Medication state calibration method and device based on interaction feedback and electronic equipment

Abstract

The application discloses a medication state calibration method and device based on interaction feedback and electronic equipment, and relates to the technical field of medical information processing, in particular to the field of chronic disease management. The method comprises the following steps: collecting interaction feedback data between a user and an interaction system; performing semantic analysis on the interaction feedback data by using a large language model to obtain feature information related to the actual medication state of the user; determining a dynamic adherence coefficient according to the feature information, wherein the dynamic adherence coefficient is used to represent the real-time execution degree of the user to the medication plan; and calibrating the medication state estimation model of the user online according to the dynamic adherence coefficient to obtain a target medication state, wherein the target medication state comprises an updated estimated value of the remaining amount of the medicine. The application solves the technical problem of distorted medication adherence evaluation caused by the fact that the existing technology only relies on the purchase behavior to infer the medication state.

Description

Technical Field

[0001] This application relates to the field of medical information processing technology, and more particularly to the field of chronic disease management. Specifically, it relates to a method, device, and electronic device for calibrating medication status based on interactive feedback. Background Technology

[0002] In the field of chronic disease management, medication adherence monitoring mainly relies on users' medication purchase behavior data. The system infers users' medication status by collecting medication purchase records.

[0003] However, there is a significant discrepancy between medication purchase behavior and actual medication use. Users may hoard medication or fail to take it regularly after purchase, leading to a mismatch between the inferred medication status based on purchase behavior and the actual situation. When user adherence declines, existing technologies struggle to identify the specific reasons, resulting in a lack of targeted interventions.

[0004] Therefore, existing technologies suffer from the problem of distorted medication adherence assessment due to relying solely on drug purchase behavior to infer medication status. Summary of the Invention

[0005] This application provides a medication status calibration method, apparatus, and electronic device based on interactive feedback, to at least solve the technical problem of distorted medication adherence assessment caused by relying solely on medication purchase behavior to infer medication status in the prior art.

[0006] According to one aspect of the embodiments of this application, a medication status calibration method based on interactive feedback is provided, comprising: collecting interactive feedback data between a user and an interactive system, the interactive feedback data including text data and / or voice data; performing semantic parsing on the interactive feedback data using a large language model to obtain feature information related to the user's actual medication status; determining a dynamic compliance coefficient based on the feature information, the dynamic compliance coefficient being used to characterize the user's real-time execution of the medication plan; and calibrating the user's medication status estimation model online based on the dynamic compliance coefficient to obtain a target medication status, wherein the target medication status includes an updated estimate of the remaining drug quantity, and the medication status estimation model is a mathematical model used to estimate the remaining drug quantity based on drug purchase behavior and medication compliance.

[0007] Optionally, the system collects interactive feedback data between the user and the interactive system, including: obtaining the user's interactive text data through an instant messaging platform interface; and / or obtaining the voice data of telephone follow-ups through an outbound calling system and transcribing the voice data into text data.

[0008] Optionally, a large language model is used to perform semantic parsing on the interactive feedback data to obtain feature information related to the user's actual medication status. This includes: using the interactive feedback data as model input data, performing semantic parsing using a large language model to obtain the parsing results; extracting feature information from the parsing results, whereby the feature information is used to characterize the user's current medication status, and the feature information includes at least the following dimensions: medication adherence tendency, symptom grading score, risk score, and semantic confidence.

[0009] Optionally, before using the large language model to perform semantic parsing on the interaction feedback data, the method further includes: obtaining the interaction text sequence between the user and the interaction system; recalling similar cases in a pre-built historical case library based on the interaction text sequence, and filling the context of the large language model with at least one recalled target historical case, wherein each historical case in the historical case library is represented as a combination of historical interaction text and corresponding business vector.

[0010] Optionally, determining the dynamic compliance coefficient based on feature information includes: obtaining semantic confidence from the feature information, where semantic confidence is used to characterize the degree of confidence of the large language model in the semantic parsing results; obtaining the most similar case recalled from the historical case library based on the interaction text sequence between the user and the interaction system; updating the semantic confidence based on the similarity between the most similar case and the interaction text sequence to obtain the effective confidence, where the effective confidence is equal to the square root of the product of semantic confidence and similarity; determining the instantaneous compliance coefficient based on medication adherence tendency, symptom rating score, and risk score from the feature information; obtaining the user's historical average compliance baseline value and the time interval between the current moment and the last effective text interaction; and determining the dynamic compliance coefficient with time decay based on the effective confidence, time interval, instantaneous compliance coefficient, and historical average compliance baseline value.

[0011] Optionally, the transient compliance coefficient is determined based on the medication adherence tendency, symptom grading score, and risk score in the feature information, including: obtaining the weight coefficients corresponding to each of the medication adherence tendency, symptom grading score, and risk score in the feature information; calculating the sum of the products of the medication adherence tendency, symptom grading score, and risk score with their respective weight coefficients, and adding a bias term to obtain a weighted sum; inputting the weighted sum into a soft-addition function for numerical transformation to obtain a transient compliance coefficient greater than zero, wherein the soft-addition function is used to map any real number to a positive real number.

[0012] Optionally, the dynamic compliance coefficient with time decay is determined based on the effective confidence level, time interval, instantaneous compliance coefficient, and historical average compliance benchmark value. This includes: obtaining a time decay factor, wherein the time decay factor is equal to the negative value of the natural constant multiplied by a preset time decay constant and then multiplied by the exponent of the time interval; calculating the product of the effective confidence level and the time decay factor to obtain a first weighting coefficient; determining a second weighting coefficient based on the first weighting coefficient, wherein the second weighting coefficient is equal to 1 minus the first weighting coefficient; calculating the product of the first weighting coefficient and the instantaneous compliance coefficient to obtain a first product value; calculating the product of the second weighting coefficient and the historical average compliance benchmark value to obtain a second product value; and adding the first product and the second product to obtain the dynamic compliance coefficient.

[0013] Optionally, the user's medication status estimation model is calibrated online based on the dynamic compliance coefficient to obtain the target medication status, including: obtaining the estimated amount of medication remaining for the user at historical time points; obtaining the user's medication purchase detection results at the current time point, wherein the medication purchase detection results are used to indicate whether the user has purchased medication at the current time point; determining the estimated amount of medication remaining for the user at the current time point based on the estimated amount of medication remaining for the user at historical time points, the dynamic compliance coefficient, the estimated daily medication consumption, the medication purchase detection results, a preset perturbation factor, and a preset noise factor; updating the user's continuous medication status sequence based on the estimated amount of medication remaining for the user at the current time point, and calibrating the user's medication status estimation model online based on the updated continuous medication status sequence and the estimated amount of medication remaining for the user at the current time point to obtain the target medication status.

[0014] Optionally, based on the user's historical drug remaining quantity estimate, dynamic compliance coefficient, estimated daily medication consumption, drug purchase detection results, preset disturbance factor, and preset noise factor, the estimated drug remaining quantity for the user at the current moment is determined, including: multiplying the dynamic compliance coefficient by the estimated daily medication consumption to obtain the compliance-adjusted daily consumption; determining the drug purchase replenishment quantity based on the drug purchase detection results, wherein when the drug purchase detection results indicate that the user should purchase medication at the current moment, the drug purchase replenishment quantity is equal to the quantity of medication purchased, and when the drug purchase detection results indicate that the user should not purchase medication at the current moment, the drug purchase replenishment quantity is zero; subtracting the compliance-adjusted daily consumption from the historical drug remaining quantity estimate, and adding the drug purchase replenishment quantity, preset disturbance factor, and preset noise factor, to obtain the estimated drug remaining quantity for the user at the current moment.

[0015] Optionally, after calibrating the user's medication status estimation model online based on the dynamic compliance coefficient to obtain the target medication status, the user's risk level is determined based on the target medication status. The risk level is used to characterize the probability of the user experiencing medication interruption risk. The prediction confidence level is determined based on the target medication status. The prediction confidence level is used to quantify the accuracy of the user's medication status estimation model in estimating the current medication status. A differentiated diversion scheduling strategy is determined based on the risk level and the prediction confidence level.

