Intelligent voice reminder system for medication adherence in cataract patients

Abstract

The present application relates to the technical field of medical health information and speech synthesis, and specifically discloses an intelligent voice reminding system for medication compliance of cataract patients, which comprises a user interaction terminal, a cloud collaborative processing platform and a medical data interface. The user interaction terminal determines the user's interactive state through multi-modal perception and local behavior recognition and reports it; the cloud platform dynamically decides the reminding time and content strategy through a dynamic strategy decision unit based on the user portrait, the medication task and the user state, and synthesizes personalized voice content by a multi-modal content generation unit for broadcasting. The system realizes precise timing reminding by sensing the user state, and dynamically generates differentiated reminding content based on the personalized portrait, effectively alleviating reminding fatigue, improving the acceptance of reminding and the efficiency of medication compliance management.

Description

Technical Field

[0001] This invention belongs to the field of medical and health information and speech synthesis technology, specifically relating to an intelligent voice reminder system for cataract patients to improve medication adherence. Background Technology

[0002] In the field of chronic disease management, patient medication adherence is a key factor affecting treatment effectiveness and prognosis. For patients who require long-term, regular medication, ensuring they take their medication on time and in the correct dosage is one of the core challenges facing healthcare management. Smart reminder technology is an important auxiliary means to improve adherence.

[0003] For post-cataract surgery patients, whose medication regimens are typically complex and lengthy, higher demands are placed on the accuracy and user-friendliness of reminder systems. Intelligent voice reminder systems, through voice interaction, aim to assist patients in completing the medication process in a more natural and acceptable way, reducing missed or incorrect doses due to forgetfulness or operational difficulties.

[0004] Existing intelligent voice reminder solutions mostly employ preset, fixed-content voice broadcast patterns. This mechanical and repetitive reminder method lacks emotional interaction and content variation. Long-term use can easily lead to auditory and psychological fatigue in patients, thereby reducing their attention to and willingness to respond to the reminder information, ultimately weakening the actual effectiveness of the reminder system. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent voice reminder system for cataract patients to improve medication adherence, thereby solving the technical contradiction in the prior art where fixed voice reminder patterns, due to the lack of emotional interaction and content variation, cause reminder fatigue in patients, which in turn reduces reminder attention and willingness to respond, ultimately weakening the effectiveness of medication adherence management.

[0006] To achieve the above objectives, this invention provides an intelligent voice reminder system for cataract patients' medication adherence. The system includes a user interaction terminal, a cloud-based collaborative processing platform, and a medical data interface. The user interaction terminal is deployed on the patient's side to collect user status data and perform voice interaction and physical reminders; the cloud-based collaborative processing platform is used for multimodal data analysis, dynamic reminder strategy generation, and personalized content synthesis; the medical data interface is used to securely acquire and synchronize the patient's standardized medication regimen and historical adherence records.

[0007] The user interaction terminal integrates a main control module, a multimodal perception module, a voice interaction module, and a physical execution module. The multimodal perception module continuously collects environmental audio data and user visual data from preset viewpoints. The main control module embeds a local lightweight behavior recognition engine, which processes the raw data collected by the multimodal perception module in real time. The specific execution steps are as follows: First, feature extraction is performed on the environmental audio data to identify whether it contains user voice activity; second, key point detection and posture analysis are performed on the user visual data to identify whether the user is in a static sitting or lying position; finally, the voice activity recognition results and user posture recognition results are combined to generate a local behavior label representing the user's current interactive state. This label includes three states: idle and interactive, busy and temporarily uninterrupted, and user not detected after leaving the area.

[0008] The cloud-based collaborative processing platform comprises a user profile modeling unit, a dynamic strategy decision-making unit, and a multimodal content generation unit. The user profile modeling unit receives medication plan data and historical adherence records from medical data interfaces, as well as interaction history data uploaded from user interaction terminals, to construct and continuously update personalized user profiles for patients. This user profile includes at least three dimensions: medication knowledge level, preference for reminder methods, and historical interaction response patterns. Medication knowledge level is quantitatively assessed by evaluating the accuracy of patients' answers to questions regarding drug names, usage, and side effects in their historical voice Q&A sessions. Reminder method preference is statistically summarized by analyzing the rate of positive feedback or subsequent adherence improvement rates from patients' proactive responses to different tones, announcer voices, and reminder content structures in historical records. Historical interaction response patterns are modeled by statistically analyzing the patient's response delay time and confirmation rate to reminders under different time periods and user behavior tags.

