An outpatient whole cycle intelligent follow-up and recheck reminding system

Abstract

This invention relates to the field of medical information technology, specifically disclosing an intelligent follow-up and re-examination reminder system for outpatients throughout their entire treatment cycle. Targeting the management of chronic brain diseases, this invention integrates personalized pharmacokinetic modeling and a cross-validation mechanism of subjective and objective data. By continuously collecting and analyzing patients' micro-movement and vital sign data, combined with an individualized drug concentration theoretical model, it achieves accurate assessment of changes in the patient's condition. Simultaneously, it utilizes subjective and objective cross-validation to ensure data reliability and adopts an event-driven model, transforming traditional periodic follow-up into real-time response management based on key pathological events. The system also features an intelligent re-examination scheduling function, which can automatically adjust the re-examination plan and generate emergency re-examination reminders based on changes in the patient's condition and drug efficacy assessment results, forming a fully automated closed-loop management process. This invention improves the accuracy, timeliness, and patient satisfaction of chronic brain disease management, bringing a new breakthrough to the field of medical information technology.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and more specifically, to an intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle. Background Technology

[0002] In the long-term management of chronic brain diseases such as epilepsy or Parkinson's disease, clinicians face the challenge of obtaining continuous, objective information about changes in a patient's condition between outpatient follow-ups. Daily fluctuations in the disease, especially mild or transient pathological signs, are crucial for assessing the effectiveness of current treatment regimens. However, the collection of this information relies primarily on the patient's subjective recollection and diary entries, which can be affected by memory biases, inaccurate descriptions, and poor adherence to record-keeping, thus limiting the data basis for physicians to precisely adjust medication dosages or optimize treatment strategies.

[0003] Currently, commonly used solutions in the industry include regular outpatient follow-ups, telephone follow-ups, and some AI-based follow-up plan generation systems. Some solutions introduce wearable devices to monitor patients' macroscopic physiological indicators, such as activity levels, sleep patterns, or heart rate, aiming to provide more objective health data. For example, patent publication number CN114023432A discloses a Parkinson's disease follow-up system that assesses patients by collecting their movement data; patent publication number CN112216361A proposes an AI-based follow-up plan generation method, device, terminal, and medium. These methods, to some extent, optimize the standardization and efficiency of the follow-up process.

[0004] However, the aforementioned traditional methods still have certain limitations when dealing with specific clinical problems. First, the fixed-cycle follow-up model may not be able to capture acute or subacute changes caused by fluctuations in drug efficacy or disease progression in a timely manner, potentially leading to intervention lag. Second, existing wearable device data applications are mostly focused on assessing general health status; the specificity of feature extraction and identification for subtle prodromal signs of specific brain diseases, such as mild absence seizures in epilepsy or early subtle manifestations of resting tremor in Parkinson's disease, needs improvement. Furthermore, existing technical solutions typically treat physiological sign monitoring and pharmacokinetic analysis as two separate processes, lacking a mechanism for real-time correlation analysis between the occurrence of objective signs and the theoretical blood drug concentration in the patient. Therefore, it is difficult to directly determine whether the occurrence of a symptom is directly related to the "wearing-off phenomenon" or insufficient drug concentration.

[0005] Based on the above problems, the present invention aims to solve the problem in the prior art that there is a lack of correlation and verification between objectively monitored weak pathological signs and individualized drug concentration theoretical models, so as to realize an event-driven precise follow-up visit reminder mechanism. Summary of the Invention

[0006] In view of this, in order to solve the problems mentioned in the background technology, a full-cycle intelligent follow-up and re-visit reminder system for outpatients is proposed.

[0007] The objective of this invention can be achieved through the following technical solution: This invention provides an intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, comprising: a pharmacokinetic modeling module, used to acquire historical digital prescription information and preset distributed pharmacokinetic database parameters, and calculate and generate a theoretical decay time axis of individual drug concentration.

[0008] The physiological feature extraction module is used to receive triaxial accelerometer signals and gyroscope angular velocity signals collected by the target wearable device, extract frequency domain and time domain features, and output micro-motion sequence feature stream.

[0009] The anomaly detection module is used to compare a preset database of early signs of brain diseases with a micro-movement sequence feature stream to generate an anomaly feature matrix and anomaly classification labels.

[0010] The beacon encapsulation module is used to encapsulate the anomaly feature matrix, anomaly classification labels, and extracted anomaly occurrence timestamps to generate micro-anomaly beacon data packets.

[0011] The questionnaire distribution module is used to parse micro-anomaly beacon data packets to extract anomaly classification tags, match them with a pre-set subjective question and answer questionnaire library, and assemble and issue instant question and answer confirmation instructions.

