidance instruction based on the spatial positioning distance and sends it to the reminder device group. After receiving the medication retrieval signal from the reminder device group, the module updates the action abnormality interruption label to a medication taken status confirmation label. The weak network status synchronization module is used to perform causal synchronization of weak network status based on the medication status confirmation tag, encapsulate data in combination with the acquired local vector clock, generate data instructions including synchronization data packets, and broadcast the data instructions to other node devices in the local area network.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images