[0009] The dynamic strategy decision-making unit is the core decision-making hub of the system. Based on the real-time user profile output by the user profile modeling unit, the current medication task node, and the local behavior tags reported by the user interaction terminal, it generates dynamic reminder strategies. This unit internally operates a hierarchical decision-making logic. The first layer is the timing decision layer, whose decision logic is as follows: when the medication task node reaches the preset reminder time window, it queries the local behavior tags reported by the user interaction terminal; if the tag is "busy and temporarily uninterrupted," the decision is to delay the reminder execution and start an adjustable waiting timer, continuously monitoring changes in the behavior tags until the timer expires; if the tag is "left and no user detected," the decision is to trigger a one-time physical reminder as a primary notification; if the tag is "idle and interactive," the decision is to immediately execute a voice reminder. The second layer is the content strategy layer, which is activated after the timing decision layer determines that the voice reminder should be executed immediately. Its decision logic is as follows: based on the user profile's level of medication knowledge, if the evaluation value is lower than the preset threshold, an educational reinforcement content strategy is selected; if the evaluation value is higher than or equal to the preset threshold, it is further based on historical interaction response patterns. If the analysis shows that the recent response delay has an increasing trend, an attention-awakening content strategy is selected; otherwise, a regular confirmation content strategy is selected.

[0010] The multimodal content generation unit receives specific content strategy instructions from the dynamic strategy decision-making unit and calls upon preference dimension data from the user profile to synthesize personalized reminder content in real time. This unit includes a structured speech corpus, a speech synthesis engine, and a feedback question generator. The structured speech corpus stores various text templates categorized by content strategy type, with each template reserving variables in key information slots. For educational reinforcement content strategies, the corresponding text template includes basic information such as medication actions, drug names, and dosages, as well as embedded variables such as brief explanations of the importance of this step or common misconceptions. For attention-awakening content strategies, the text template includes patient-related or concrete scenario-based questions at the beginning. The speech synthesis engine synthesizes speech from the final text after filling in the variables, based on the preferred tone parameters analyzed from the user profile. The feedback question generator is automatically activated after each voice reminder broadcast, generating a closed-ended voice question to confirm whether the patient has performed the medication action. The question is expressed in different sentence structures in each interaction but has the same meaning. For example, this time it is "Have you finished instilling the eye drops?", and next time it may be "Have you finished instilling the eye drops?"

[0011] The physical execution module is integrated into the user interaction terminal and is controlled by the main control module. When the timing decision layer of the dynamic strategy decision unit outputs a decision to trigger a physical reminder, or when the feedback question broadcast by the voice interaction module does not receive a voice confirmation response from the user within a preset time, the main control module sends a trigger command to the physical execution module. The physical execution module includes a micro-motor-driven eccentric wheel mechanism, which, when activated, causes the device to produce a non-sharp vibration alert lasting several seconds.

[0012] The voice interaction module is also integrated into the user interaction terminal, including an audio codec, a local wake-word detection submodule, and a cloud-based semantic understanding relay submodule. Its workflow is as follows: The local wake-word detection submodule runs continuously in standby mode; when a preset wake-word is detected or an active reminder command is received from the main control module, audio acquisition is initiated and uploaded to the cloud-based collaborative processing platform; the cloud-based collaborative processing platform performs semantic recognition and understanding of the speech, and sends the understanding results or the speech content to be played down to the user interaction terminal; the audio codec of the voice interaction module ultimately completes the audio playback or acquisition encoding. For immediately executed voice reminder commands issued by the dynamic strategy decision-making unit, the content is synthesized by the multimodal content generation unit and then played through this module. For user-initiated voice inquiries, the semantics are processed by the cloud, and the multimodal content generation unit or the medical knowledge base interface can provide and play the response content.

[0013] Furthermore, the historical interaction response pattern modeling in the user profile modeling unit employs a sliding window statistical method based on time decay weights. Specifically, the system divides the past 30 days into six consecutive 5-day time windows, assigning a decay weight to each window, with windows closer to the current date having a higher weight. Within each time window, the system statistically analyzes the average response delay time and confirmation rate of patients in an idle, interactive state for voice prompts. The final historical interaction response pattern output consists of two weighted averages: one for the weighted average response delay time and the other for the weighted average confirmation rate. The content strategy layer of the dynamic strategy decision-making unit compares the weighted average response delay time of the latest time window with the value of the previous time window to determine if there is an increasing trend.

[0014] Furthermore, when selecting an attention-awakening content strategy, the content strategy layer of the dynamic strategy decision-making unit includes contextual element parameters in its instructions. Based on these parameters, the multimodal content generation unit selects matching question text fragments from its contextual question variable library. This variable library is constructed and updated through anonymized analysis of patients' historical voice interaction records, extracting frequently mentioned or concerned daily life elements such as weather, kinship terms, and favorite TV program names, and transforming these elements into general contextual templates.

