A smart system to facilitate the administration of subcutaneous injection drugs for people with operational disabilities
Through an intelligent system that combines semantic structure analysis and blood glucose recording, individualized injection dosage control and abnormal response identification are achieved for patients with vision and memory impairment. This solves the problem of inaccurate injection time and dosage reminders in traditional systems and improves medication safety for people with operational impairments.
Patent Information
- Application Number
- CN202611236826.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional intelligent systems for subcutaneous injection drugs cannot meet the personalized needs of patients with both visual and memory impairment. They lack multimodal information input/output capabilities and voice interaction capabilities, resulting in inaccurate injection time and dosage reminders, inability to identify abnormal injection events, and risks of injection delays and dosage errors.
The system uses a semantic structure parsing module to parse voice interaction commands, calculates individual digestive response thresholds based on blood glucose records, generates individualized injection parameters, monitors injection execution through voice and motion feedback, identifies blood glucose response delays, locates abnormal events, and provides error correction guidance.
It improves the accuracy and safety of injection procedures, reduces errors and risks caused by operational obstacles, and enhances the ability to identify and automatically repair abnormal responses.
Smart Images

Figure CN123120192A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of drug use management technology, and in particular to an intelligent system that facilitates the administration of subcutaneous injection drugs by people with operational disabilities. Background Technology
[0002] The field of medication use management technology involves the entire process of intervention and management for patients during medication use. Core aspects include medication reminders, dosage control, adherence improvement, error prevention, and medication data recording and analysis. Research focuses on improving the safety and accuracy of patient medication use through information technology and intelligent methods. With the development of an aging society and the continuous rise in the prevalence of chronic diseases, the target population for medication use management is becoming increasingly diversified, especially among individuals requiring long-term subcutaneous injections, such as those with diabetes. Its application scenarios have expanded to multiple sub-fields, including intelligent reminder systems, injection assistance devices, and human-computer interaction terminals. Existing technologies partially integrate dosage adjustment display functions in medication use management systems, and some products feature enlarged fonts to aid visual impairment. However, significant shortcomings remain in terms of reminder functions, intelligent interaction, and long-term adherence assurance.
[0003] Traditional intelligent systems for administering subcutaneous injections to individuals with operational impairments refer to systems designed to assist patients with decreased vision, memory loss, and reduced operational abilities in administering subcutaneous medications. These systems typically help patients determine the injection dosage by using a dosage display window, an additional magnifying glass component, or mechanical limiting structures. Some systems also integrate a fixed alarm clock or simple button prompts to remind patients of injection time. Some systems further utilize simplified mechanical buttons to replace complex operations and reduce the probability of misoperation. Overall, traditional technologies primarily rely on physical structure design or simple electronic prompts for functionality, lacking multimodal information input / output capabilities and voice interaction capabilities. They cannot fully meet the comprehensive assistance needs of patients with memory decline and visual impairments for time management and dosage confirmation during injections.
[0004] Traditional reminder systems primarily rely on physical structure design and simple electronic reminders to assist patients in completing subcutaneous injections. Lacking voice interaction capabilities, they are ill-suited for patients with both visual and memory impairments. Injection time and dosage reminders are typically set in a fixed manner, lacking adaptation to users' semantic understanding. User commands are often incomplete or grammatically incorrect, leading to frequent interaction failures. Dosage setting relies mainly on visual judgment and static parameter settings, lacking individualized adjustment capabilities. Furthermore, injection records do not establish a closed-loop linkage between behavioral patterns and blood glucose changes, resulting in a lack of automatic identification and response mechanisms for abnormal injection events. This leads to risks such as injection delays, dosage errors, and missing subsequent abnormal responses during medication administration. Summary of the Invention
[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide an intelligent system that facilitates the administration of subcutaneous injection drugs by people with operational disabilities. The technical solution is as follows:
[0006] On the one hand, an intelligent system is provided to facilitate the administration of subcutaneous injection drugs by people with operational disabilities, the system comprising: The semantic structure parsing module collects voice interaction commands, verifies word order rules, and extracts current meal quantity parameters. The dose response control module calculates the blood glucose variation gradient difference based on the current meal size parameter, extracts the median value of the blood glucose variation gradient difference to construct an individual digestive response threshold, compares it with the current meal size parameter, determines the injection dose offset adjustment amount, and generates the current injection parameter. The injection execution monitoring module constructs a standardized voice prompt stream based on the current injection parameters, plays the standardized voice prompt stream, collects key feedback signals and inertial sensor data, generates an injection confirmation step completion status bit and an auxiliary action recording bit, and obtains injection execution record information. The response delay analysis module monitors blood glucose values and calculates the blood glucose response delay time difference based on the injection execution record information, and identifies and marks abnormal blood glucose events. The abnormal repair guidance module verifies the consistency between the injection confirmation step completion status bit and the auxiliary action recording bit for the abnormal blood glucose event. When there is a discrepancy, it locates the user behavior interruption point and constructs abnormal correction reminder information.
[0007] As a further aspect of the present invention, the current meal quantity parameter includes the estimated total amount of carbohydrates, food glycemic index classification identifier, and relative coordinates of eating time; the current injection parameter includes the preset value of basal insulin, the dynamic compensation amount of blood glucose gradient, and the individual sensitivity correction coefficient; the injection execution record information includes the drug mixing acceleration feature code, the spatial attitude tilt angle of the injection device, and the needle suspension stability measurement value; the abnormal blood glucose event includes the blood glucose response delay window overflow flag, the drug absorption curve deviation judgment value, and the confidence level of missed injection risk; and the abnormal correction reminder information includes error correction guidance audio data frames, key layers of action demonstration, and process node reset index signals.
[0008] As a further aspect of the present invention, the semantic structure parsing module includes: The syntactic dependency feature extraction submodule collects user voice interaction commands and converts them into digitized text character sequences. It performs word segmentation and part-of-speech tagging on the text character sequences, determines the grammatical attribute category of each word, establishes directed connections based on the dominance and subdominance relationships between each word, calculates the hierarchical depth of each node in the syntactic structure, marks the dependency paths between core predicate nodes and argument nodes, and generates a syntactic dependency tree structure graph. The word order compliance verification submodule parses the syntactic dependency tree structure diagram, locates the action words marked as core predicate nodes and the object nouns as argument nodes, extracts the position index values of the action words and object nouns in the original sequence, performs position index difference operation, verifies whether the action words are at the beginning of the sequence and the object nouns are after them, and generates a word order rule verification pass signal when the operation result meets the preset constraint structure logic. The meal quantity parameter entity extraction submodule, in response to the word order rule verification pass signal, retrieves numerical characters and quantifier fragments that have a semantic modification relationship with the object noun in the text character sequence, calls the food measurement unit database to map the quantifier fragments to standard quality units, performs the product operation of numerical characters and standard quality units, quantifies the food intake scale, and generates the current meal quantity parameter.
[0009] As a further aspect of the present invention, the dose response regulation module includes: The response benchmark extraction submodule, based on the current meal quantity parameter, calls the blood glucose records of historical continuous periods, calculates the variation gradient difference obtained by subtracting the blood glucose value before eating from the blood glucose value two hours after the meal for each record, calls the time series sliding window algorithm to smooth the variation gradient difference, filter out discrete noise, locate the stable interval with the minimum variance in the sequence, and extract the statistical median level within the interval to construct the individual digestive response threshold. The drug solution compensation quantification submodule calls the individual digestive response threshold and the current meal quantity parameter, maps the current meal quantity parameter to the standard glucose equivalent value through the preset load conversion logic, performs a cross comparison operation between the individual digestive response threshold and the standard glucose equivalent value, quantifies the metabolic demand gap, and converts the metabolic demand gap into a drug solution compensation value in combination with the preset insulin sensitivity coefficient, generating the injection dose offset adjustment amount. The dosing instruction generation submodule obtains the user-preset basal insulin baseline value, performs an arithmetic correction on the basal insulin baseline value by superimposing the injection dose offset adjustment amount, calls the safety logic to verify whether the corrected value is within the safe range for a single dose, and when the verification passes, formats the value into standardized instruction data including dose units and execution time to generate the current injection parameters.