[0016] Optionally, a differentiated triage and scheduling strategy is determined based on the risk level and the predicted confidence level, including: when the detected risk level is lower than a first risk threshold and the predicted confidence level is higher than a first confidence threshold, the user's status is determined to be stable, and a standardized reminder from the automated intelligent interaction module is triggered; when the detected risk level is higher than a second risk threshold and the predicted confidence level is higher than a second confidence threshold, the user is determined to have a risk of medication shortage, and a manual telephone follow-up task is generated; when the detected risk level is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, or the predicted confidence level is lower than a third confidence threshold, the user's status is determined to be fluctuating, and the user is assigned to a manual online communication channel for follow-up, wherein both the first confidence threshold and the second confidence threshold are greater than the third confidence threshold, and the first confidence threshold and the second confidence threshold are equal or unequal.

[0017] According to another aspect of the embodiments of this application, a medication status calibration device based on interactive feedback is also provided, comprising: a collection unit for collecting interactive feedback data between a user and an interactive system, the interactive feedback data including text data and / or voice data; a parsing unit for performing semantic parsing on the interactive feedback data using a large language model to obtain feature information related to the user's actual medication status; a first processing unit for determining a dynamic compliance coefficient based on the feature information, the dynamic compliance coefficient being used to characterize the user's real-time execution of the medication plan; and a second processing unit for calibrating the user's medication status estimation model online based on the dynamic compliance coefficient to obtain a target medication status, wherein the target medication status includes an updated estimate of the remaining drug quantity, and the medication status estimation model is a mathematical model used to estimate the remaining drug quantity based on drug purchase behavior and medication compliance.

[0018] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described medication status calibration method based on interactive feedback.

[0019] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the above-described medication status calibration method based on interactive feedback.

[0020] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described medication status calibration method based on interactive feedback.

[0021] According to the technical solution of this application, interactive feedback data (including text data and / or voice data) between the user and the interactive system is collected, and semantic analysis is performed using a large language model to obtain feature information related to the user's actual medication status. Based on this, a dynamic compliance coefficient representing the user's real-time execution of the medication plan can be determined according to this feature information, and the medication status estimation model can be calibrated online based on this coefficient to obtain the target medication status including the updated estimate of the remaining drug quantity.

[0022] As can be seen from the above, firstly, by introducing user-initiated interactive feedback data, direct information reflecting the user's actual medication use can be obtained, rather than relying solely on the indirect indicator of medication purchase behavior. Secondly, using a large language model to perform semantic parsing of natural language interactions helps extract characteristic information such as medication adherence tendency from users' daily communication, providing a data foundation for subsequent calibration. Thirdly, the dynamic adherence coefficient calculated based on interactive feedback can reflect the user's degree of adherence to the medication plan in real time, thus helping to overcome the lag in medication purchase behavior data. Finally, online calibration of the medication status estimation model and updating the remaining drug quantity estimate using this dynamic adherence coefficient ensures that the target medication status output by the model can respond promptly to changes in the user's actual medication use behavior. The technical solution of this application at least solves the problem of distorted medication adherence assessment caused by existing technologies that rely solely on medication purchase behavior to infer medication status. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0024] Figure 1 This is a flowchart of an optional medication status calibration method based on interactive feedback according to an embodiment of this application;

[0025] Figure 2This is a flowchart of a medication status calibration system based on interactive feedback, provided according to an embodiment of this application.

[0026] Figure 3 This is a schematic diagram of an optional medication status calibration device based on interactive feedback according to an embodiment of this application. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] It should also be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) collected in this application are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with the relevant laws, regulations, and standards of the relevant regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, this system has interfaces with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface, and the relevant information is obtained only after receiving consent from the aforementioned user or organization.

[0030] According to an embodiment of this application, an embodiment of a medication status calibration method based on interactive feedback is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0031] According to the embodiments of this application, an interactive feedback-based medication status calibration system can be used as the execution subject of the interactive feedback-based medication status calibration method of this application embodiment. The system can be a software system or an embedded system combining software and hardware. Of course, the execution subject of the method in the embodiments of this application can also be other forms of execution subject, such as devices, equipment, etc. It should be known by those skilled in the art that this application does not particularly limit the specific form of the execution subject.

[0032] Figure 1 This is a flowchart of an optional medication status calibration method based on interactive feedback according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0033] Step S101: Collect interaction feedback data between the user and the interactive system. The interaction feedback data includes text data and / or voice data.

[0034] For example, interactive feedback data can be the original interactive content generated by users during proactive communication or passive follow-up with the chronic disease management platform, including text data and / or voice data. Text data originates from text messages entered by users through instant messaging channels (such as mini-program chat interfaces or application-embedded dialog boxes); voice data originates from recordings of calls between users and customer service personnel during telephone follow-ups or voice messages actively recorded by users. By collecting the aforementioned interactive feedback data, the medication status calibration system can obtain information on users' proactively disclosed medication experiences, symptom changes, and adherence.

[0035] In some embodiments, the medication status calibration system can acquire user interaction text data in real time through an instant messaging platform interface. For example, when using a chronic disease management application, a user might enter natural language messages such as "I've been feeling dizzy after taking my medication lately, and I'm a little hesitant to take it" or "I forgot to bring my medication on a business trip this week and missed two days" through the built-in chat window. The medication status calibration system receives these text messages in real time by calling the instant messaging platform's application programming interface and stores them as interactive feedback data in the system. This data collection method requires no additional user intervention and can collect feedback information imperceptibly during daily use, which helps improve the timeliness of data collection and user engagement.

[0036] In other embodiments, the medication status calibration system can also acquire voice data from telephone follow-ups via an outbound calling system and transcribe the voice data into text data. For example, when the system determines that a user has a potential risk of running out of medication based on medication purchase behavior data, it automatically triggers a manual telephone follow-up task. During the follow-up, the conversation between the user and customer service personnel is recorded in real time. After acquiring this voice data, the medication status calibration system calls a voice-to-text service to convert the call recording into text format. The converted text data has the same format as the text data collected through instant messaging channels, facilitating subsequent unified semantic analysis using a large language model. This data collection method is suitable for elderly users who have difficulty using text interaction or high-risk user scenarios that require in-depth communication, helping to expand the system's service coverage.

[0037] Step S102: Use a large language model to perform semantic parsing on the interactive feedback data to obtain feature information related to the user's actual medication status.

[0038] For example, a medication status calibration system can utilize a large language model to perform semantic analysis on the collected interactive feedback data to extract feature information related to the user's actual medication status. A large language model is a deep learning-based natural language processing model capable of understanding the semantic connotations of user-input natural language text and extracting structured key information. This feature information characterizes the user's current medication behavior and subjective feelings, including at least dimensions such as medication adherence tendency, symptom rating, risk score, and semantic confidence. Through the semantic understanding capabilities of the large language model, the medication status calibration system transforms unstructured user dialogue into structured quantitative indicators, providing a data foundation for subsequent adherence coefficient calculations.

[0039] In some embodiments, the medication status calibration system can use user text messages collected through instant messaging channels as input to a large language model for direct semantic parsing. For example, when a user inputs "My stomach always hurts after taking medication lately, and I don't really want to take it anymore," the large language model parses the medication adherence tendency as "decreased," the symptom rating as "moderate," the risk rating as "relatively high," and the semantic confidence as "high." When a user inputs "My medication is almost finished; I'll buy more tomorrow," the large language model parses the medication adherence tendency as "normal" and the risk rating as "low." The medication status calibration system can directly extract feature information from the text actively input by the user without additional manual annotation or rule configuration, thus facilitating rapid response to changes in the user's status.

[0040] In other embodiments, the medication status calibration system can also use the transcribed text data from telephone follow-ups as input to a large language model for semantic analysis. For example, a user might say in a telephone follow-up, "I've been too busy with work lately and always forget to take my medication, and the medication is a bit expensive." After the outbound calling system transcribes the call recording into text, the medication status calibration system inputs the transcribed text into the large language model. The model analyzes the medication adherence tendency as "decreased" (reason: forgetting), the symptom rating as "none," and the risk rating as "moderate," while also identifying economic stress as a potential influencing factor. This embodiment is suitable for scenarios where users are unable to input text. By combining speech transcription with large language model analysis, it helps to extract features from the voice interaction content.