[0015] Furthermore, the system operates within a closed-loop optimization framework. Each complete reminder interaction event, including reminder triggering, content playback, user feedback, and whether it is ultimately recorded as successful medication use in the medical data interface, is recorded as a data sample. The user profile modeling unit periodically, for example, every 7 days, performs batch analysis on all newly generated data samples, recalculating the quantitative indicators of each dimension in the user profile, thereby achieving dynamic updates to the user profile. The decision parameters and content library of the dynamic strategy decision-making unit and the multimodal content generation unit can be periodically optimized and expanded based on anonymized and desensitized data of the overall user group.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by introducing multimodal perception and local real-time behavior recognition, enables the system to perceive the user's current interaction state. The dynamic strategy decision-making unit makes reminder timing decisions based on accurate local behavior tags, avoiding ineffective voice broadcasts when the user is busy or away, and accurately delivering reminder interventions to the user at an acceptable time. This reduces potentially annoying and ineffective disturbances from the source, improving the acceptability of reminders and the humanization level of the system.

[0017] 2. This invention constructs a dynamic content generation system that deeply integrates personalized user profiles. Instead of broadcasting fixed content, the system dynamically selects different strategies, such as educational reinforcement, attention awakening, or routine confirmation, based on continuous learning of the patient's medication awareness, content preferences, and response patterns, and synthesizes personalized voice content that matches the patient's preferences. This continuous variation and targeting of content effectively breaks the auditory and psychological inertia caused by mechanical repetition, significantly alleviates reminder fatigue caused by long-term use, and maintains the patient's sense of novelty and attention to the reminder information.

[0018] 3. This invention uses physical reminders as a supplement and backup to voice reminders, and designs an interactive closed loop based on feedback confirmation. When the user does not respond to the voice interaction in a timely manner, the system provides a gentle reminder through vibration; at the same time, each reminder is accompanied by a confirmation question with varying sentence structure, ensuring that the completion of the interaction has a clear endpoint. This multi-channel, strong feedback design ensures reliable delivery and execution confirmation of reminder signals in complex home environments, effectively reducing the probability of missed doses due to simple forgetfulness or distraction, thereby systematically improving the management efficiency and reliability of medication adherence. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical architecture of the intelligent voice reminder system for cataract patients' medication adherence proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the dynamic strategy decision-making unit in this invention; Figure 3 This is a logical flow diagram of the user profile modeling unit in this invention; Figure 4 This is a schematic diagram of the interaction relationship and data flow between the multimodal content generation unit and the voice interaction module in this invention; Figure 5 This is a schematic diagram illustrating the operating principle of the system closed-loop optimization framework in this invention. Detailed Implementation

[0020] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 This invention proposes an intelligent voice reminder system for cataract patients to improve medication adherence. The system consists of three main parts: a user interaction terminal deployed on the patient's side, a cloud-based collaborative processing platform located in a remote server cluster, and a medical data interface that connects to the medical institution's information system. These three components interact in real-time via a secure and encrypted communication protocol, jointly achieving intelligent, personalized, and high-adherence management of cataract patients' home medication use, such as eye drops.

[0021] The user interaction terminal is the core carrier for direct physical and informational interaction between the system and the patient. It integrates a main control module, a multimodal perception module, a voice interaction module, and a physical execution module. The multimodal perception module includes a wide-angle camera pointing to a preset viewing angle and a high-sensitivity omnidirectional microphone array. The wide-angle camera continuously captures visual images centered on the patient's usual sitting or lying area at a frame rate of 15 frames per second and a resolution of 640×480 pixels. The omnidirectional microphone array simultaneously acquires ambient audio signals within a 3-meter radius at a sampling rate of 16 kHz.

[0022] The main control module incorporates a local lightweight behavior recognition engine. This engine employs an embedded neural network model to process the raw data stream output by the multimodal perception module with low latency. Its processing flow first performs a short-time Fourier transform on the audio signal, extracting Mel-frequency cepstral coefficient features, which are then input into a binary classification convolutional neural network to determine whether valid human voice activity exists in the current audio segment. The output is a Boolean value: present or absent. Secondly, it performs human keypoint detection on the visual image, using a lightweight pose estimation algorithm to identify key coordinate points of the head, shoulders, elbows, and hips. Based on the spatial distribution and motion vectors of these coordinate points, it determines whether the user is in a static sitting or lying position. The threshold for this determination is set at a displacement of less than 5 pixels for each keypoint within two consecutive seconds. Finally, the main control module logically merges the outputs of the two subtasks: if both "human activity is present" and "the user is in a static sitting / lying position," a local behavior tag "idle and interactive" is generated; if "no human activity is present" but the user is in a static sitting / lying position, a tag "busy and temporarily uninterrupted" is generated; if "no human key points are detected" or "human key points are in a state of violent movement," a tag "no user detected upon leaving the site" is generated. This local behavior tag is uploaded to the cloud-based collaborative processing platform every 5 seconds as a key input for dynamic reminder strategy decisions.