[0012] The cross-validation module receives subjective feedback confirmation signals and supplementary symptom description text in response to immediate rebuttal confirmation instructions, fuses the abnormal feature matrix, and extracts the timestamp of the abnormal occurrence to establish the effective confirmation time point of the abnormal event.

[0013] The drug efficacy correlation analysis module is used to map the effective confirmation time point of abnormal events to the theoretical decay time axis of individual drug concentration to obtain the current predicted blood drug concentration value, calculate the time difference from the single most recent target drug administration time point, and generate efficacy intervention early warning indicators when the time difference falls into the preset drug efficacy exhaustion period segment and the current predicted blood drug concentration value is lower than the preset lower limit blocking edge threshold.

[0014] The follow-up appointment scheduling module is used to respond to efficacy intervention early warning indicators. It injects the preset drug efficacy decay alarm code statement and supplementary symptom description text into the intercepted outpatient follow-up appointment data frame set, and generates follow-up reminder information with priority identifier.

[0015] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) By collecting continuous micro-movement sign data and comparing it with a preset brain disease precursor feature database, the present invention can use a sequence pattern recognition algorithm to detect events that are close to specific pathological patterns (such as epileptic minor absence or Parkinson's resting tremor) in the feature space in a continuous feature stream. This processing method can objectively quantify and classify brief and slight physical abnormalities that are difficult for the human eye to detect or that patients may ignore, thereby realizing the continuous and automated capture of subtle fluctuations in the condition that are difficult to cover in traditional follow-up models.

[0016] (2) This invention establishes a closed-loop confirmation logic by immediately sending an instant rebuttal confirmation command associated with the abnormality classification label to the patient after an objective abnormal event is detected, and by performing subjective-objective cross-validation using feedback response data. This mechanism binds the objective data points captured by the sensor with the patient's true subjective feelings at that moment. If the patient confirms the relevant feelings, the system marks this event as a high-confidence true positive abnormal event. This approach uses the patient's subjective feedback to filter out artifacts or interference that may be caused by non-pathological activities, improving the reliability of the event data on which subsequent clinical judgments are based.

[0017] (3) This invention maps the effective confirmation time point of the abnormal event established through cross-validation of subjective and objective factors onto the theoretical decay time axis of individual drug concentration generated based on digital prescriptions and pharmacokinetic parameters. This method can calculate the theoretical blood drug concentration of the patient when the abnormality occurs and determine whether it falls into the period of drug efficacy exhaustion. This correlation analysis mechanism establishes a complete evidence chain from "objective signs" to "subjective feelings" and then to "drug concentration", providing a quantitative basis for judging whether the fluctuation of symptoms is caused by drug efficacy decay, and realizing the dynamic evaluation of the effectiveness of drug treatment.

[0018] (4) This invention, by responding to efficacy intervention early warning indicators, forcibly reconstructs the original follow-up appointment logic pathway, generates and publishes follow-up reminder information with priority identifiers. This method can transform the traditional periodic follow-up mode based on fixed time into a dynamic intervention mode driven by key pathological events. When the system determines that the patient's symptom onset is highly correlated with insufficient drug efficacy, it can automatically trigger clinical recall actions, advancing the intervention timing from a fixed "checkpoint" after the fact to the "immediate point" when the problem occurs, realizing the transformation from passive response to active intervention in chronic disease management. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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.

[0020] Figure 1 This is a schematic diagram of the system module structure connection of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides an intelligent follow-up and re-visit reminder system for outpatients throughout their entire lifecycle, comprising: a pharmacokinetic modeling module, a physiological feature extraction module, an anomaly detection module, a beacon encapsulation module, a questionnaire distribution module, a cross-validation module, a pharmacodynamic correlation analysis module, and a re-visit scheduling module.

[0023] The physiological feature extraction module is connected to the anomaly detection module, the anomaly detection module is connected to the beacon encapsulation module, the beacon encapsulation module is connected to the questionnaire distribution module, both the beacon encapsulation module and the questionnaire distribution module are connected to the cross-validation module, both the pharmacokinetic modeling module and the cross-validation module are connected to the pharmacodynamic correlation analysis module, and the pharmacodynamic correlation analysis module is connected to the follow-up visit scheduling module.

[0024] The pharmacokinetic modeling module is used to acquire historical digital prescription information and preset distributed pharmacokinetic database parameters to calculate and generate the theoretical decay time axis of individual drug concentrations.

[0025] In a specific embodiment of the present invention, historical digital prescription information and preset distributed pharmacokinetic database parameters are obtained, and an individual drug concentration theoretical decay time axis is calculated and generated, including: parsing historical digital prescription information to obtain drug type identifier, single dose and medication timestamp set.

[0026] Half-life parameters and target concentration decay models are obtained by querying a distributed pharmacokinetic database based on drug type identifiers.