[0041] Step S103: Determine the dynamic compliance coefficient based on the feature information. The dynamic compliance coefficient is used to characterize the user's real-time execution of the medication plan.

[0042] For example, the medication status calibration system can determine a dynamic compliance coefficient based on feature information obtained from semantic parsing. The dynamic compliance coefficient is a time-varying numerical indicator used to quantify the user's real-time adherence to the prescribed medication plan. This coefficient integrates multiple dimensions, including the user's expressed medication intention, symptom perception, and the model's confidence in the parsing results, reflecting the user's actual medication behavior tendencies over a recent period. Through this dynamic compliance coefficient, the medication status calibration system transforms the user's natural language feedback into a calculable quantitative indicator, providing a basis for subsequent model calibration.

[0043] In some embodiments, the medication status calibration system can determine the transient compliance coefficient based on medication adherence tendency, symptom grading score, and risk score from the feature information. For example, when the large language model parses the user message "I've been feeling dizzy lately after taking medication, so I reduced the dosage by half" and finds a medication adherence tendency of "decreased," a symptom grading score of "moderate," and a risk score of "high," the medication status calibration system obtains the weight coefficients corresponding to each dimension, calculates the weighted sum, and performs numerical transformation to obtain an transient compliance coefficient of 0.35. This value reflects the user's current low compliance status, and the medication status calibration system uses this value to determine that the user has a medication behavior deviation.

[0044] In other embodiments, the medication status calibration system can further integrate the user's historical average compliance baseline and the time interval between the current moment and the last effective interaction to obtain a dynamic compliance coefficient, based on the calculation of the instantaneous compliance coefficient. For example, a user's historical average compliance baseline is 0.75, the current instantaneous compliance coefficient is 0.35, and 15 days have passed since the last effective interaction. The medication status calibration system first calculates the time decay factor based on the effective confidence level and the time interval, then calculates the real-time information weight and the historical information weight respectively, and finally adds the weighted instantaneous compliance coefficient to the historical baseline to obtain a dynamic compliance coefficient of 0.52. This dynamic compliance coefficient reflects both the declining trend of compliance shown in the user's current feedback and the user's long-term historical performance, avoiding drastic fluctuations in status estimation due to a single abnormal feedback.

[0045] Step S104: The user's medication status estimation model is calibrated online based on the dynamic compliance coefficient to obtain the target medication status. The target medication status includes the updated estimated amount of remaining medication. The medication status estimation model is a mathematical model used to estimate the amount of remaining medication based on purchasing behavior and medication compliance.

[0046] For example, a medication status calibration system can calibrate a user's medication status estimation model online based on a dynamic adherence coefficient to obtain a target medication status. The medication status estimation model is a mathematical model used to estimate the remaining amount of medication based on purchasing behavior and medication adherence. This model continuously tracks the user's medication consumption and replenishment through a state update equation. The target medication status includes the updated estimated remaining amount of medication, reflecting the user's medication inventory status at the current moment. The medication status calibration system uses the dynamic adherence coefficient as model input and makes real-time corrections to the state estimation, making the model output closer to the user's actual medication usage.

[0047] In some embodiments, the medication status calibration system can obtain the estimated remaining amount of medication for a user at a historical time and the user's medication purchase detection result at the current time. The medication status calibration system multiplies the dynamic compliance coefficient by the estimated daily average medication consumption to obtain the compliance-adjusted daily average consumption, and subtracts this consumption from the historical estimated remaining amount of medication. If the medication purchase detection result indicates that the user has purchased medication, the medication status calibration system also includes the purchased quantity in the calculation, and simultaneously adds a preset perturbation factor and a preset noise factor to obtain the estimated remaining amount of medication for the user at the current time. The medication status calibration system updates the user's continuous medication status sequence based on this estimate and uses the updated sequence as part of the target medication status.

[0048] In other embodiments, the medication status calibration system updates the status when a user has not purchased medication. After adjusting the daily average consumption based on a dynamic compliance coefficient, the system subtracts this consumption from the historical remaining medication quantity. If the result is negative, it indicates that the user may be out of medication. The system uses this negative value as a basis for an out-of-medication warning and simultaneously calibrates the parameters of the medication status estimation model online based on the updated estimated remaining medication quantity. The system uses the updated continuous medication status sequence and the estimated remaining medication quantity together as the target medication status for subsequent risk assessment and triage / schedule decisions.

[0049] In some embodiments, collecting interaction feedback data between the user and the interactive system includes: obtaining the user's interactive text data through an instant messaging platform interface; and / or obtaining voice data from telephone follow-ups through an outbound calling system and transcribing the voice data into text data.

[0050] For example, the medication status calibration system can collect interaction feedback data between the user and the interactive system through two channels. The instant messaging platform interface refers to the data pathway between the medication status calibration system and a third-party instant messaging service (such as a mini-program chat function or an in-app message module), used to obtain text messages actively entered by the user. The outbound calling system refers to the connection module between the medication status calibration system and a telephone call center or automated voice outbound calling service, used to obtain call recordings during telephone follow-ups. The medication status calibration system can use one of these channels alone or both channels simultaneously to collect data, adapting to the usage habits of different user groups and the communication needs of different scenarios.

[0051] In some embodiments, a large language model is used to perform semantic parsing on the interactive feedback data to obtain feature information related to the user's actual medication status. This includes: using the interactive feedback data as model input data, performing semantic parsing on the large language model to obtain the parsing result; extracting feature information from the parsing result, the feature information being used to characterize the user's current medication status, and the feature information including at least the following dimensions: medication adherence tendency, symptom grading score, risk score, and semantic confidence.

[0052] For example, a medication status calibration system can utilize a large language model to perform semantic parsing on the collected interactive feedback data to extract structured feature information related to the user's actual medication status. A large language model is a natural language processing model based on deep neural networks that can understand the semantic connotations of user input text and output quantitative analysis results.

[0053] In some embodiments, the medication status calibration system can directly input user text messages collected through instant messaging channels into a large language model for semantic parsing. After performing word segmentation, semantic understanding, and sentiment analysis on the text, the large language model outputs structured parsing results. The medication status calibration system extracts feature information from the four dimensions mentioned above from the parsing results. The system combines these dimensions into a feature vector for subsequent compliance coefficient calculation. This method eliminates the need for manual annotation or rule configuration, facilitating the automated extraction of key information from unstructured text.

[0054] In other embodiments, the medication status calibration system inputs the transcribed text data from telephone follow-ups into a large language model for semantic parsing. The transcribed text typically contains longer dialogues and more complex semantic information, such as multi-turn question-and-answer exchanges between the user and customer service personnel. The medication status calibration system uses the complete transcribed text as input, from which the large language model identifies key information points expressed by the user. The system extracts features such as medication adherence tendency, symptom rating, risk score, and semantic confidence from the parsing results. For potential speech recognition errors in the transcribed text, the large language model can correct them by incorporating contextual semantics, ensuring the robustness of feature extraction. The extracted features are then used as input data for subsequent calculations of dynamic compliance coefficients.

[0055] In some embodiments, before using a large language model to perform semantic parsing on the interaction feedback data, the method further includes: obtaining the interaction text sequence between the user and the interaction system; recalling similar cases in a pre-built historical case library based on the interaction text sequence; and filling the context of the large language model with at least one recalled target historical case, wherein each historical case in the historical case library is represented as a combination of historical interaction text and a corresponding business vector.

[0056] For example, before using a large language model for semantic parsing, the medication status calibration system performs a similar case recall step to enhance the parsing effect. The historical case library is a pre-built dataset storing historical interaction records and their corresponding annotation results. Each historical case is represented as a combination of historical interaction text and a corresponding business vector. The historical interaction text records the user's past feedback, while the corresponding business vector contains labeled features such as medication adherence tendency, symptom rating, and risk rating from that interaction. By recalling the historical cases most similar to the current user, the medication status calibration system uses the business vectors from those cases as prior knowledge to populate the context of the large language model, guiding the model to more accurately parse the current interaction content.