[0010] As a further aspect of the present invention, the process of minimizing the variance stable interval in the positioning sequence specifically involves: setting a sliding observation window covering a preset time span; controlling the sliding observation window to perform a sliding operation on the time sequence composed of varying gradient differences with a fixed step size; calculating the statistical variance of all data points within each local data segment captured by the sliding observation window; constructing a discrete evolution sequence including the statistical variance value corresponding to each position; retrieving continuous index intervals in the discrete evolution sequence whose values are continuously lower than a preset stability judgment threshold; mapping the continuous index intervals back to the original time axis; and selecting the interval with the largest span and the smallest average variance value as the variance minimization stable interval; wherein, the preset stability judgment threshold is set by collecting baseline blood glucose fluctuation data of the standard reference group under resting fasting conditions, calculating the variance statistical distribution characteristics of the baseline blood glucose fluctuation data, and selecting the preset quantile value in the variance statistical distribution characteristics as the preset stability judgment threshold; The process of mapping the current meal quantity parameter to a standard glucose equivalent value through a preset load conversion logic specifically involves: parsing the current meal quantity parameter to obtain the carbohydrate mass value and the food glycemic index category; calling the conversion coefficient mapping table in memory and extracting the corresponding bioavailability correction coefficient according to the food glycemic index category; performing the product operation of the carbohydrate mass value and the bioavailability correction coefficient, and using the resulting product as the standard glucose equivalent value.
[0011] As a further aspect of the present invention, the injection execution monitoring module includes: The prompt instruction construction submodule, based on the current injection parameters, calls the instruction syntax template in the natural language generation library to perform word order reconstruction, anchors the task verb that identifies the injection behavior to the starting node of the syntax structure, maps the current injection parameters to the object slot after the action node, drives the speech synthesis engine to convert the constructed text sequence into audio waveform data with acoustic features, and generates a standardized voice prompt stream. The interactive feedback verification submodule sets the signal acquisition window according to the audio duration of the standardized voice prompt stream, synchronously scans the level state of the physical button interface during the audio playback process, acquires the voltage pulse sequence generated when the user performs the confirmation operation, calls the time domain filtering algorithm to remove jitter noise and extract the falling edge feature, compares the timing deviation between the effective trigger time and the instruction end time, and generates the injection confirmation step completion status bit. The sensor data fusion submodule, for the period when the injection confirmation step completion status bit is set to an effective state, uses a six-axis inertial sensor to read acceleration and angular velocity components at a preset frequency, calls attitude calculation logic to calculate the spatial tilt angle and jitter amplitude of the device during the injection process, constructs an auxiliary action recording bit, performs spatiotemporal alignment and data packet structured encapsulation of the auxiliary action recording bit and the injection confirmation step completion status bit, and generates injection execution record information.
[0012] As a further aspect of the present invention, the response delay analysis module includes: The latency parameter extraction submodule opens a glucose monitoring time window for the start time defined by the injection execution record information, reads real-time subcutaneous interstitial fluid glucose concentration data at a preset sampling frequency, calculates the concentration decrease slope between adjacent sampling points, locates the moment when the slope first exceeds the effective response benchmark, performs time difference calculation between the time point and the drug administration end time in the injection execution record information, and generates the current blood glucose response delay time difference. Deviation Index Calculation Submodule: Calls the current blood glucose response delay time difference, retrieves the average time window parameter of blood glucose response corresponding to the historical injection behavior and the historical response stability variance, and simultaneously extracts the blood glucose decline gradient factor from the real-time monitoring data to calculate and obtain the blood glucose response deviation assessment index; The abnormal event marking submodule calls a preset upper limit threshold for abnormal judgment, compares the blood glucose response deviation assessment index with the upper limit threshold for abnormal judgment, and extracts the maximum deviation value and corresponding timestamp within the current monitoring period for cases where the blood glucose response deviation assessment index is greater than the upper limit threshold for abnormal judgment, thereby constructing an abnormal blood glucose event marker.
[0013] As a further aspect of the present invention, the process of setting the upper limit threshold for anomaly determination specifically involves retrieving a set of normal injection samples that did not trigger an alarm mechanism within a historical time period, calculating the corresponding blood glucose response deviation assessment index for each data point in the sample set, constructing a probability density statistical model based on the blood glucose response deviation assessment index, calculating the one-sided distribution critical value corresponding to the preset confidence level in the statistical model, and determining the critical value as the upper limit threshold for anomaly determination.
[0014] As a further aspect of the present invention, the anomaly repair guidance module includes: The state consistency verification submodule, based on the abnormal blood glucose event marker, retrieves the corresponding injection confirmation step completion status bit and auxiliary action recording bit, performs timing alignment, calculates the time deviation value and logical state XOR value of the two triggering times, compares the obtained values with the preset consistency judgment benchmark, identifies the logical matching status of physical action and confirmation signal, and generates a behavior logic consistency discriminant. The sequence matching and tracing submodule responds to the behavior logic consistency discriminant indicating an inconsistent state, extracts the motion trajectory vector sequence included in the auxiliary action record bit, performs sliding window matching between the sequence and the standard injection program template library, calculates the feature distance between the actual action and the standard template frame by frame, locates the moment when the distance value first exceeds the preset effective execution tolerance, marks the standard step number corresponding to the moment as the deviation point of the execution path, and generates the user behavior interruption node index value. The guidance information reconstruction submodule calls the index value of the user behavior interruption node as the address pointer of the instruction library, retrieves the remaining process sequence data that has not been executed after the pointer, extracts the corresponding text script and action keyframe, drives the speech synthesis unit to convert the text script into error correction guidance audio, synchronously calls the graphics rendering unit to generate a demonstration layer of the correct gesture, encapsulates the error correction guidance audio and demonstration layer into a multimedia data stream, and generates abnormal correction reminder information.
[0015] As a further aspect of the present invention, the process of setting the effective execution tolerance specifically involves obtaining historically verified qualified injection action trajectory samples, calculating the Euclidean distance set between the feature vector of each sample trajectory and the feature vector of the standard template, statistically analyzing the arithmetic mean and standard deviation of the distance set, and selecting the value of the arithmetic mean superimposed with a preset multiple of the standard deviation as the effective execution tolerance.
[0016] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention enhances the structural recognition ability of voice commands for elderly users by extracting semantic structure information from voice commands to construct a word order constraint template. It combines historical blood glucose records with current meal quantity parameters to calculate gradient difference and extract median level, constructing an individual digestive response threshold to achieve individualized control of injection dosage. By synchronously collecting voice prompts and action feedback signals, it establishes an injection execution record and monitors blood glucose response delay based on this record. By comparing the current delay time with the historical average response time difference, it accurately identifies abnormal response events. Based on the user behavior sequence interruption point location results, it regenerates voice and image commands to complete the prompt correction process in abnormal scenarios, improves the perception and automatic repair capabilities of abnormal injection behavior, and reduces injection errors and subsequent risks caused by operational obstacles. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1This is a schematic diagram of an intelligent system provided by an embodiment of the present invention that facilitates the administration of subcutaneous injection drugs by people with operational disabilities; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the semantic structure parsing module in this invention; Figure 4 This is a flowchart of the dose response control module in this invention; Figure 5 This is a flowchart of the injection execution monitoring module in this invention; Figure 6 This is a flowchart of the response delay analysis module in this invention; Figure 7 This is a flowchart of the anomaly repair guidance module in this invention. Detailed Implementation
[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.