[0023] The cloud-based collaborative processing platform, serving as the system's intelligent hub, comprises three core functional modules: a user profile modeling unit, a dynamic strategy decision-making unit, and a multimodal content generation unit. The user profile modeling unit receives data from three channels: the first channel is a medical data interface, which periodically synchronizes standardized medication regimen data for patients via a secure API compliant with the HL7 medical information exchange standard. This data includes the generic name of the medication, the dosage per infusion, the daily frequency of administration, the medication time window (e.g., 9:00 AM ± 30 minutes), and the total number of days in the treatment course. The second channel, also from the medical data interface, retrieves historical adherence records from medical institutions documented during past medication cycles. These records are stored as a sequence of Boolean values, indicating whether each daily medication administration was completed on time. The third channel receives historical interaction data uploaded from the user's interactive terminal, including the timestamp of each reminder event, trigger type, summary of the broadcast content, user voice reply text, response latency (the time interval from the end of the broadcast to receiving a confirmation reply), and whether the system ultimately determines the medication administration was successful.

[0024] Please refer to the attached document. Figure 3The user profiling modeling unit performs structured processing on the aforementioned multi-source heterogeneous data to construct and maintain a three-dimensional personalized user profile. The first dimension is medication knowledge level, quantified by the system periodically (e.g., weekly) posing three closed-ended questions about the user's current medication regimen, such as "Should I use eye drops in my left or right eye today?", "How many times a day should I use this eye drop?", and "How many minutes should I close my eyes after using the eye drops?". The user's voice responses are converted to text through cloud-based semantic understanding and matched against standard answers to calculate the accuracy rate. This accuracy rate, after smoothing and filtering, serves as the real-time evaluation value for the medication knowledge level dimension, ranging from 0 to 1. The second dimension is reminder method preference, modeled as follows: The system records users' proactive feedback behaviors to different reminder content in historical interactions, including explicitly expressing evaluative statements such as "I like this voice" or "Try a different way of saying it next time," as well as improvements in compliance behavior after specific types of reminders (e.g., a compliance rate increase of more than 10% the day after a certain type of reminder). The system statistically analyzes the positive review rate and compliance improvement rate for each combination of timbre, tone intensity, and content structure, and forms a preference vector after weighted averaging. The third dimension is the historical interaction response pattern, which is modeled using a sliding window statistical method based on time decay weights. Specifically, the system divides the past 30 days into 6 consecutive 5-day time windows, namely... (Last 5 days) (Days 6-10), ... (Days 26-30), and assign decay weights to each window with a weight sequence of 0.5, 0.3, 0.1, 0.05, 0.03, and 0.02. In each window... Within this window, the system only counts voice notification events that occur when the local behavior label is "idle and interactive," and calculates the average response latency Ti and average confirmation rate within that window. The final weighted average response delay time is... The weighted average confirmation rate is This modeling method ensures that recent behavior has a much greater impact on the current profile than long-term behavior, giving the profile a high degree of timeliness and sensitivity.

[0025] The dynamic strategy decision-making unit is the core of the entire system's intelligent decision-making. Please refer to the appendix. Figure 2The unit's input consists of three parts: first, real-time 3D profile data output by the user profile modeling unit; second, information on the current medication task node to be executed, including drug name, target eye, and planned execution time; and third, the latest local behavior tags reported by the user interaction terminal. Internally, it operates with a strict hierarchical decision-making logic. The first layer is the timing decision layer, whose workflow is as follows: When the system clock enters the preset reminder time window for a certain medication task node, the timing decision layer immediately queries the latest local behavior tags. If the tag is "Busy, temporarily uninterrupted," the decision result is "delay execution," and a waiting timer with an initial duration of 2 minutes is started. During this period, the system checks the newly uploaded behavior tags every 5 seconds. Once the tag changes to "Idle, interactive," the timer is immediately terminated and a voice reminder is triggered. If the timer expires without a change, a voice reminder is forcibly triggered, but the volume is reduced by 10% to show respect. If the tag is "Leaving, no user detected," the decision result is "trigger physical reminder," and a physical reminder command is sent to the main control module of the user interaction terminal. If the label is "Idle and Interactive", the decision is "Immediately Execute Voice Reminder" and the second-level content strategy layer is activated.