[0057] In other embodiments, the medication status calibration system fills the context of the large language model with multiple recalled similar cases. The system sets a similarity threshold, recalls all historical cases with similarity greater than the threshold, sorts them by similarity from high to low, and then sequentially fills the business vectors of the top-ranked historical cases into the context of the large language model. For example, when the current user's interaction text is short or semantically ambiguous, the reference value of a single similar case is limited. By introducing comprehensive information from multiple similar cases, the medication status calibration system provides the large language model with richer prior knowledge, thereby improving the accuracy and robustness of semantic parsing. The system fills the context of the large language model with at least one recalled target historical case before performing subsequent semantic parsing operations.

[0058] In some embodiments, determining the dynamic compliance coefficient based on feature information includes: obtaining semantic confidence from the feature information, where semantic confidence is used to characterize the degree of confidence of the large language model in the semantic parsing results; obtaining the most similar case recalled from the historical case library based on the interaction text sequence between the user and the interaction system; updating the semantic confidence based on the similarity between the most similar case and the interaction text sequence to obtain an effective confidence, wherein the effective confidence is equal to the square root of the product of semantic confidence and similarity; determining the instantaneous compliance coefficient based on medication adherence tendency, symptom rating score, and risk score from the feature information; obtaining the user's historical average compliance baseline value and the time interval between the current moment and the last effective text interaction; and determining the dynamic compliance coefficient with time decay based on the effective confidence, time interval, instantaneous compliance coefficient, and historical average compliance baseline value.

[0059] For example, the historical average adherence baseline reflects the user's long-term average adherence level. The time interval is used to measure the timeliness of the current interaction feedback. By integrating the above factors, the medication status calibration system obtains a dynamic adherence coefficient that comprehensively reflects the user's real-time medication adherence.

[0060] In some embodiments, when the most similar case is highly similar to the current interactive text, the effective confidence level is close to the semantic confidence level; when the difference between the two is large, the effective confidence level decreases significantly. The medication status calibration system obtains the weight coefficients corresponding to each dimension based on the medication adherence tendency, symptom grading score, and risk score in the feature information, calculates the weighted sum, and performs a soft-addition function transformation to obtain an instantaneous compliance coefficient greater than zero.

[0061] In other embodiments, the medication status calibration system determines a dynamic compliance coefficient with fusion time decay based on effective confidence, time interval, instantaneous compliance coefficient, and historical average compliance benchmark. The medication status calibration system first calculates a time decay factor, which decays exponentially with increasing time interval. The system then calculates the product of the effective confidence and the time decay factor to obtain a first weighting coefficient, which reflects the reliability of the real-time interactive feedback. The system uses this first weighting coefficient as the weight for real-time information, subtracts the first weighting coefficient from 1 as the weight for historical information, multiplies this by the instantaneous compliance coefficient and the historical average compliance benchmark, and sums the results to obtain the dynamic compliance coefficient. When the effective confidence is high and the time interval is short, the dynamic compliance coefficient mainly relies on the real-time interactive feedback; when the effective confidence is low or the time interval is long, the dynamic compliance coefficient relies more on the historical benchmark.

[0062] In some embodiments, determining the transient compliance coefficient based on medication adherence tendency, symptom grading score, and risk score in the feature information includes: obtaining the weight coefficients corresponding to each of the medication adherence tendency, symptom grading score, and risk score in the feature information; calculating the sum of the products of the medication adherence tendency, symptom grading score, and risk score with their respective weight coefficients, and adding a bias term to obtain a weighted sum; inputting the weighted sum into a soft-addition function for numerical transformation to obtain a transient compliance coefficient greater than zero, wherein the soft-addition function is used to map any real number to a positive real number.

[0063] In some embodiments, the medication status calibration system obtains the weight coefficients corresponding to each dimension and then calculates a weighted sum. For example, the weight coefficient for medication adherence tendency is set to 0.5, the weight coefficient for symptom grading score is set to 0.3, and the weight coefficient for risk score is set to 0.2. The medication status calibration system multiplies the medication adherence tendency value by its weight, adds the symptom grading score multiplied by its weight, adds the risk score multiplied by its weight, and finally adds a bias term to obtain the weighted sum. If the weighted sum is negative, directly using it as input to the soft-addition function will result in a positive number close to zero; if the weighted sum is positive, the soft-addition function output is approximately equal to that positive value. The medication status calibration system ensures that the output instantaneous compliance coefficient is always greater than zero through the soft-addition function transformation.

[0064] In other embodiments, the medication status calibration system dynamically adjusts the weighting coefficients based on different business scenarios. When the system focuses on identifying early signs of declining adherence, the system increases the weighting coefficient of medication adherence tendency, making this dimension dominant in the weighted sum. When the system focuses on assessing the impact of the user's health status on adherence, the system increases the weighting coefficient of symptom rating. The system can also configure differentiated combinations of weighting coefficients based on the user's disease type or medication category. After inputting the calculated weighted sum into a soft summing function, the system obtains an instantaneous adherence coefficient greater than zero; a higher coefficient value indicates a higher current level of user adherence.

[0065] In some embodiments, determining a dynamic compliance coefficient with time decay based on effective confidence level, time interval, instantaneous compliance coefficient, and historical average compliance benchmark value includes: obtaining a time decay factor, wherein the time decay factor is equal to the negative value of the natural constant multiplied by a preset time decay constant and then multiplied by the exponent of the time interval; calculating the product of the effective confidence level and the time decay factor to obtain a first weighting coefficient; determining a second weighting coefficient based on the first weighting coefficient, wherein the second weighting coefficient is equal to 1 minus the first weighting coefficient; calculating the product of the first weighting coefficient and the instantaneous compliance coefficient to obtain a first product value; calculating the product of the second weighting coefficient and the historical average compliance benchmark value to obtain a second product value; and adding the first product and the second product to obtain the dynamic compliance coefficient.

[0066] For example, the medication status calibration system determines the dynamic compliance coefficient through a weighted fusion method based on effective confidence level, time interval, instantaneous compliance coefficient, and historical average compliance benchmark value. The time decay factor is an exponential decay amount calculated based on the time interval, whose value approaches zero as the time interval increases, and is used to characterize the degree of time-sensitivity decay of historical interaction information. The first weighting coefficient is the product of the effective confidence level and the time decay factor, used to quantify the weight proportion of real-time interaction feedback in the fusion calculation. The second weighting coefficient is equal to 1 minus the first weighting coefficient, used to quantify the weight proportion of historical benchmark information in the fusion calculation. Through the above weighted fusion mechanism, the medication status calibration system dynamically balances real-time feedback and historical information.

[0067] In some embodiments, after obtaining the time decay factor, the medication status calibration system calculates a first weighting coefficient and a second weighting coefficient. The system calculates a first product value by multiplying the first weighting coefficient by the instantaneous compliance coefficient, and a second product value by multiplying the second weighting coefficient by the historical average compliance baseline value. The first and second product values ​​are then added together to obtain the dynamic compliance coefficient. When the time interval between the user's last effective text interaction is short, the time decay factor is close to 1, the first weighting coefficient is dominated by the effective confidence level, and the dynamic compliance coefficient mainly depends on the instantaneous compliance coefficient. The medication status calibration system prioritizes trusting the user's real-time interactive feedback. When the time interval is long, the time decay factor approaches 0, the first weighting coefficient approaches 0, and the second weighting coefficient approaches 1. The dynamic compliance coefficient is mainly determined by the historical average compliance baseline value, and the medication status calibration system relies more on the user's historical performance.

[0068] In other embodiments, the medication status calibration system adjusts the fusion weights based on different values ​​of effective confidence. When the large language model has a high degree of confidence in the semantic parsing results, the effective confidence is close to 1, the first weighting coefficient is dominated by the time decay factor, and the weight of real-time interactive feedback is high. When the semantic parsing results are ambiguous or unclear, the effective confidence is low, the first weighting coefficient is reduced accordingly, and the medication status calibration system automatically reduces the weight of real-time interactive feedback, relying more on the historical average compliance baseline value for estimation. Through the above mechanism, the medication status calibration system can quickly respond to changes in user status when the confidence of interactive feedback is high, and maintain the stability of status estimation when the confidence of interactive feedback is low, avoiding drastic fluctuations in compliance coefficients due to a single abnormal feedback.