[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0021] like Figure 1-2 As shown, this embodiment of the invention provides an intelligent system that facilitates the application of subcutaneous injection drugs by people with operational disabilities. The system includes a semantic structure parsing module, a dose response control module, an injection execution monitoring module, a response delay analysis module, and an anomaly repair guidance module. The semantic structure parsing module collects user voice interaction commands, calls the syntactic analysis tree in the natural language processing library to decompose the voice interaction commands into structures, and verifies whether they conform to the preset word order rule that the action item is at the beginning of the sentence and the object item follows the action item. When the verification is successful, the current meal quantity parameter is extracted from the voice interaction command. The dose response control module retrieves historical continuous blood glucose records based on the current meal size parameters and calculates the difference in blood glucose change gradient between two hours after the meal and before the meal. It then calls the sliding window algorithm to extract the median level within the stable interval to construct the individual digestive response threshold. Finally, it performs a cross-comparison operation between the individual digestive response threshold and the current meal size parameters to generate the injection dose offset adjustment amount. Based on the injection dose offset adjustment amount, it updates and generates the current injection parameters. The injection execution monitoring module, based on the insulin subcutaneous injection task, places the task verb at the beginning of the sentence and the current injection parameters after the action item to construct a standardized voice prompt stream. During the playback of the standardized voice prompt stream, it simultaneously collects key feedback signals and inertial sensor data, and generates injection confirmation step completion status bits and auxiliary action recording bits respectively to obtain injection execution record information. The response delay analysis module continuously monitors real-time blood glucose values within a specified detection window after the injection operation record occurs, based on the injection execution record information. It calculates the blood glucose response delay time difference between the current blood glucose value and the previous effective response curve, performs the difference calculation between the blood glucose response delay time difference and the average blood glucose response time window parameter corresponding to the historical injection behavior, and marks abnormal blood glucose events that exceed the upper limit threshold. The abnormal repair guidance module performs an injection confirmation step to verify the consistency between the status bit and the auxiliary action record bit for records marked as abnormal blood glucose events. When the verification result is inconsistent, it calls the sequence matching algorithm to locate the user behavior interruption point, calls the process sequence located after the user behavior interruption point in the memory to regenerate the voice and image instructions, and obtains the abnormal correction reminder information.
[0022] The current meal quantity parameters include the estimated total amount of carbohydrates, food glycemic index classification labels, and relative coordinates of eating time. The current injection parameters include the preset value of basal insulin, the amount of dynamic compensation for blood glucose gradient, and the individual sensitivity correction coefficient. The injection execution record information includes the drug mixing acceleration feature code, the spatial attitude tilt angle of the injection device, and the needle suspension stability measurement value. Abnormal blood glucose events include the blood glucose response delay window overflow flag, the drug absorption curve deviation judgment value, and the confidence level of missed injection risk. The abnormal correction reminder information includes error correction guidance audio data frames, key layers of action demonstration, and process node reset index signals.
[0023] Specifically, such as Figure 2 , 3 As shown, the semantic structure parsing module includes: The syntactic dependency feature extraction submodule collects user voice interaction commands and converts them into digitized text character sequences. It performs word segmentation and part-of-speech tagging on the text character sequences, determines the grammatical attribute category of each word, establishes directed connections based on the dominance and subdominance relationships between each word, calculates the hierarchical depth of each node in the syntactic structure, marks the dependency paths between core predicate nodes and argument nodes, and generates a syntactic dependency tree structure graph. In process control, a high-sensitivity omnidirectional microphone array is used to acquire user-issued voice interaction commands, such as "inject insulin corresponding to 200 grams of rice," at a sampling rate of 16000 Hz and quantization precision. During this process, a real-time buffer status monitoring process is initiated to dynamically manage the write pointer of the audio data stream, preventing circular buffer overflows during analog-to-digital conversion service interruptions. The acquired analog sound wave signals are converted into a pulse-code modulation (PCM) digital audio stream via an analog-to-digital converter. Next, an automatic speech recognition engine is invoked. This engine contains an acoustic model and a language model. The acoustic model employs a deep convolutional neural network structure. The input layer receives audio feature frames processed by Mel-frequency cepstral coefficients. Acoustic features are extracted through five convolutional layers, with each layer having a 3x3 kernel size and a stride of 1. Modified linear units (MLUs) are used as the activation function, and max pooling is employed to reduce dimensionality. The fully connected layers have 1024 nodes. The output layer uses a Softmax function to output the phoneme probability distribution, ultimately decoding the text character sequence "Injecting 200 grams of rice corresponds to insulin". Subsequently, word segmentation is performed on this text character sequence using a directed acyclic graph scanning algorithm based on a prefix dictionary. All possible segmentation paths are constructed, and dynamic programming is used to find the path with the highest probability, segmenting the sequence into independent word units such as "inject", "200", "gram", "rice", "corresponding to", and "insulin". Part-of-speech tagging is then performed. Note that a Hidden Markov Model (HMM) is used to calculate the conditional probability of each part-of-speech tag sequence under a given vocabulary sequence. The optimal state sequence is decoded using the Viterbi algorithm. For example, "inject" is tagged as a verb, "200" as a numeral, "gram" as a quantifier, and "rice" as a noun. Based on this, dependency parsing is performed using a transformation-based dependency parsing algorithm. A buffer stack containing all words and a stack for storing temporary processing nodes are initialized. The analyzer predicts the "move," "left arc," or "right arc" action using a multilayer perceptron classifier based on the feature vectors of the current top word in the stack and the top word in the buffer stack. If the prediction is "left arc," the dominance relationship between the top word in the buffer stack and the top word in the buffer stack is established. For example, establishing the dominance relationship between "inject" and the top word in the buffer stack... The core predicate node is "insulin", which governs "insulin" as a direct object argument node. At the same time, "rice" modifies "insulin" as a noun phrase, forming a "verb-object" and "modifier-head" relationship. Finally, the generated dependency relationship list is traversed, and the distance from the root node to each leaf node is calculated to determine the level depth. "Injection" is marked as the root node at depth 0, "insulin" as the direct object at depth 1, and "rice" as the modifier at depth 2. The dependency path from the core predicate "injection" to the argument "insulin" is extracted, and a syntactic dependency tree structure graph containing node attributes (part of speech, word meaning) and edge attributes (dependency relationship type) is constructed. After confirming that there are no broken nodes in the tree structure through the data flow integrity verification control logic, it is output to the next level cache.
[0024] The word order compliance verification submodule parses the syntactic dependency tree structure diagram, locates the action words marked as core predicate nodes and the object nouns as argument nodes, extracts the position index values of the action words and object nouns in the original sequence, performs position index difference operation, verifies whether the action words are at the beginning of the sequence and the object nouns are after them, and generates a word order rule verification pass signal when the operation result meets the preset constraint structure logic. A depth-first search algorithm is used to retrieve nodes marked as "Root" or core predicate attributes in the tree structure, identifying "injection" as the action word. Simultaneously, child nodes with connecting edges labeled "dobj" (direct object) under this core predicate node are retrieved, identifying "rice" or "insulin" as object nouns. In process control, the metadata index table of the original text character sequence is accessed, extracting the starting position index value of the action word "injection" in the sequence and setting it to 0 (indicating it's at the beginning of the sequence). The starting position index value of the object noun "rice" in the sequence is extracted and set to 4 (indicating it's in the middle of the sequence). A position index difference operation is performed, subtracting the position index value of the action word from the position index value of the object noun. In the example above, the operation is 4 minus 0, resulting in a positive value of 4. Then, a preset limit is called... The structure logic rule is defined such that compliance is only determined when the action word index value is 0 and the position index difference is greater than 0, i.e., the instruction must begin with an imperative verb and the object must follow. At this time, the execution flow control state machine will intervene in the verification process. If the logical judgment does not meet the conditions, the state machine will immediately trigger an abnormal interrupt signal to prevent illegal calls to subsequent modules. In this embodiment, since the index of "inject" is 0 and 4 is greater than 0, the logical judgment is true, verifying that the action word is at the beginning of the sequence and the object noun is after it. If the user instruction is "200 grams of rice injected", the index of "inject" is not 0, and the verification will fail. When the calculation result meets the above preset logic, the submodule generates a high-level word order rule verification pass signal, encapsulates it as a Boolean state bit "True" and sends it to the bus to activate the subsequent parameter extraction process.