[0026] The content strategy layer begins operation upon receiving the instruction to "immediately execute voice reminder." Its decision-making logic first reads the medication knowledge level assessment value from the user profile. If this value is below a preset threshold of 0.6, it determines that the patient's understanding of this medication use may be insufficient, and selects the "educational reinforcement content strategy." If the assessment value is greater than or equal to 0.6, it further analyzes the weighted average response latency T in historical interaction response patterns. The system compares the average response latency T1 of the latest time window W1 with T2 of the previous window W2. If the increase exceeds 20%, it is determined that the patient's recent attention has declined, and the "attention-awakening content strategy" is selected; otherwise, the "routine confirmation content strategy" is selected. This decision-making process ensures that the strategy selection of reminder content is always dynamically aligned with the patient's current cognitive state and attention level.

[0027] After receiving specific policy instructions from the content policy layer, the multimodal content generation unit initiates the personalized content synthesis process. Please refer to the appendix. Figure 4This unit maintains a structured speech corpus, which stores hundreds of text templates categorized by strategy type. For the "Educational Reinforcement Content Strategy," the template structure is: "[Address], it's [time] now, it's time to administer [medication name]. This time it's [left / right] eye ​​drops, [dosage] drops. After administering, please close your eyes and rest for 2 minutes so the medication can work better." The [Address] variable is filled with "Grandpa Zhang" or "Aunt Li," etc., based on the user's registration information; [time] is the current system time; [medication name], [left / right], and [dosage] are all extracted from the current medication task node. Additionally, a knowledge prompt slot is added at the end of the template, such as "Many people forget to close their eyes, but this step is important to prevent the eye drops from running out too quickly." For the "Attention Awakening Content Strategy," the template begins with a scenario-based question variable, such as "The sun is shining brightly outside today, are you going to the balcony to sunbathe? Before that, let me administer your eye drops." This scenario element comes from a dynamically updated scenario variable library. The library is constructed as follows: The system periodically performs natural language processing on all users' anonymized voice interaction records, extracting frequently occurring entity nouns such as "grandson," "weather," and "TV series 'The World'," and generalizing them into templates, such as "I heard you are very concerned about [relative's] studies; you can call him after you finish your IV drip." The multimodal content generation unit matches the most relevant question text fragments from the library based on the contextual element parameters carried in the content strategy instructions.

[0028] After the text template is filled in, the speech synthesis engine starts. This engine is an end-to-end text-to-speech model based on deep learning, supporting multiple preset timbres. The system selects the timbre parameters with the highest historical positive feedback rates from patients (e.g., "gentle female voice" or "steady male voice") based on the user's preferences, and synthesizes the filled text into a natural and fluent audio file. Finally, the feedback question generator is automatically activated, randomly selecting one question from a closed question library containing 20 different sentence structures, such as "Have you finished putting in the eye drops?", "Have you finished putting in the eye drops?", "Have you finished putting in the eye drops?", and appending the question after the main notification content to form a complete broadcast sequence.

[0029] The voice interaction module is responsible for playing the synthesized speech and acquiring user voice data. This module includes a low-power audio codec, a local wake-word detection submodule, and a cloud-based semantic understanding relay submodule. During non-reminder periods, the module is in standby mode, running only the local wake-word detection submodule. This submodule uses keyword spotting technology to continuously monitor preset wake-words such as "little assistant," consuming less than 1 milliwatt. Once a wake-word is detected, or an active reminder command is received from the main control module, the module immediately activates the audio codec, begins acquiring user voice data, and encodes it into a 16kHz, 16-bit PCM format data stream, which is then uploaded to the cloud-based collaborative processing platform via the TLS 1.3 protocol. The cloud platform uses a large language model to perform semantic parsing of the speech, understands the user's intent, and generates corresponding response text. This text is then sent to the multimodal content generation unit for speech synthesis, or directly calls the medical knowledge base interface to obtain standard answers before synthesis, and finally downloads it to the user's interactive terminal for playback. For proactive reminders triggered by the dynamic strategy decision unit, the synthesized complete speech sequence (main reminder + feedback question) is played directly through the audio codec of this module, and the volume is automatically adjusted according to the ambient noise level, ranging from 50 to 70 decibels.