[0069] In some embodiments, the user's medication status estimation model is calibrated online based on the dynamic compliance coefficient to obtain the target medication status, including: obtaining the user's estimated remaining drug quantity at historical time points; obtaining the user's drug purchase detection result at the current time point, wherein the drug purchase detection result is used to indicate whether the user has purchased drugs at the current time point; determining the user's estimated remaining drug quantity at the current time point based on the user's estimated remaining drug quantity at historical time points, the dynamic compliance coefficient, the estimated daily average drug consumption, the drug purchase detection result, a preset perturbation factor, and a preset noise factor; updating the user's continuous medication status sequence based on the user's estimated remaining drug quantity at the current time point, and calibrating the user's medication status estimation model online based on the updated continuous medication status sequence and the user's estimated remaining drug quantity at the current time point to obtain the target medication status.

[0070] For example, the medication status calibration system calibrates the user's medication status estimation model online based on the dynamic compliance coefficient to update the user's estimated remaining medication quantity. The continuous medication status sequence is a data sequence formed by arranging the estimated remaining medication quantity at different times in chronological order, reflecting the dynamic changes in the user's medication inventory. The preset disturbance factor is an adjustment term added by the system based on actual conditions, used to simulate potential regular deviations in user medication behavior. The preset noise factor is a random error term set by the system to characterize the uncertainty in the status estimation process. Through comprehensive calculation of the above multiple factors, the medication status calibration system obtains the estimated remaining medication quantity at the current time and calibrates the model online.

[0071] In some embodiments, the medication status calibration system obtains the estimated remaining amount of medication for a user at a historical time and the user's medication purchase detection result at the current time. The system multiplies the dynamic compliance coefficient by the estimated daily average medication consumption to obtain the compliance-adjusted daily average consumption. If the medication purchase detection result indicates that the user purchased medication at the current time, the system includes the purchased quantity as part of the medication replenishment amount in the calculation; if the result indicates that the user did not purchase medication, the replenishment amount is zero. The system subtracts the compliance-adjusted daily average consumption from the estimated remaining amount of medication at a historical time, and then adds the replenishment amount, a preset perturbation factor, and a preset noise factor to obtain the estimated remaining amount of medication for the user at the current time. The system updates the medication continuous status sequence based on this estimate.

[0072] In other embodiments, the medication status calibration system calibrates the medication status estimation model online based on the updated continuous medication status sequence. The system appends the user's current remaining medication quantity estimate as a new element to the end of the continuous medication status sequence, forming an updated sequence. The updated continuous medication status sequence and the user's current remaining medication quantity estimate are considered important components of the target medication status. The system uses the remaining medication quantity estimate in the target medication status to correct the model parameters, enabling the model to better fit the user's actual medication behavior. The system outputs the target medication status to the triage and scheduling module for subsequent risk assessment and intervention strategy development.

[0073] In some embodiments, determining the estimated remaining amount of medication for a user at the current moment based on the user's historical medication balance estimate, dynamic compliance coefficient, estimated daily medication consumption, medication purchase monitoring results, a preset perturbation factor, and a preset noise factor includes: multiplying the dynamic compliance coefficient by the estimated daily medication consumption to obtain a compliance-adjusted daily consumption; determining the medication replenishment amount based on the medication purchase monitoring results, wherein the medication replenishment amount equals the medication quantity when the medication purchase monitoring results indicate that the user will purchase medication at the current moment, and the medication replenishment amount is zero when the medication purchase monitoring results indicate that the user will not purchase medication at the current moment; and subtracting the compliance-adjusted daily consumption from the historical medication balance estimate, and adding the medication replenishment amount, the preset perturbation factor, and the preset noise factor to obtain the user's estimated remaining amount of medication at the current moment.

[0074] For example, the medication status calibration system calculates the estimated remaining amount of medication for the user at the current moment using a state update equation based on historical estimates of remaining medication, dynamic compliance coefficients, estimated daily medication consumption, purchase detection results, preset disturbance factors, and preset noise factors. The estimated daily medication consumption is the daily recommended dosage determined according to the drug instructions or doctor's orders. The compliance-adjusted daily consumption is the product of the dynamic compliance coefficient and the estimated daily medication consumption, reflecting the deviation between the user's actual medication use and the doctor's orders. The purchase replenishment quantity is the amount of medication to be replenished based on purchase detection results; it equals the purchase quantity when the user purchases medication, and is zero otherwise. Through the above calculations, the medication status calibration system achieves dynamic updates to the estimated remaining medication amount.

[0075] In some embodiments, the medication status calibration system estimates the remaining medication amount when the user has not purchased any medication. The system multiplies the dynamic compliance coefficient by the estimated daily average medication consumption to obtain the compliance-adjusted daily average consumption. The system then subtracts the compliance-adjusted daily average consumption from the historical remaining medication amount estimate, and adds a preset perturbation factor and a preset noise factor to obtain the user's estimated remaining medication amount at the current moment. When the dynamic compliance coefficient is low, the compliance-adjusted daily average consumption decreases accordingly, and the rate of decrease in remaining medication amount slows down, reflecting a situation where the user's actual medication consumption is lower than the prescribed dosage. The system uses this estimated value as the remaining medication amount at the current moment.

[0076] In other embodiments, the medication status calibration system estimates the remaining amount of medication in scenarios where a user has purchased medication. The system first calculates the remaining amount after consumption based on a scenario where no medication has been purchased, and then adds the purchased quantity as a replenishment amount to the calculation. The system adds the remaining amount after consumption to the replenishment amount, a preset disturbance factor, and a preset noise factor to obtain an estimate of the user's remaining medication amount at the current moment. When the user purchases a large quantity of medication, the estimated remaining amount may show a significant jump. The system uses this estimate as the remaining amount of medication at the current moment and updates the continuous medication status sequence. Through this mechanism, the medication status calibration system can simultaneously track the user's medication consumption and purchase behavior, achieving continuous dynamic monitoring of medication inventory.

[0077] Optionally, after calibrating the user's medication status estimation model online based on the dynamic compliance coefficient to obtain the target medication status, the user's risk level is determined based on the target medication status. The risk level is used to characterize the probability of the user experiencing medication interruption risk. The prediction confidence level is determined based on the target medication status. The prediction confidence level is used to quantify the accuracy of the user's medication status estimation model in estimating the current medication status. A differentiated diversion scheduling strategy is determined based on the risk level and the prediction confidence level.

[0078] For example, the medication status calibration system determines a user's risk level and prediction confidence based on the estimated remaining amount of medication in the target medication status and the continuous medication status sequence. The risk level characterizes the probability of a user experiencing medication shortage; the lower the remaining amount and the faster the consumption rate, the higher the risk level. The prediction confidence quantifies the accuracy of the medication status estimation model's estimation of the current medication status; its value is derived from the error covariance during the status update process. A smaller error covariance indicates a more reliable model estimation. The medication status calibration system comprehensively evaluates the user's status through these two dimensions, providing a decision-making basis for subsequent triage and scheduling.

[0079] In some embodiments, the medication status calibration system calculates a risk level based on the target medication status and determines a diversion and scheduling strategy based on the risk level. The system compares the estimated remaining drug quantity with a preset remaining quantity threshold, matches the consumption trend in the continuous medication status sequence with a preset consumption pattern, and comprehensively calculates the risk level. When the risk level is low, it indicates that the user's drug inventory is sufficient and consumption is regular; the system determines the user's status is stable and triggers a standardized reminder from the automated intelligent interaction module. When the risk level is high, it indicates that the user's drug inventory is insufficient or consumption is abnormal; the system determines that the user is at risk of running out of medication and generates a manual telephone follow-up task.