[0025] The meal quantity parameter entity extraction submodule responds to the word order rule verification pass signal, retrieves numerical characters and quantifier fragments that have a semantic modification relationship with the object noun in the text character sequence, calls the food measurement unit database to map the quantifier fragments to standard quality units, performs the product operation of numerical characters and standard quality units, quantifies the food intake scale, and generates the current meal quantity parameter; A secondary scan was performed on the text character sequence "Injecting insulin corresponding to 200 grams of rice". Using the rule matching pattern in named entity recognition technology, the numeric character "200" with a "nummod" (quantity modifier) dependency and the quantifier fragment "gram" with a "clf" (quantifier) dependency were retrieved to the left, with the identified object noun "rice" as the center. Subsequently, a pre-set database of dietary measurement units was called. This database stores the conversion ratios between common dietary quantifiers and standard international units (grams or milliliters) in key-value pairs, as shown in Table 1. The standard mass unit mapping coefficient corresponding to the quantifier fragment "gram" was queried in the database, and the coefficient was 1.0. If the quantifier is "liang" (two ounces), the coefficient is 50. 0; During the mapping process, type safety process control is implemented to check the floating-point format compliance of the extracted numerical characters to prevent non-numerical characters from being mixed into the calculation unit; Next, the product operation of the numerical character and the standard mass unit is performed, converting the extracted numerical character "200" into the floating-point number 200.0, and multiplying it by the mapping coefficient 1.0. The calculation formula is 200.0 multiplied by 1.0, and the result is 200.0, which is the quantified food intake scale; Finally, this value is marked as the current meal quantity parameter, with the unit uniformly set to grams, and a data packet containing the food type ID (rice) and mass value (200.0) is generated. This process effectively transforms the vague food quantity described in natural language into a precise physical quantity that can be processed by a computer.
[0026] Table 1 Mapping Table of Dietary Measurement Units
[0027] As shown in Table 1, the corresponding mapping coefficient is retrieved based on the quantifier identifier, which is used to convert non-standard quantifiers into standard weights. For example, "1 bowl" will be mapped to a standard reference mass of 250 grams.
[0028] Specifically, such as Figure 2 , 4 As shown, the dose response control module includes: The response benchmark extraction submodule, based on the current meal quantity parameter, calls the blood glucose records of historical continuous periods. For each record, it calculates the gradient difference obtained by subtracting the blood glucose value before eating from the blood glucose value two hours after the meal. It calls the time series sliding window algorithm to smooth the gradient difference, filter out discrete noise, locate the stable interval with the minimum variance in the sequence, and extract the statistical median level within the interval to construct the individual digestive response threshold. The process of minimizing the variance in the localization sequence is as follows: A sliding observation window covering a preset time span is set, and the sliding observation window is controlled to perform a sliding operation on the time series composed of varying gradient differences with a fixed step size; for each local data segment captured by the sliding observation window, the statistical variance of all data points within it is calculated; a discrete evolution sequence including the statistical variance value corresponding to each position is constructed; continuous index intervals with values consistently below a preset stability threshold are retrieved from the discrete evolution sequence; the continuous index intervals are mapped back to the original time axis, and the interval with the largest span and smallest average variance value is selected as the variance minimization stable interval; the preset stability threshold is set by collecting baseline blood glucose fluctuation data of the standard reference group under resting fasting conditions, calculating the variance statistical distribution characteristics of the baseline blood glucose fluctuation data, and selecting the preset quantile values in the variance statistical distribution characteristics as the preset stability threshold. The system retrieves historical blood glucose records from the past 30 consecutive periods in memory. For each historical record, it extracts the blood glucose reading at the time of eating (e.g., 5.0 mmol / L) and the blood glucose reading two hours after eating (e.g., 8.5 mmol / L). It then performs a subtraction operation to calculate the gradient difference, i.e., 8.5 minus 5.0 equals 3.5 mmol / L, thus constructing a time series consisting of 30 gradient differences. Next, it calls a time series sliding window algorithm, which runs in a controlled iterative loop. A boundary check controller ensures the window index does not exceed the bounds, and the window size is set to 5 data points. The sequence is then smoothed using a moving average, for example, by calculating the arithmetic mean of points 1 to 5 in the sequence to replace the value of point 3, thereby filtering out discrete noise caused by random factors. Subsequently, it locates the stable interval for minimizing variance in the sequence, as follows: a sliding observation window covering a 7-day (i.e., 7 data points) time span is set, and this window is controlled with a fixed step size of 1 day on the smoothed gradient difference sequence. Slide upwards; for each sliding window, extract 7 local data points and calculate their statistical variance. For example, if the data in a window is [3.4, 3.5, 3.3, 3.5, 3.4, 3.6, 3.4], calculate its mean as 3.44 and its variance as 0.008, constructing a dispersion evolution sequence. Search within the dispersion evolution sequence for continuous index intervals where the values are consistently below a preset stability threshold. The preset stability threshold is obtained by collecting 24-hour baseline blood glucose fluctuation data from 100 standard reference group participants in a resting fasting state, calculating the variance statistical distribution of this dataset, selecting the value at the 95th percentile, and setting it to 0.05 as the stability threshold. Find a segment in the dispersion sequence where the variance of 5 consecutive windows is below 0.05, map it back to the original time axis, select the data within this span, calculate the statistical median level, set it to 3.2 mmol / L, and construct this value as the individual digestive response threshold. This threshold represents the typical blood glucose increase for this meal amount in a stable state.
[0029] The drug solution compensation quantification submodule calls the individual digestive response threshold and the current meal volume parameter. Through the preset load conversion logic, it maps the current meal volume parameter to the standard glucose equivalent value, performs a cross-comparison calculation between the individual digestive response threshold and the standard glucose equivalent value, quantifies the metabolic demand gap, and combines the preset insulin sensitivity coefficient to convert the metabolic demand gap into a drug solution compensation value, generating the injection dose offset adjustment amount. The process of mapping the current meal quantity parameter to the standard glucose equivalent value through the preset load conversion logic is as follows: parsing the current meal quantity parameter to obtain the carbohydrate mass value and the food glycemic index category; calling the conversion coefficient mapping table in memory, extracting the corresponding bioavailability correction coefficient according to the food glycemic index category; performing the product operation of the carbohydrate mass value and the bioavailability correction coefficient, and using the product as the standard glucose equivalent value. The system receives the current meal quantity parameter (e.g., 200 grams of rice) and the individual digestive response threshold (e.g., 3.2 mmol / L). It parses the current meal quantity parameter, consults a food composition database to obtain the carbohydrate content of "rice" as 25.9%, calculates the carbohydrate mass value (200 multiplied by 0.259 equals 51.8 grams), and identifies its glycemic index category as "high GI". In process control, it calls the conversion coefficient mapping table in memory, extracts the bioavailability correction coefficient based on the "high GI" category, and sets the coefficient corresponding to high GI to 1.0 (0.8 for medium GI and 0.6 for low GI). In this step, the coefficient retrieval logic is controlled by a numerical range validator. If the extracted coefficient exceeds the preset biologically reasonable range (e.g., 0.1 to 1.5), a default value rollback mechanism is automatically triggered. Finally, it performs a product operation, multiplying 51.8 grams by 1.0 to obtain 51.8 as the standard glucose equivalent value. Subsequently, a cross-comparison operation is performed to quantify the metabolic demand gap. Using a pre-defined load conversion logic, which sets the expected increase in blood glucose to 0.8 mmol / L for every 10 grams of standard glucose equivalent (this coefficient is based on a general physiological model), 51.8 grams is expected to cause an increase in blood glucose of 4.144 mmol / L. The difference between the expected increase of 4.144 and the individual digestive response threshold of 3.2 is calculated, i.e., 4.144 minus 3.2 equals 0.944 mmol / L. This difference represents the metabolic demand gap not covered by the body's own regulation. Finally, combined with a pre-defined insulin sensitivity coefficient, it is set that each unit of insulin injected by the user can reduce blood glucose by 2.0 mmol / L. The metabolic demand gap of 0.944 is divided by the insulin sensitivity coefficient of 2.0, resulting in 0.472. This is rounded up to 0.5 units according to a stepping rule to generate an injection dose offset adjustment, used to compensate for additional blood glucose fluctuations caused by food intake.