[0030] The physical execution module, serving as a supplement and backup mechanism to the voice prompts, is integrated into the base of the user interaction terminal. This module consists of a miniature DC motor (8 mm in diameter), an eccentric mass block, and a flexible silicone shell. When the main control module receives a "trigger physical prompt" command from the cloud-based collaborative processing platform, or if no valid user voice confirmation (i.e., no affirmative keywords such as "drip done" or "completed") is detected within 15 seconds after the voice interaction module finishes broadcasting the feedback question, the main control module sends a high-level trigger signal to the physical execution module. The miniature motor then starts, driving the eccentric wheel to rotate, generating a gentle vibration with a frequency of 120 Hz and an amplitude of 0.5 mm, lasting for 5 seconds. This vibration intensity is sufficient to cause a slight resonance on a desktop or bedside table, thus reaching the patient's tactile senses without producing harsh noise or causing fright, making it particularly suitable for elderly cataract patients with limited vision.

[0031] The entire system operates within a closed-loop optimization framework; please refer to the appendix. Figure 5Each complete reminder interaction event, from the moment the dynamic strategy decision-making unit makes a decision to the final record of "successful medication" or "missed dose" in the medical data interface, is encapsulated as a structured data sample. This sample contains dozens of fields, such as decision time, strategy type, content template ID, voice ID, user response delay, confirmation result, and final compliance status. The user profile modeling unit is equipped with a batch processing scheduler that executes a full profile update task every 7 days. This task iterates through all new samples generated in the past 7 days, recalculates various indicators of medication knowledge level, preference vector, and response pattern, and overwrites the old profile. In addition, the system backend has a swarm intelligence optimization engine that periodically (e.g., monthly) analyzes the aggregated data of all anonymized users. For example, if it finds that the "attention-awakening" strategy significantly improves the average compliance rate among users over 75 years old compared to other strategies, the default strategy threshold for that age group is automatically adjusted; if a newly constructed scenario question template receives higher approval ratings in A / B testing, it is added to the public corpus for all users to use. This mechanism, which combines dynamic updates to individual profiles with feedback from collective knowledge, ensures the continuous evolution of system capabilities and the enhancement of universality.

[0032] In summary, this embodiment constructs an intelligent voice reminder system that proactively adapts to the individual differences and real-time status of cataract patients through precise multimodal perception, rigorous hierarchical decision-making, deeply personalized multimodal content generation, and closed-loop feedback optimization. This system fundamentally solves the reminder fatigue problem caused by traditional fixed-pattern reminders, transforming each medication reminder into a precise, warm, and effective health intervention, thereby steadily maintaining and improving patient medication adherence over long-term use.

[0033] Example 2: Building upon Example 1, this example introduces an adaptive reminder intensity adjustment mechanism based on ambient light and circadian rhythms to further enhance the system's robustness and user experience in complex home environments. This mechanism primarily affects the broadcast parameters of the voice interaction module and the vibration intensity of the physical execution module, with its control logic embedded in the timing decision layer of the dynamic strategy decision unit.

[0034] Specifically, the multimodal perception module of the user interaction terminal integrates a digital ambient light sensor in addition to the existing camera and microphone. This sensor collects the illuminance value of the device's location every 10 seconds, in lux, with a measurement range of 0 to 65,535 lux. The main control module compares this illuminance value with a preset circadian rhythm model. This model does not simply divide day and night according to the system time, but combines geographical location information (provided by the user during initial setup) with astronomical algorithms to calculate the precise sunrise and sunset times of the day, and defines three illumination intervals accordingly: the night interval (1 hour after sunset to 1 hour before sunrise), with typical illuminance below 50 lux; the twilight transition interval (1 hour before sunrise to 1 hour after sunrise and 1 hour before sunset to 1 hour after sunset), with an illuminance range of 50 to 500 lux; and the daytime interval (the rest of the time), with illuminance typically above 500 lux.

[0035] After making a decision to "immediately execute a voice reminder" or "trigger a physical reminder," the timing decision layer of the dynamic strategy decision unit will additionally read the current ambient illuminance value and the corresponding light range. If the current period is nighttime, the system will automatically activate "nighttime silent mode." In this mode, the maximum volume of the voice interaction module is forcibly limited to 45 decibels, and the voice synthesis engine will select a dedicated nighttime tone with a slower speech rate and a lower tone; at the same time, the vibration intensity of the physical execution module is reduced to 50% of the original setting, that is, the amplitude is reduced to 0.25 mm, to avoid disturbing the patient's sleep in the middle of the night. If the current period is the transition between day and night, the system activates "mild mode," with the maximum voice volume at 55 decibels and the vibration intensity maintained at normal; if the current period is daytime, "standard mode" is activated, and all parameters are executed according to the settings in Example 1.