[0080] In other embodiments, the medication status calibration system determines the triage strategy based on both risk level and predicted confidence level. The system extracts the error covariance from the status update process and converts it into predicted confidence level. When the predicted confidence level is high, it indicates that the model's estimation of the current state is relatively reliable, and the system directly makes triage decisions based on the risk level. When the predicted confidence level is low, it indicates that the model estimation has significant uncertainty, and the system does not directly make decisions based on the risk level. Instead, it assigns users to a human online communication channel for follow-up, further confirming the user's actual medication status manually. Based on the combination of risk level and predicted confidence level, the system assigns users to the corresponding intervention channel among the automated intelligent interaction module, the human online communication channel, or the human telephone follow-up channel.

[0081] Optionally, a differentiated triage and scheduling strategy is determined based on the risk level and the predicted confidence level, including: when the detected risk level is lower than a first risk threshold and the predicted confidence level is higher than a first confidence threshold, the user's status is determined to be stable, and a standardized reminder from the automated intelligent interaction module is triggered; when the detected risk level is higher than a second risk threshold and the predicted confidence level is higher than a second confidence threshold, the user is determined to have a risk of medication shortage, and a manual telephone follow-up task is generated; when the detected risk level is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, or the predicted confidence level is lower than a third confidence threshold, the user's status is determined to be fluctuating, and the user is assigned to a manual online communication channel for follow-up, wherein both the first confidence threshold and the second confidence threshold are greater than the third confidence threshold, and the first confidence threshold and the second confidence threshold are equal or unequal.

[0082] For example, the medication status calibration system triages users to different intervention channels based on a combination of risk level and predicted confidence level. A first risk threshold is the boundary value used to distinguish between low and medium risk, and a second risk threshold is the boundary value used to distinguish between medium and high risk, with the second risk threshold being greater than the first risk threshold. A first confidence threshold is the minimum confidence requirement required for a stable state, a second confidence threshold is the confidence requirement required for high-risk management, and a third confidence threshold is the upper limit of confidence for a fluctuating state. The medication status calibration system achieves accurate user triage through the joint judgment of these two indicators.

[0083] In some embodiments, the medication status calibration system determines that the user's status is stable when the risk level is below a first risk threshold and the prediction confidence level is above the first confidence threshold. In this case, the user has sufficient medication inventory, consumes medication regularly, and the model's status estimation is relatively reliable, requiring no manual intervention. The medication status calibration system triggers standardized reminders from the automated intelligent interaction module, such as sending regular medication reminder messages to the user via an instant messaging platform. When the risk level is above a second risk threshold and the prediction confidence level is above the second confidence threshold, the medication status calibration system determines that the user is at risk of running out of medication. In this case, the user's medication inventory is insufficient, and the model's judgment is reliable, requiring deep manual intervention. The medication status calibration system generates a manual telephone follow-up task, where customer service personnel communicate with the user by phone to understand the specific reasons and provide intervention support.

[0084] In other embodiments, when the risk level is in the middle range (greater than or equal to the first risk threshold and less than or equal to the second risk threshold), the medication status calibration system determines that the user's status is fluctuating, regardless of the prediction confidence level. The system then assigns the user to a human online communication channel for follow-up, where customer service personnel communicate with the user via text or voice through an instant messaging platform to confirm the medication status. When the prediction confidence level is lower than the third confidence threshold, the system also determines that the user's status is fluctuating, regardless of the risk level, and assigns the user to a human online communication channel. Since the third confidence threshold is lower than the first and second confidence thresholds, this condition covers scenarios with high model estimation uncertainty. Through this three-level triage mechanism, the medication status calibration system directs stable users to automated channels, fluctuating users to online human channels, and high-risk users to telephone human channels, thus optimizing the allocation of service resources.

[0085] For example, Figure 2 This is a flowchart of a medication status calibration system based on interactive feedback, provided according to an embodiment of this application. Figure 2 As shown, the medication status calibration system includes a data acquisition and feature analysis module, a feature quantification module, a medication adherence calibration and inference module, and a diversion scheduling module. These modules are connected in sequence to form a closed-loop feedback loop.

[0086] Optionally, the data acquisition and feature parsing module in the medication status calibration system first acquires the user's interactive text data through an instant messaging platform interface and obtains the voice data from telephone follow-ups through an outbound calling system, then transcribes the voice data into text data. This module takes the interactive text data and the transcribed text as input and uses a large language model to perform semantic parsing on the unstructured data, extracting features related to the user's actual medication feedback. During the parsing process, this module acquires the current user's interactive text sequence and recalls similar cases from a pre-built historical case library. Each historical case is represented as a combination of historical interactive text and a corresponding business vector. This module fills the context of the large language model with the recalled content to improve output accuracy. The parsing result vector output by this module... It should include at least the following dimensions: medication adherence tendency, symptom severity, risk score, and semantic confidence.

[0087] Optionally, the feature quantization module in the medication status calibration system receives the feature vectors output by the data acquisition and feature parsing modules. This is transformed into a correction signal driven by real-time interaction, outputting a dynamic compliance coefficient. For example, this module first calculates the effective confidence level. Using similarity to assess the confidence level of the original semantics Perform calibration, i.e. ,in To determine the cosine similarity between the most similar historical cases recalled, the original confidence level is constrained by external prior knowledge. Then, the module calculates the instantaneous compliance coefficient. ,in These are the components of each dimension in the feature vector (including medication adherence tendency, symptom severity, and risk score). Here are the weight coefficients for each dimension, and b is the bias term. Perform numerical transformation to ensure This aligns with the physical meaning of the compliance coefficient. Finally, the module integrates time decay to obtain the final dynamic compliance coefficient. ,in This serves as a long-term historical average compliance benchmark for users. This represents the time interval between the current moment and the last valid text interaction. The preset time decay constant, This is a time decay factor. When both confidence and timeliness are high, the medication status calibration system prioritizes trusting real-time user interaction information. This dynamic compliance coefficient... The subsequent state update equation is input to compensate for the estimation bias caused by the lag in external observations (such as drug purchase records).

[0088] Optionally, the medication adherence calibration and inference module in the medication status calibration system receives the dynamic adherence coefficient output by the feature quantization module. The remaining amount of medicine is continuously estimated using a state update equation. The state update equation is as follows: ,in This is an estimate of the remaining amount of medicine on day t. This is an estimated daily dosage. This represents the average daily consumption based on the compliance coefficient. This is the result of the drug purchase test (a value of 1 indicates that the drug was purchased, and 0 indicates that the drug was not purchased). For the quantity of medicines purchased, For the observed amount of medicine purchased and replenished, This represents the system noise term.

[0089] Optionally, the triage scheduling module in the medication status calibration system is based on the hidden state probability distribution and estimated drug remaining quantity output by the medication adherence calibration and inference module. The system calculates the risk level R, with a higher risk level indicating a greater likelihood of medication interruption for the user. This module also calculates the prediction confidence score C, used to quantify the reliability of the model's current inference. This prediction confidence score is derived from the error covariance during the state update process; a smaller error covariance indicates a more reliable model estimate and a higher prediction confidence score, reflecting the accuracy of the state update. Based on the user's risk level and prediction confidence score, the medication status calibration system implements a differentiated resource allocation strategy: for users with stable status and low risk, standardized reminders from the intelligent interaction module are triggered; for users with abnormally fluctuating status or high risk, a manual telephone follow-up task is automatically generated.

[0090] After the new feedback data generated by the intervention is produced, the medication status calibration system inputs the new feedback data back into the data acquisition and feature analysis module to start a new round of model calibration, forming a closed-loop feedback mechanism for continuous optimization.

[0091] According to another aspect of the embodiments of this application, Figure 3 This is a schematic diagram of an optional medication status calibration device based on interactive feedback according to an embodiment of this application, such as... Figure 3As shown, the medication status calibration device based on interactive feedback includes: a data acquisition unit 301, used to acquire interactive feedback data between the user and the interactive system, the interactive feedback data including text data and / or voice data; a parsing unit 302, used to perform semantic parsing on the interactive feedback data using a large language model to obtain feature information related to the user's actual medication status; a first processing unit 303, used to determine a dynamic compliance coefficient based on the feature information, the dynamic compliance coefficient being used to characterize the user's real-time execution of the medication plan; and a second processing unit 304, used to calibrate the user's medication status estimation model online based on the dynamic compliance coefficient to obtain a target medication status, wherein the target medication status includes an updated estimate of the remaining drug quantity, and the medication status estimation model is a mathematical model used to estimate the remaining drug quantity based on drug purchase behavior and medication compliance.