[0030] The dosing instruction generation submodule obtains the user's preset basal insulin baseline value, performs an arithmetic correction on the basal insulin baseline value by adding the injection dose offset adjustment, calls the safety logic to verify whether the corrected value is within the safe range for a single dose, and when the verification passes, formats the value into standardized instruction data including the dose unit and execution time to generate the current injection parameters. The system retrieves the preset basal insulin baseline value from the non-volatile storage area of the user configuration file, setting the baseline value for the current period to 10.0 units. An arithmetic correction is performed on this baseline value; this correction operation is encapsulated in an atomic transaction control block to ensure data consistency during calculation. The previously generated injection dose offset adjustment (e.g., 0.5 units) is summed with the basal insulin baseline value of 10.0 units, i.e., 10.0 plus 0.5 equals 10.5 units. Subsequently, the system calls the safety logic verification module, reads the preset upper limit of the single-dose safety range, sets it to 20.0 units, and the lower limit, sets it to 2.0 units. The system then compares the modified values. If the value 10.5 is within the closed interval between 2.0 and 20.0, the verification is considered successful. If the value is greater than or equal to 2.0 and less than or equal to 20.0, the clamping logic will be triggered to limit the dosage and trigger an alarm. When the verification is successful, the current timestamp (e.g., January 22, 2026, 12:00:00) is obtained. The value "10.5", the unit "unit", and the execution time "12:00:00" are formatted into a standardized hexadecimal instruction data packet. The header contains a synchronization frame header, the middle contains the data payload, and the tail contains a cyclic redundancy check code. Finally, the current injection parameters are generated and prepared to be sent to the actuator or prompt module.
[0031] Specifically, such as Figure 2 , 5 As shown, the injection execution monitoring module includes: The prompt instruction construction submodule, based on the current injection parameters, calls the instruction syntax template in the natural language generation library to perform word order reconstruction, anchors the task verb that identifies the injection behavior to the starting node of the syntax structure, maps the current injection parameters to the object slot after the action node, drives the speech synthesis engine to convert the constructed text sequence into audio waveform data with acoustic features, and generates a standardized voice prompt stream; The system calls a command syntax template from the natural language generation library, with the template format "[task verb][object][dosage][unit]"; it performs word order reconstruction, anchoring the task verb "inject" (identifying the injection action) to the start node of the syntax structure, mapping "insulin" to the object slot after the action node, mapping the parameter value "10.5" to the dosage slot, and mapping "unit" to the unit slot, combining these to generate the text sequence "Please inject 10.5 units of insulin"; in the process control, it drives the speech synthesis engine, whose resource scheduling is managed by a priority controller. The synthesis task is completed within the time slice required for real-time processing. The engine first converts the text sequence from word to phoneme, predicts prosodic features (such as pitch, duration, and energy), and then uses a vocoder based on a deep neural network, such as WaveNet or HiFi-GAN, to convert the acoustic feature parameters into 16-bit, 44.1 kHz audio waveform data with natural human voice characteristics. The generated audio data is then subjected to gain adjustment and dynamic range compression to ensure clarity, and finally a standardized voice prompt stream is generated, ready to be played through a speaker to guide user operation.
[0032] The interactive feedback verification submodule sets the signal acquisition window according to the audio duration of the standardized voice prompt stream, synchronously scans the level status of the physical button interface during the audio playback process, acquires the voltage pulse sequence generated when the user performs the confirmation operation, calls the time domain filtering algorithm to remove jitter noise and extract the falling edge feature, compares the timing deviation between the effective trigger time and the instruction end time, and generates the injection confirmation step completion status bit. The module parses the header information of the standardized voice prompt stream audio file to obtain its audio duration, for example, 3.5 seconds. Based on this duration, a signal acquisition window is set, with the window opening synchronously with the start of audio playback. The window duration is set to the audio duration plus a 3-second redundant waiting time, for a total of 6.5 seconds. During audio playback, the module synchronously scans the GPIO level of the physical confirmation button interface at a frequency of 100 Hz. The priority control logic of the interrupt service routine ensures that the button signal can preempt low-priority tasks and be responded to immediately; the default state is a high level of 3.3 volts. When the user presses the button, the level is pulled low to 0. The system generates a voltage pulse sequence. After acquiring the pulse signal, a time-domain filtering algorithm is called, employing debouncing logic. This means that a low-level state is continuously detected for more than 20 milliseconds before it is considered a valid button press, eliminating microsecond-level noise caused by mechanical contact jitter. The falling edge feature of the valid button action is extracted, and the valid trigger time is recorded. Subsequently, the timing deviation between the valid trigger time and the instruction end time is compared. If the trigger time is within the 3-second waiting window after the instruction playback ends, the interaction logic is determined to be valid. A binary injection confirmation step completion status bit is generated and set to 1 (valid status), indicating that the user has confirmed and is ready to perform the injection.
[0033] The sensor data fusion submodule, for the period when the injection confirmation step completion status bit is set to an effective state, uses a six-axis inertial sensor to read acceleration and angular velocity components at a preset frequency, calls attitude calculation logic to calculate the spatial tilt angle and jitter amplitude of the device during the injection process, constructs an auxiliary action recording bit, performs spatiotemporal alignment and data packet structured encapsulation of the auxiliary action recording bit and the injection confirmation step completion status bit, and generates injection execution record information. The six-axis inertial sensor (including a three-axis accelerometer and a three-axis gyroscope) is activated, and the sensor bus synchronization control signal is simultaneously initiated to ensure strict alignment of data in each axis in the time domain. During process control, the acceleration and angular velocity components of the device in the X, Y, and Z axes are read at a preset frequency of 200 Hz. Based on the acquired raw data, the attitude calculation logic is invoked, and a quaternion complementary filtering algorithm is used to fuse the low-frequency stability characteristics of the accelerometer and the high-frequency dynamic characteristics of the gyroscope to calculate the Euler angles (pitch, roll, and yaw) of the device in space in real time. This determines the spatial tilt angle of the device, for example, for detecting… The device is tilted vertically downwards at 45 degrees (consistent with abdominal injection posture), and the variance of the three-axis acceleration is calculated to quantify the jitter amplitude. An auxiliary motion recording bit is constructed, and the calculated maximum tilt angle data, average jitter amplitude, and motion duration characteristics are written into it. Next, a spatiotemporal alignment operation is performed to match the timestamp of the auxiliary motion recording bit with the generation time of the injection confirmation step completion status bit, ensuring that the two belong to the same operation event. Finally, the aligned data is encapsulated according to a predefined protocol structure to generate injection execution record information containing motion feature vectors and confirmation status, which is then stored in the historical database.