[0036] Furthermore, this adaptive mechanism is linked to historical interaction response patterns in the user profile. For example, if the system detects that a patient's average confirmation rate is significantly lower during the night than during the day, it will label them as "night-sensitive" in their profile. Subsequently, when this patient triggers an alert during the night, the system not only activates a night-time silent mode but also further adjusts its content strategy: downgrading the previously selected "attention-awakening" strategy to "routine confirmation," because contextual questions may be considered unnecessary disturbances at night. This adjustment mechanism, which deeply integrates environmental awareness with individual behavioral patterns, allows the system to respect patients' circadian rhythms and privacy to the greatest extent possible without sacrificing the effectiveness of alerts, making it particularly suitable for elderly cataract patients who require long-term, frequent medication.

[0037] Example 3: Based on the system architecture of Example 1, this example adds support for complex medication plans involving multiple drugs and multiple eyes, and strengthens the medication error prevention mechanism. This capability is mainly achieved by expanding the data model of the medical data interface, enhancing the conflict detection logic of the dynamic strategy decision-making unit, and enriching the warning templates of the multimodal content generation unit.

[0038] At the medical data interface level, the system no longer supports single-drug dosing regimens but can receive and parse composite dosing regimens that include multiple drugs, different eyes, and staggered time points. For example, a patient might use drug A (left eye, 3 times daily) and drug B (both eyes, twice daily) simultaneously, with at least a 5-minute interval between doses. To address this, the medical data interface's data model has been expanded, with each medication task node adding a "drug ID," a "target eye set" (which can be left eye, right eye, or both eyes), and a "minimum dosing interval" field. Upon receiving such composite regimens, the cloud-based collaborative processing platform performs time-series planning to generate a globally conflict-free queue of alert events.

[0039] Before making a timing decision, the dynamic strategy decision-making unit first executes a medication conflict detection subroutine. This subroutine checks whether the time interval between the current medication task to be executed and the previously successfully executed medication task meets the "minimum medication interval" requirement. If not, for example, if the patient just instilled medication A 3 minutes ago and it's time to instill medication B, but the required interval is 5 minutes, the decision-making unit will temporarily overwrite the local behavior label of this reminder as "busy and temporarily uninterrupted" and start a precise countdown timer for the remaining required interval time (2 minutes in this example). During this period, even if the multimodal perception module reports "idle and interactive," the system will not trigger a reminder, thus eliminating potential eye irritation or medication efficacy interference caused by medication mixing or insufficient intervals.

[0040] The multimodal content generation unit has correspondingly expanded its structured speech corpus, adding two new templates: "Drug Administration Sequence Warning" and "Eye Identification Reinforcement." When the system detects that a patient is about to perform a multi-step medication procedure (e.g., instilling medication A into the left eye first, then instilling medication B into both eyes 5 minutes later), if the patient asks in advance "What should I do next?", the system will adopt an educational reinforcement strategy and announce: "You have just instilled medication A into your left eye. Now you need to wait 3 minutes before you can instill medication B. This time, you need to instill medication into both eyes." When reminding the patient to administer medication into both eyes, the feedback question generator will generate more targeted confirmation questions, such as "Have you finished instilling the eye drops into both your left and right eyes?", rather than the general "Have you finished?". This refined content design greatly reduces the risk of medication errors caused by eye confusion or missed steps, providing a higher level of safety for cataract patients with complex conditions.

Claims

1. An intelligent voice reminder system for improving medication adherence in cataract patients, characterized in that, include: The user interaction terminal, deployed on the patient's side, is used to collect user status data and perform voice interaction and physical reminders; A cloud-based collaborative processing platform is used for multimodal data analysis, dynamic reminder strategy generation, and personalized content synthesis. A medical data interface for securely acquiring and synchronizing patients’ standardized medication regimens and historical adherence records; The user interaction terminal integrates a main control module, a multimodal perception module, a voice interaction module, and a physical execution module; The cloud-based collaborative processing platform includes a user profile modeling unit, a dynamic strategy decision-making unit, and a multimodal content generation unit.