[0092] Optionally, the acquisition unit 301 includes: a data acquisition subunit, used to acquire user interactive text data through an instant messaging platform interface; and / or, acquire voice data of telephone follow-up through an outbound calling system, and transcribe the voice data into text data.

[0093] Optionally, the parsing unit 302 includes: a parsing subunit, used to use interactive feedback data as model input data, perform semantic parsing using a large language model, and obtain parsing results; and a feature extraction subunit, used to extract feature information from the parsing results. The feature information is used to characterize the user's current medication status, and the feature information includes at least the following dimensions: medication adherence tendency, symptom grading score, risk score, and semantic confidence.

[0094] Optionally, the medication status calibration device based on interactive feedback further includes: an acquisition unit for acquiring the interactive text sequence between the user and the interactive system; and a recall unit for recalling similar cases in a pre-built historical case library based on the interactive text sequence, and filling at least one target historical case recalled into the context of the large language model, wherein each historical case in the historical case library is represented as a combination of historical interactive text and corresponding business vector.

[0095] Optionally, the first processing unit 303 includes: a first acquisition subunit for acquiring semantic confidence in the feature information, wherein semantic confidence is used to characterize the degree of confidence of the large language model in the semantic parsing results; a second acquisition subunit for acquiring the most similar case recalled from the historical case library based on the interaction text sequence between the user and the interaction system; an update subunit for updating the semantic confidence based on the similarity between the most similar case and the interaction text sequence to obtain an effective confidence, wherein the effective confidence is equal to the square root of the product of semantic confidence and similarity; a first determination subunit for determining the instantaneous compliance coefficient based on the medication adherence tendency, symptom grading score, and risk score in the feature information; a third acquisition subunit for acquiring the user's historical average compliance baseline value and the time interval between the current moment and the last effective text interaction; and a second determination subunit for determining the dynamic compliance coefficient with fusion time decay based on the effective confidence, time interval, instantaneous compliance coefficient, and historical average compliance baseline value.

[0096] Optionally, the first determining subunit includes: a first acquisition submodule, used to acquire the weight coefficients corresponding to medication adherence tendency, symptom grading score, and risk score in the feature information; a first calculation submodule, used to calculate the sum of the products of medication adherence tendency, symptom grading score, and risk score with their respective weight coefficients, and add a bias term to obtain a weighted sum; and a second calculation submodule, used to input the weighted sum into a soft addition function for numerical transformation to obtain an instantaneous compliance coefficient greater than zero, wherein the soft addition function is used to map any real number to a positive real number.

[0097] Optionally, the second determining subunit includes: a second acquisition submodule for acquiring a time decay factor, wherein the time decay factor is equal to the negative value of the natural constant multiplied by a preset time decay constant and then multiplied by an exponent of the time interval; a third calculation submodule for calculating the product of the effective confidence level and the time decay factor to obtain a first weighting coefficient; a fourth calculation submodule for determining a second weighting coefficient based on the first weighting coefficient, wherein the second weighting coefficient is equal to 1 minus the first weighting coefficient; a fifth calculation submodule for calculating the product of the first weighting coefficient and the instantaneous compliance coefficient to obtain a first product value; a sixth calculation submodule for calculating the product of the second weighting coefficient and the historical average compliance benchmark value to obtain a second product value; and a seventh calculation submodule for adding the first product and the second product to obtain a dynamic compliance coefficient.

[0098] Optionally, the second processing unit 304 includes: a second acquisition subunit, used to acquire the estimated amount of medicine remaining for the user at a historical time; and to acquire the user's purchase detection result at the current time, wherein the purchase detection result is used to indicate whether the user has purchased medicine at the current time; a third determination subunit, used to determine the estimated amount of medicine remaining for the user at the current time based on the estimated amount of medicine remaining for the user at a historical time, dynamic compliance coefficient, estimated daily medication consumption, purchase detection result, preset disturbance factor, and preset noise factor; and a fourth determination subunit, used to update the user's continuous medication state sequence based on the estimated amount of medicine remaining for the user at the current time, and to calibrate the user's medication state estimation model online based on the updated continuous medication state sequence and the estimated amount of medicine remaining for the user at the current time to obtain the target medication state.

[0099] Optionally, the fourth determining subunit includes: an eighth calculation submodule, used to multiply the dynamic compliance coefficient by the estimated daily average medication consumption to obtain the compliance-adjusted daily average consumption; a ninth calculation submodule, used to determine the medication replenishment quantity based on the medication purchase test results, wherein when the medication purchase test results indicate that the user should purchase medication at the current time, the medication replenishment quantity is equal to the medication quantity purchased, and when the medication purchase test results indicate that the user has not purchased medication at the current time, the medication replenishment quantity is zero; and a tenth calculation submodule, used to subtract the compliance-adjusted daily average consumption from the estimated remaining medication quantity at historical time points, and add the medication replenishment quantity, a preset disturbance factor, and a preset noise factor to obtain the estimated remaining medication quantity for the user at the current time.

[0100] Optionally, the medication status calibration device based on interactive feedback further includes: a first determining unit, used to determine the user's risk level according to the target medication status, the risk level being used to characterize the probability of the user experiencing medication interruption risk; a second determining unit, used to determine the prediction confidence level according to the target medication status, the prediction confidence level being used to quantify the accuracy of the user's medication status estimation model in estimating the current medication status; and a third determining unit, used to determine a differentiated diversion scheduling strategy according to the risk level and the prediction confidence level.

[0101] Optionally, the third determining unit includes: a first determining subunit, used to determine that the user's state is stable and trigger a standardized reminder from the automated intelligent interaction module when the detected risk level is lower than a first risk threshold and the predicted confidence level is higher than a first confidence threshold; a second determining subunit, used to determine that the user has a risk of medication interruption and generate a manual telephone follow-up task when the detected risk level is higher than a second risk threshold and the predicted confidence level is higher than a second confidence threshold; and a third determining subunit, used to determine that the user's state is fluctuating and assign the user to a manual online communication channel for follow-up when the detected risk level is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, or when the predicted confidence level is lower than a third confidence threshold. The first confidence threshold and the second confidence threshold are both greater than the third confidence threshold, and the first confidence threshold and the second confidence threshold may be equal or unequal.

[0102] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described medication status calibration method based on interactive feedback.

[0103] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the above-described medication status calibration method based on interactive feedback.

[0104] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described medication status calibration method based on interactive feedback.

[0105] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0106] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0107] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0108] The units described as separate components may or may not be physically separate. 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 units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0109] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0110] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0111] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for calibrating medication status based on interactive feedback, characterized in that, include: Collect interactive feedback data between the user and the interactive system, including text data and / or voice data; The interactive feedback data is semantically parsed using a large language model to obtain feature information related to the user's actual medication status. Determining a dynamic compliance coefficient based on the aforementioned feature information, the dynamic compliance coefficient characterizing the user's real-time adherence to the medication plan includes: obtaining semantic confidence from the feature information, the semantic confidence characterizing the degree of confidence of the large language model in the semantic parsing results; obtaining the most similar case recalled from the historical case library based on the interaction text sequence between the user and the interaction system; updating the semantic confidence based on the similarity between the most similar case and the interaction text sequence to obtain an effective confidence, wherein the effective confidence is equal to the square root of the product of the semantic confidence and the similarity; determining an instantaneous compliance coefficient based on the medication adherence tendency, symptom grading score, and risk score in the feature information; obtaining the user's historical average compliance baseline value and the time interval between the current moment and the last effective text interaction; and determining a dynamic compliance coefficient with time decay based on the effective confidence, the time interval, the instantaneous compliance coefficient, and the historical average compliance baseline value. The target medication status is obtained by calibrating the user's medication status estimation model online based on the dynamic compliance coefficient. The target medication status includes the updated estimated amount of remaining medication. The medication status estimation model is a mathematical model used to estimate the amount of remaining medication based on purchasing behavior and medication compliance.