[0034] Specifically, such as Figure 2 , 6 As shown, the response delay analysis module includes: The latency parameter extraction submodule opens a glucose monitoring time window for the start time defined by the injection execution record information, reads real-time subcutaneous interstitial fluid glucose concentration data at a preset sampling frequency, calculates the concentration decrease slope between adjacent sampling points, locates the moment when the slope first exceeds the effective response benchmark, performs time difference calculation between the time point and the drug administration end time in the injection execution record information, and generates the current blood glucose response delay time difference. For the dosing end time defined in the injection execution record (e.g., 12:05:00), a glucose monitoring time window is opened starting from this time, with a duration set to 60 minutes. The operation of this time window is monitored by a watchdog timer to prevent unexpected suspension of the monitoring process. During this period, subcutaneous interstitial fluid glucose concentration data is read by a continuous glucose monitoring sensor at a preset sampling frequency of once every 3 minutes. The concentration decrease slope between two adjacent sampling points is calculated in real time, using the formula (concentration value at the later time minus concentration value at the previous time) divided by the time interval (3 minutes). The effective response baseline slope is set at -0.1 mmol / L / min. When the calculated slope value is first less than this baseline value (i.e., the rate of decrease exceeds 0.1), the timestamp corresponding to the sampling point (e.g., 12:30:00) is locked as the response start point. The time difference between this time point and the end time of administration is calculated, i.e., 12:30:00 minus 12:05:00, resulting in 25 minutes (i.e., 25.0 minutes). This value is generated as the current blood glucose response delay time difference. This parameter reflects the actual onset speed at which insulin begins to exert its blood glucose-lowering effect after being injected into the body.
[0035] The deviation index calculation submodule: calls the current blood glucose response delay time difference, retrieves the average time window parameter of blood glucose response corresponding to historical injection behavior and the historical response stability variance, and simultaneously extracts the blood glucose descent gradient factor from real-time monitoring data, using the formula: ; The blood glucose response deviation assessment index is obtained through calculation. in, Represents the glycemic response bias assessment index. This represents the time difference in the current blood glucose response delay, obtained by calculating the time interval between the end of drug administration and the start of blood glucose decline. The parameter representing the average time window of blood glucose response is obtained by retrieving the arithmetic mean of the response delay times from historical normal injection records. Representing the historical response stability variance, it is obtained by calculating the statistical variance of historical response delay time data. Representing the The normalized blood glucose decline gradient factor for the nth sampling point is calculated by... The real-time blood glucose decrease rate at each sampling point is obtained by dividing by a preset standard maximum metabolic rate constant. The total number of sampling points is obtained by counting the number of valid data points within the statistical monitoring window. The normalized baseline drift correction value is obtained by detecting the sensor baseline drift signal strength and dividing it by the sensor's maximum allowable drift tolerance. The system retrieves the average blood glucose response time window parameters corresponding to the user's past 100 normal injection behaviors from the historical database, based on the current blood glucose response delay time difference. In process control, the arithmetic mean is set to 20.0 minutes. Simultaneously, it retrieves the historical response stability variance, i.e., the variance value of historical response time data, set to 4.0. It then extracts all data within the post-dose monitoring window from the real-time monitoring data. The blood glucose decrease gradient factor for each sampling point is set, and the total number of sampling points is determined. The normalized gradient factor is obtained by setting the gradient factor to 10 and obtaining the real-time rate of decline at each point. The normalized gradient factor is obtained by dividing the rate by the preset standard maximum metabolic rate constant (e.g., 2.0 mmol / L / min). The sum of the factors at the 10 points is set to 0.8. At the same time, the normalized baseline drift correction value is obtained by detecting the intensity of the sensor baseline drift signal and dividing it by the maximum allowable drift tolerance (e.g., 0.5 mmol / L), which is set to 0.1. Subsequently, the submodule executes the following computational process described in the textual logic to obtain the blood glucose response deviation assessment index. Throughout the entire computation process, the non-zero division check control logic continuously runs to prevent mathematical calculation anomalies: First, the absolute value of the difference between the current blood glucose response delay time difference and the average blood glucose response time window parameter is calculated, i.e., the absolute value of the difference between 25.0 and 20.0 is calculated, yielding a value of 5.0; second, the arithmetic square root of the historical response stability variance is calculated, i.e., the square root of 4.0 is taken, yielding a value of 2.0; third, the absolute value of 5.0 obtained above is divided by the square root value of 2.0, yielding the first intermediate result of 2.5; next, the arithmetic square root of the ratio of the current blood glucose response delay time difference to the average blood glucose response time window parameter is calculated, i.e., 25.0 divided by 20.0 is taken as... The result is 1.25. Taking the square root of 1.25 gives approximately 1.118. Then, the sum of the normalized blood glucose decline gradient factors of all sampling points (0.8) is added to the square root value of 1.118 to obtain the second intermediate result of 1.918. Subsequently, the product of the first intermediate result of 2.5 and the second intermediate result of 1.918 is performed to obtain the third intermediate result of 4.795. Finally, the absolute value of the normalized baseline drift correction value, i.e., the absolute value of 0.1 (0.1), is calculated, and the third intermediate result of 4.795 is subtracted from this absolute value of 0.1 to obtain the final blood glucose response bias assessment index of 4.695. By introducing the joint weighting of historical variance and real-time gradient, this index can reflect the degree of current metabolic abnormality of an individual more sensitively than the traditional single time difference index.
[0036] The abnormal event marking submodule calls the preset upper limit threshold for abnormal judgment, compares the blood glucose response deviation assessment index with the upper limit threshold for abnormal judgment, and extracts the maximum deviation value and corresponding timestamp within the current monitoring period for cases where the blood glucose response deviation assessment index is greater than the upper limit threshold for abnormal judgment, and constructs an abnormal blood glucose event marker. The process of setting the upper limit threshold for anomaly judgment is as follows: retrieve the normal injection sample set that did not trigger the alarm mechanism in the historical period, calculate the corresponding blood glucose response deviation assessment index for each data in the sample set, construct a probability density statistical model based on the blood glucose response deviation assessment index, calculate the one-sided distribution critical value corresponding to the preset confidence level in the statistical model, and determine the critical value as the upper limit threshold for anomaly judgment. The system invokes a preset upper threshold for anomaly detection, set to 3.5. It then compares the calculated blood glucose response deviation assessment index of 4.695 with this threshold. This comparison is executed by hysteresis comparison control logic to prevent state fluctuations at the threshold threshold. Since 4.695 is greater than 3.5, the system determines the current state to be abnormal. In response, it extracts the value with the largest deviation (4.695) and its corresponding timestamp within the current monitoring period, constructing an abnormal blood glucose event record containing "Event Type: Response Delay Anomaly," "Deviation: 4.695," and "Time: Current Time." The data structure is labeled; the process for setting the upper limit threshold for anomaly judgment is as follows: retrieve 500 normal injection samples that have not triggered the alarm mechanism in the past three months, and calculate the corresponding blood glucose response deviation assessment index for each sample data according to the aforementioned logic; based on these 500 index values, construct a probability density statistical model (e.g., a normal distribution model), and calculate the mean and standard deviation of the model; select the critical value of the one-sided distribution corresponding to a confidence level of 99% (i.e., the mean plus 2.33 times the standard deviation), set the calculation result to 3.5, and determine this value as the upper limit threshold for anomaly judgment to ensure that the false alarm rate is controlled within 1%.