2. The intelligent voice reminder system for cataract patient medication adherence according to claim 1, characterized in that, The user profile modeling unit is used to receive medication plan data and historical adherence records from the medical data interface, as well as interaction history data uploaded from the user interaction terminal, to build and continuously update the patient's personalized user profile. The user profile includes at least the dimensions of medication knowledge level, reminder method preference, and historical interaction response pattern. The dynamic strategy decision-making unit is used to generate a dynamic reminder strategy based on the real-time user profile output by the user profile modeling unit, the current medication task node, and the local behavior tags reported by the user interaction terminal. The dynamic strategy decision-making unit internally operates a hierarchical decision-making logic, including a timing decision layer and a content strategy layer; The multimodal content generation unit is used to receive specific content strategy instructions issued by the dynamic strategy decision unit and call the preference dimension data in the user profile to synthesize personalized reminder content in real time.

3. The intelligent voice reminder system for cataract patient medication adherence according to claim 2, characterized in that, The multimodal perception module is used to continuously collect environmental audio data and user visual data from a preset perspective; The main control module has a built-in local lightweight behavior recognition engine, which is used to process the raw data collected by the multimodal perception module in real time and generate local behavior tags that represent the user's current interactive state. The local behavior tags include three states: idle and interactive, busy and temporarily inactive, and no user detected when the user is away. The voice interaction module is integrated into the user interaction terminal and is used to perform voice content playback and user voice collection and uploading. The physical execution module is integrated into the user interaction terminal and controlled by the main control module, and is used to generate physical vibration prompts under specific conditions.

4. The intelligent voice reminder system for cataract patient medication adherence according to claim 3, characterized in that, The local lightweight behavior recognition engine processes the raw data in real time as follows: First, feature extraction is performed on the environmental audio data to identify whether it contains user voice activity; Secondly, key point detection and posture analysis are performed on user visual data to identify whether the user is in a static sitting or lying position. Finally, the local behavior labels are generated by combining the voice activity recognition results and the user posture recognition results.

5. The intelligent voice reminder system for cataract patient medication adherence according to claim 4, characterized in that, The decision-making logic of the timing decision layer of the dynamic strategy decision-making unit is as follows: When the medication task node reaches the preset reminder time window, query the local behavior tag; If the tag is "Busy and cannot be interacted with temporarily", the decision is to delay the execution of the reminder and start a waiting timer to continuously monitor changes in the behavior tag; If the tag indicates that no user was detected upon leaving the premises, the decision is to trigger a physical alert. If the tag is idle and interactive, the decision is to immediately execute a voice prompt and activate the content strategy layer.

6. The intelligent voice reminder system for cataract patient medication adherence according to claim 5, characterized in that, The decision-making logic of the content strategy layer is as follows: Based on the evaluation value of the medication knowledge level dimension in the user profile, if the evaluation value is lower than the preset threshold, an educational reinforcement content strategy is selected. If the evaluation value is higher than or equal to the preset threshold, then based on the historical interaction response pattern, if the analysis shows that the recent response delay has an increasing trend, then the attention-awakening content strategy is selected; otherwise, the regular confirmation content strategy is selected.

7. The intelligent voice reminder system for cataract patient medication adherence according to claim 6, characterized in that, The multimodal content generation unit includes a structured speech corpus, a speech synthesis engine, and a feedback question generator; The structured speech corpus stores various text templates with reserved variables in key information slots, categorized according to content strategy types. The speech synthesis engine is used to synthesize speech from the final text after variable filling based on the preferred timbre parameters analyzed from the user profile. The feedback question generator is automatically activated after each voice reminder broadcast to generate a closed-ended voice question to confirm the execution of the medication action. The sentence structure of the closed-ended voice question is different in each interaction, but the semantics are the same.

8. The intelligent voice reminder system for cataract patient medication adherence according to claim 7, characterized in that, The text template corresponding to the educational reinforcement content strategy includes, in addition to basic information such as medication action, drug name and dosage, a brief explanation of the importance of the medication step or a reminder of common misconceptions. The text template corresponding to the attention-awakening content strategy includes patient-related or concrete scenario-based question variables at the beginning.

9. The intelligent voice reminder system for cataract patient medication adherence according to claim 8, characterized in that, The historical interaction response pattern modeling in the user profile modeling unit adopts a sliding window statistical method based on time decay weight; The specific process is as follows: Divide the past 30 days into 6 consecutive 5-day time windows, and assign a decay weight to each window, with the window closer to the current date having a higher weight; Within each time window, the average response delay time and confirmation rate of patients to voice prompts in an idle and interactive state were statistically analyzed. The final output consists of two weighted averages: the weighted average response delay time and the weighted average confirmation rate.

10. The intelligent voice reminder system for cataract patient medication adherence according to claim 9, characterized in that, The content strategy layer of the dynamic strategy decision unit determines whether there is an increasing trend in response latency by comparing the weighted average response latency of the latest time window with the value of the previous time window.

Images