2. The method according to claim 1, characterized in that, Collect user interaction feedback data with the interactive system, including: Obtain user interaction text data through instant messaging platform interfaces; and / or, The system obtains voice data from telephone follow-ups through an outbound calling system and transcribes the voice data into text data.

3. The method according to claim 1, characterized in that, The interactive feedback data is semantically parsed using a large language model to obtain feature information related to the user's actual medication status, including: The interactive feedback data is used as input data for the model, and semantic parsing is performed using a large language model to obtain the parsing results; The feature information is extracted from the parsing results. The feature information is used to characterize the user's current medication status. The feature information includes at least the following dimensions: medication adherence tendency, symptom grading score, risk score, and semantic confidence.

4. The method according to claim 1, characterized in that, Before using a large language model to perform semantic parsing on the interactive feedback data, the method further includes: Obtain the text sequence of interactions between the user and the interaction system; Based on the interactive text sequence, similar cases are recalled in a pre-built historical case library, and at least one target historical case is populated into the context of the large language model. Each historical case in the historical case library is represented as a combination of historical interactive text and corresponding business vector.

5. The method according to claim 1, characterized in that, Based on the medication adherence tendency, symptom severity score, and risk score in the aforementioned feature information, the transient compliance coefficient is determined, including: Obtain the weight coefficients corresponding to medication adherence tendency, symptom grading score, and risk score from the feature information; The weighted sum is obtained by calculating the sum of the products of the medication adherence tendency, the symptom rating score, and the risk score with their respective weight coefficients, plus the bias term; The weighted sum and input soft-addition function are numerically transformed to obtain an instantaneous compliance coefficient greater than zero, wherein the soft-addition function is used to map any real number to a positive real number.

6. The method according to claim 1, characterized in that, The dynamic compliance coefficient with fusion time decay is determined based on the effective confidence level, the time interval, the instantaneous compliance coefficient, and the historical average compliance baseline value, including: Obtain the time decay factor, wherein the time decay factor is equal to the negative value of the natural constant multiplied by a preset time decay constant and then multiplied by the exponent of the time interval; Calculate the product of the effective confidence level and the time decay factor to obtain the first weighting coefficient; A second weighting coefficient is determined based on the first weighting coefficient, wherein the second weighting coefficient is equal to 1 minus the first weighting coefficient; Calculate the product of the first weighting coefficient and the instantaneous compliance coefficient to obtain the first product value; Calculate the product of the second weighting coefficient and the historical average compliance benchmark value to obtain the second product value; The dynamic compliance coefficient is obtained by adding the first product to the second product.

7. The method according to claim 1, characterized in that, Based on the aforementioned dynamic compliance coefficient, the user's medication status estimation model is calibrated online to obtain the target medication status, including: Obtain the estimated remaining amount of medicine for the user at a historical time; obtain the drug purchase detection result of the user at the current time, wherein the drug purchase detection result is used to indicate whether the user has purchased medicine at the current time; Based on the user's estimated remaining drug quantity at a historical time, the dynamic compliance coefficient, the estimated daily average drug consumption, the drug purchase detection results, the preset disturbance factor, and the preset noise factor, the estimated remaining drug quantity for the user at the current time is determined. The user's medication status sequence is updated based on the estimated amount of medication remaining at the current moment, and the user's medication status estimation model is calibrated online based on the updated medication status sequence and the estimated amount of medication remaining at the current moment to obtain the target medication status.

8. The method according to claim 7, characterized in that, Based on the user's estimated remaining drug quantity at a historical time, the dynamic compliance coefficient, the estimated daily average drug consumption, the drug purchase monitoring results, a preset disturbance factor, and a preset noise factor, the estimated remaining drug quantity for the user at the current time is determined, including: Multiply the dynamic compliance coefficient by the estimated daily medication amount to obtain the compliance-adjusted daily consumption. The amount of medicine to be replenished is determined based on the results of the medicine purchase test. When the results of the medicine purchase test indicate that the user has purchased medicine at the current time, the amount of medicine to be replenished is equal to the quantity of medicine purchased. When the results of the medicine purchase test indicate that the user has not purchased medicine at the current time, the amount of medicine to be replenished is zero. The estimated amount of medicine remaining for the user at the current moment is obtained by subtracting the compliance-adjusted daily consumption from the estimated amount of medicine remaining at the historical time, and then adding the amount of medicine purchased, the preset disturbance factor, and the preset noise factor.

9. The method according to claim 1, characterized in that, After obtaining the target medication status by calibrating the user's medication status estimation model online based on the dynamic compliance coefficient, the method further includes: The user's risk level is determined based on the target medication status, and the risk level is used to characterize the probability of the user experiencing medication interruption risk; The prediction confidence level is determined based on the target medication status, and the prediction confidence level is used to quantify the accuracy of the user's medication status estimation model in estimating the current medication status. Differentiated traffic diversion and scheduling strategies are determined based on the risk level and the prediction confidence level.

10. The method according to claim 9, characterized in that, Determining the differentiated traffic diversion and scheduling strategy based on the risk level and the prediction confidence level includes: If the risk level is detected to be lower than the first risk threshold and the prediction confidence level is higher than the first confidence threshold, the user's state is determined to be stable, and a standardized reminder from the automated intelligent interaction module is triggered. If the risk level is detected to be higher than the second risk threshold and the prediction confidence level is higher than the second confidence threshold, it is determined that the user is at risk of running out of medication, and a manual telephone follow-up task is generated. If the risk level is detected to be greater than or equal to the first risk threshold and less than or equal to the second risk threshold, or if the prediction confidence level is lower than the third confidence threshold, it is determined that the user's status is fluctuating, and the user is assigned to a human online communication channel for a follow-up visit. The first confidence threshold and the second confidence threshold are both greater than the third confidence threshold, and the first confidence threshold and the second confidence threshold are equal or unequal.

11. A medication status calibration device based on interactive feedback, characterized in that, include: The acquisition unit is used to collect interactive feedback data between the user and the interactive system, the interactive feedback data including text data and / or voice data; The parsing unit is used to perform semantic parsing on the interactive feedback data using a large language model to obtain feature information related to the user's actual medication status. The first processing unit is used to determine a dynamic compliance coefficient based on the feature information, wherein the dynamic compliance coefficient is used to characterize the user's real-time execution of the medication plan; The second processing unit is used to calibrate the user's medication status estimation model online based on the dynamic compliance coefficient to obtain the target medication status, wherein the target medication status includes the updated estimated amount of remaining drugs, and the medication status estimation model is a mathematical model used to estimate the amount of remaining drugs based on drug purchase behavior and medication compliance. The first processing unit includes: a first acquisition subunit for acquiring semantic confidence in the feature information, wherein the semantic confidence is used to characterize the degree of confidence of the large language model in the semantic parsing results; a second acquisition subunit for acquiring the most similar case recalled from the historical case library based on the interaction text sequence between the user and the interaction system; an update subunit for updating the semantic confidence based on the similarity between the most similar case and the interaction text sequence to obtain an effective confidence, wherein the effective confidence is equal to the square root of the product of the semantic confidence and the similarity; a first determination subunit for determining the instantaneous compliance coefficient based on the medication adherence tendency, symptom rating score, and risk score in the feature information; a third acquisition subunit for acquiring the user's historical average compliance baseline value and the time interval between the current moment and the last effective text interaction; and a second determination subunit for determining a dynamic compliance coefficient with fusion time decay based on the effective confidence, the time interval, the instantaneous compliance coefficient, and the historical average compliance baseline value.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device containing the computer-readable storage medium performs the medication status calibration method based on interactive feedback as described in any one of claims 1 to 10.

13. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the medication status calibration method based on interactive feedback as described in any one of claims 1 to 10.

14. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the medication status calibration method based on interactive feedback as described in any one of claims 1 to 10.

Images