[0037] Specifically, such as Figure 2 , 7 As shown, the anomaly repair guidance module includes: The state consistency verification submodule, based on the abnormal blood glucose event marker, retrieves the corresponding injection confirmation step completion status bit and auxiliary action recording bit, performs timing alignment, calculates the time deviation value and logical state XOR value of the two triggering times, compares the obtained values with the preset consistency judgment benchmark, identifies the logical matching status of physical action and confirmation signal, and generates a behavior logic consistency discriminant. Retrieve the injection confirmation step completion status bit (set to 1, indicating that the user has pressed the button to confirm) and the auxiliary action recording bit (including the action features recorded by the inertial sensor) within the time period corresponding to the event; after performing timing alignment, calculate the time deviation between the physical action occurrence time and the button confirmation time, set to 2 seconds (within a reasonable range); then calculate the logical state XOR value: this operation process is controlled by the validity flag bit, and the XOR logic is only performed when both status bits are marked as "valid". The logical value (1) of the confirmation status bit and the valid action identifier in the auxiliary action recording bit (set the sensor to determine that the action amplitude is too small and does not constitute a valid injection action, the logical value is 0) are XORed; the result of 1XOR0 is 1, which represents a logical inconsistency; compare the obtained value 1 with the preset consistency judgment benchmark (0, that is, the two are required to be completely consistent) to identify that there is a logical conflict between the physical action (no valid injection) and the confirmation signal (confirmation), and generate a behavior logic consistency discriminant indicating "false confirmation", revealing that the user may have only pressed the confirmation button but did not actually perform the standard injection action.
[0038] The sequence matching and tracing submodule responds to the behavior logic consistency discriminant indicating the inconsistent state, extracts the motion trajectory vector sequence included in the auxiliary action record bit, performs sliding window matching between the sequence and the standard injection program template library, calculates the feature distance between the actual action and the standard template frame by frame, locates the moment when the distance value first exceeds the preset effective execution tolerance, marks the standard step number corresponding to the moment as the deviation point of the execution path, and generates the user behavior interruption node index value. The specific process for setting the effective execution tolerance is as follows: by obtaining historically verified qualified injection action trajectory samples, calculating the Euclidean distance set between the feature vector of each sample trajectory and the feature vector of the standard template, calculating the arithmetic mean and standard deviation of the distance set, and selecting the value of the arithmetic mean plus a preset multiple of the standard deviation as the effective execution tolerance. The system responds to inconsistent states (i.e., discriminants indicating "false confirmation" or "abnormal action") by using a consistent behavior logic discriminant. It extracts a three-dimensional motion trajectory vector sequence containing acceleration and angular velocity from the auxiliary motion recording bit. Simultaneously, it calls the standard injection procedure template library, which stores the ideal trajectory feature vectors of standard injection actions. It executes a dynamic time warping algorithm or sliding window matching, with the matching process monitored by a loop iteration controller. The matching loop is forcibly terminated when a sequence end marker is detected. The actual action sequence is compared frame-by-frame with the standard template, calculating the Euclidean distance between the actual action features and the corresponding standard template features in each frame. The distance value is detected frame-by-frame, and the moment when the distance value first exceeds a preset effective execution tolerance is identified. The distance suddenly increases at frame 150; the standard step number corresponding to this moment (e.g., "needle insertion stage") is marked as the deviation point of the execution path, and a user behavior interruption node index value (e.g., Index=3) is generated; the effective execution tolerance setting process is as follows: obtain 1000 historically verified qualified injection action trajectory samples, calculate the Euclidean distance set between the feature vector of each sample trajectory and the feature vector of the standard template; calculate the arithmetic mean (e.g., 0.5) and standard deviation (e.g., 0.1) of this distance set; select the value of the arithmetic mean plus 3 times the standard deviation, that is, 0.5 plus 3 multiplied by 0.1, to get 0.8, which is used as the effective execution tolerance to tolerate normal action differences between individual users.
[0039] The guidance information reconstruction submodule calls the user behavior interruption node index value as the address pointer of the instruction library, retrieves the remaining process sequence data that has not been executed after the pointer, extracts the corresponding text script and action keyframe, drives the speech synthesis unit to convert the text script into error correction guidance audio, synchronously calls the graphics rendering unit to generate the demonstration layer of the correct gesture, encapsulates the error correction guidance audio and demonstration layer into a multimedia data stream, and generates abnormal correction reminder information.
[0040] Using the user behavior interruption node index value (e.g., Index=3, corresponding to the "needle insertion stage") as the addressing pointer of the instruction library, in process control, the remaining process sequence data that is located after the pointer but has not yet been correctly executed is retrieved from the pre-stored standardized operation process database. That is, the text scripts of subsequent steps such as "needle insertion", "injection of medicine", and "needle removal" (e.g., "please keep the needle perpendicular to the skin") and the corresponding standard action keyframe animation data are extracted. The speech synthesis unit is driven to convert the extracted error correction text script into error correction guidance audio with a rapid tone and emphasis. The graphics rendering unit is called simultaneously to load the 3D demonstration layer of the correct gesture and mark the incorrect gesture parts with a highlight color. Finally, the error correction guidance audio and the demonstration layer are encapsulated into a time-synchronized multimedia data stream to generate an abnormal correction reminder message. The output of this message is taken over by the interface display controller, which forcibly interrupts the current non-urgent task. This message will immediately interrupt the current regular prompts and be output to the user through the display screen and speaker to correct their erroneous operation in real time.
[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the technical solution.
Claims
1. An intelligent system for facilitating the administration of subcutaneous injection drugs by people with operational disabilities, characterized in that, The system includes: The semantic structure parsing module collects voice interaction commands, verifies word order rules, and extracts current meal quantity parameters. The dose response control module calculates the blood glucose variation gradient difference based on the current meal size parameter, extracts the median value of the blood glucose variation gradient difference to construct an individual digestive response threshold, compares it with the current meal size parameter, determines the injection dose offset adjustment amount, and generates the current injection parameter. The injection execution monitoring module constructs a standardized voice prompt stream based on the current injection parameters, plays the standardized voice prompt stream, collects key feedback signals and inertial sensor data, generates an injection confirmation step completion status bit and an auxiliary action recording bit, and obtains injection execution record information. The response delay analysis module monitors blood glucose values and calculates the blood glucose response delay time difference based on the injection execution record information, and identifies and marks abnormal blood glucose events. The abnormal repair guidance module verifies the consistency between the injection confirmation step completion status bit and the auxiliary action recording bit for the abnormal blood glucose event. When there is a discrepancy, it locates the user behavior interruption point and constructs abnormal correction reminder information.
2. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 1, characterized in that, The current meal quantity parameters include the estimated total amount of carbohydrates, food glycemic index classification identifier, and relative coordinates of eating time. The current injection parameters include the preset value of basal insulin, the dynamic compensation amount of blood glucose gradient, and the individual sensitivity correction coefficient. The injection execution record information includes the drug mixing acceleration feature code, the spatial attitude tilt angle of the injection device, and the needle suspension stability measurement value. The abnormal blood glucose events include the blood glucose response delay window overflow flag, the drug absorption curve deviation judgment value, and the confidence level of missed injection risk. The abnormal correction reminder information includes error correction guidance audio data frames, key layers of action demonstration, and process node reset index signals.
3. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 1, characterized in that, The semantic structure parsing module includes: The syntactic dependency feature extraction submodule collects user voice interaction commands and converts them into digitized text character sequences. It performs word segmentation and part-of-speech tagging on the text character sequences, determines the grammatical attribute category of each word, establishes directed connections based on the dominance and subdominance relationships between each word, calculates the hierarchical depth of each node in the syntactic structure, marks the dependency paths between core predicate nodes and argument nodes, and generates a syntactic dependency tree structure graph. The word order compliance verification submodule parses the syntactic dependency tree structure diagram, locates the action words marked as core predicate nodes and the object nouns as argument nodes, extracts the position index values of the action words and object nouns in the original sequence, performs position index difference operation, verifies whether the action words are at the beginning of the sequence and the object nouns are after them, and generates a word order rule verification pass signal when the operation result meets the preset constraint structure logic. The meal quantity parameter entity extraction submodule, in response to the word order rule verification pass signal, retrieves numerical characters and quantifier fragments that have a semantic modification relationship with the object noun in the text character sequence, calls the food measurement unit database to map the quantifier fragments to standard quality units, performs the product operation of numerical characters and standard quality units, quantifies the food intake scale, and generates the current meal quantity parameter.
4. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 1, characterized in that, The dose response control module includes: The response benchmark extraction submodule, based on the current meal quantity parameter, calls the blood glucose records of historical continuous periods, calculates the variation gradient difference obtained by subtracting the blood glucose value before eating from the blood glucose value two hours after the meal for each record, calls the time series sliding window algorithm to smooth the variation gradient difference, filter out discrete noise, locate the stable interval with the minimum variance in the sequence, and extract the statistical median level within the interval to construct the individual digestive response threshold. The drug solution compensation quantification submodule calls the individual digestive response threshold and the current meal quantity parameter, maps the current meal quantity parameter to the standard glucose equivalent value through the preset load conversion logic, performs a cross comparison operation between the individual digestive response threshold and the standard glucose equivalent value, quantifies the metabolic demand gap, and converts the metabolic demand gap into a drug solution compensation value in combination with the preset insulin sensitivity coefficient, generating the injection dose offset adjustment amount. The dosing instruction generation submodule obtains the user-preset basal insulin baseline value, performs an arithmetic correction on the basal insulin baseline value by superimposing the injection dose offset adjustment amount, calls the safety logic to verify whether the corrected value is within the safe range for a single dose, and when the verification passes, formats the value into standardized instruction data including dose units and execution time to generate the current injection parameters.
5. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 4, characterized in that, The process of minimizing the stable interval of variance in the positioning sequence is as follows: a sliding observation window covering a preset time span is set, and the sliding observation window is controlled to perform a sliding operation on the time sequence composed of changing gradient differences with a fixed step size; for each local data segment intercepted by the sliding observation window, the statistical variance value of all data points inside it is calculated; a discrete evolution sequence including the statistical variance value corresponding to each position is constructed; and a continuous index interval with values continuously lower than a preset stability judgment threshold is retrieved in the discrete evolution sequence. The continuous index intervals are mapped back to the original time axis, and the interval with the largest span and the smallest average variance is selected as the stable interval with the smallest variance. The preset stability judgment threshold is set by collecting the baseline blood glucose fluctuation data of the standard reference group under the resting fasting state, calculating the variance statistical distribution characteristics of the baseline blood glucose fluctuation data, and selecting the preset quantile value in the variance statistical distribution characteristics as the preset stability judgment threshold. The process of mapping the current meal quantity parameter to a standard glucose equivalent value through a preset load conversion logic specifically involves: parsing the current meal quantity parameter to obtain the carbohydrate mass value and the food glycemic index category; calling the conversion coefficient mapping table in memory and extracting the corresponding bioavailability correction coefficient according to the food glycemic index category; performing the product operation of the carbohydrate mass value and the bioavailability correction coefficient, and using the resulting product as the standard glucose equivalent value.
6. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 1, characterized in that, The injection execution monitoring module includes: The prompt instruction construction submodule, based on the current injection parameters, calls the instruction syntax template in the natural language generation library to perform word order reconstruction, anchors the task verb that identifies the injection behavior to the starting node of the syntax structure, maps the current injection parameters to the object slot after the action node, drives the speech synthesis engine to convert the constructed text sequence into audio waveform data with acoustic features, and generates a standardized voice prompt stream. The interactive feedback verification submodule sets the signal acquisition window according to the audio duration of the standardized voice prompt stream, synchronously scans the level state of the physical button interface during the audio playback process, acquires the voltage pulse sequence generated when the user performs the confirmation operation, calls the time domain filtering algorithm to remove jitter noise and extract the falling edge feature, compares the timing deviation between the effective trigger time and the instruction end time, and generates the injection confirmation step completion status bit. The sensor data fusion submodule, for the period when the injection confirmation step completion status bit is set to an effective state, uses a six-axis inertial sensor to read acceleration and angular velocity components at a preset frequency, calls attitude calculation logic to calculate the spatial tilt angle and jitter amplitude of the device during the injection process, constructs an auxiliary action recording bit, performs spatiotemporal alignment and data packet structured encapsulation of the auxiliary action recording bit and the injection confirmation step completion status bit, and generates injection execution record information.
7. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 1, characterized in that, The response delay analysis module includes: The latency parameter extraction submodule opens a glucose monitoring time window for the start time defined by the injection execution record information, reads real-time subcutaneous interstitial fluid glucose concentration data at a preset sampling frequency, calculates the concentration decrease slope between adjacent sampling points, locates the moment when the slope first exceeds the effective response benchmark, performs time difference calculation between the time point and the drug administration end time in the injection execution record information, and generates the current blood glucose response delay time difference. Deviation Index Calculation Submodule: Calls the current blood glucose response delay time difference, retrieves the average time window parameter of blood glucose response corresponding to the historical injection behavior and the historical response stability variance, and simultaneously extracts the blood glucose decline gradient factor from the real-time monitoring data to calculate and obtain the blood glucose response deviation assessment index; The abnormal event marking submodule calls a preset upper limit threshold for abnormal judgment, compares the blood glucose response deviation assessment index with the upper limit threshold for abnormal judgment, and extracts the maximum deviation value and corresponding timestamp within the current monitoring period for cases where the blood glucose response deviation assessment index is greater than the upper limit threshold for abnormal judgment, thereby constructing an abnormal blood glucose event marker.
8. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 7, characterized in that, The process of setting the upper limit threshold for anomaly determination is as follows: retrieve a set of normal injection samples that did not trigger the alarm mechanism within a historical period, calculate the corresponding blood glucose response deviation assessment index for each data point in the sample set, construct a probability density statistical model based on the blood glucose response deviation assessment index, calculate the one-sided distribution critical value corresponding to the preset confidence level in the statistical model, and determine the critical value as the upper limit threshold for anomaly determination.
9. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 1, characterized in that, The anomaly repair guidance module includes: The state consistency verification submodule, based on the abnormal blood glucose event marker, retrieves the corresponding injection confirmation step completion status bit and auxiliary action recording bit, performs timing alignment, calculates the time deviation value and logical state XOR value of the two triggering times, compares the obtained values with the preset consistency judgment benchmark, identifies the logical matching status of physical action and confirmation signal, and generates a behavior logic consistency discriminant. The sequence matching and tracing submodule responds to the behavior logic consistency discriminant indicating an inconsistent state, extracts the motion trajectory vector sequence included in the auxiliary action record bit, performs sliding window matching between the sequence and the standard injection program template library, calculates the feature distance between the actual action and the standard template frame by frame, locates the moment when the distance value first exceeds the preset effective execution tolerance, marks the standard step number corresponding to the moment as the deviation point of the execution path, and generates the user behavior interruption node index value. The guidance information reconstruction submodule calls the index value of the user behavior interruption node as the address pointer of the instruction library, retrieves the remaining process sequence data that has not been executed after the pointer, extracts the corresponding text script and action keyframe, drives the speech synthesis unit to convert the text script into error correction guidance audio, synchronously calls the graphics rendering unit to generate a demonstration layer of the correct gesture, encapsulates the error correction guidance audio and demonstration layer into a multimedia data stream, and generates abnormal correction reminder information.
10. The intelligent system for facilitating the administration of subcutaneous injection drugs to people with operational disabilities according to claim 9, characterized in that, The process of setting the effective execution tolerance is as follows: by obtaining historically verified qualified injection action trajectory samples, calculating the Euclidean distance set between the feature vector of each sample trajectory and the feature vector of the standard template, statistically calculating the arithmetic mean and standard deviation of the distance set, and selecting the value of the arithmetic mean plus a preset multiple of the standard deviation as the effective execution tolerance.