[0027] The set of single-dose dose and administration timestamp is substituted into the target concentration decay model, and the predicted blood drug concentration on the predicted sequence slice is calculated by combining the half-life parameter. The theoretical decay time axis of individual drug concentration is then fitted to generate the model.

[0028] Specifically, in one embodiment, this method is executed by a processing module deployed on the server side of a medical information system. First, the processing module accesses and reads the electronic medical record database for outpatient clinics specializing in brain diseases via an internal data bus. For a specific patient identifier, it retrieves all historical digital prescription information. This historical digital prescription information is a structured data record. The processing module calls a pre-defined parser, such as an XML-based parsing script, to extract key fields from the historical digital prescription information, parsing to obtain a unique drug type identifier, which can be in the form of a national drug code. Simultaneously, it extracts the single-dose dosage for each medication administration, quantifying it in mg. It then iterates through all medication records, compiling a set of medication timestamps containing all precise administration times. Next, the processing module uses the extracted drug type identifier as a query key to initiate a secure application programming interface (API) call to an external distributed pharmacokinetic database. The distributed pharmacokinetic database returns a set of pharmacokinetic parameters associated with drug class identifiers. The processing module receives and parses this parameter set, extracting the half-life parameter characterizing the drug's metabolic rate and a target concentration decay model describing the change in blood drug concentration over time. This model is typically a first-order kinetic elimination model function. Finally, the processing module performs an iterative calculation process. It traverses each discrete time point in the dosing timestamp set, taking each dosing time point as a starting point, and substituting the corresponding single-dose dose, bioavailability, and volume of distribution obtained from the distributed pharmacokinetic database into the target concentration decay model. The processing module combines the half-life parameter to calculate the elimination rate constant that determines the decay rate, and then calculates the predicted blood drug concentration on a series of predicted sequence slices after the single dosing event. By linearly superimposing the concentration decay curves generated by all dosing events in the dosing timestamp set, the processing module finally fits and generates a theoretical decay time axis of individual drug concentration stored in the form of a time series array, covering a preset future period.

[0029] In this embodiment, the total predicted blood drug concentration at any target time point t is calculated by superimposing the residual blood drug concentrations generated from each historical dosing. The total predicted blood drug concentration is given by the following formula:

[0030]

[0031] Among them, the elimination rate constant From half-life parameter The calculation shows that:

[0032]

[0033] In the above formula, This represents the total predicted blood drug concentration for an individual at the target time point t. n is the total number of drug administration events in the set of drug administration timestamps. Let be the single-dose dose for the i-th dosing event. F represents the bioavailability of the drug, a dimensionless proportionality coefficient. It is the apparent volume of distribution of the drug. This is the drug elimination rate constant. t is the target time point for concentration prediction. This is the specific timestamp of the i-th dosing event. Let Heaviside step function be the function of the Heaviside step function. Its value is 1 if it is true and 0 otherwise, ensuring that only the effects of past dosing events are calculated. The half-life parameter is obtained from a distributed pharmacokinetic database.

[0034] The dimensional consistency of the formula is guaranteed, where The unit is mass per volume, for example, mg / L. The right side of the equation... The terms have the same mass per volume dimension, while the exponential term... It is a dimensionless pure number because The unit is one-third of time, and The time units cancel each other out.

[0035] Further explanation of the bioavailability in the above formula The physical and pharmacological significance. In pharmacokinetics, not all drug doses orally ingested by the patient ( All of these can completely enter the systemic circulation and exert their therapeutic effect. Bioavailability It is a dimensionless percentage constant (its value ranges from...). The ratio (between) is used to quantify the effective drug dose that actually enters the bloodstream and the theoretical dose. It is determined directly from the drug instructions approved by the national drug regulatory authority or authoritative pharmacology databases. For example, for intravenously injected drugs... The value is set to ( (Enters the bloodstream); however, for conventional oral antiepileptic drugs (such as carbamazepine) or anti-Parkinson's disease drugs (such as levodopa), due to gastrointestinal absorption loss and the first-pass effect of the liver, their... The typical value is usually set at Between. By introducing into the numerator of the formula The system can accurately filter out unabsorbed and ineffective drug doses in the body, ensuring the accuracy of subsequent calculations of predicted blood drug concentrations. This accurately reflects the "effective circulating blood drug concentration" that can cross the blood-brain barrier and inhibit subtle signs of brain aberrations, thereby improving the theoretical accuracy of pharmacodynamic correlation analysis.

[0036] To further explain in detail the Heaviside step function introduced in the above formula The mathematical functions and time-gated logic of the drug. In the calculation of concentration decay superposition for multiple consecutive doses, it is essential to ensure that "future dosing events do not affect past blood drug concentration calculations." Helvetica step function. This acts as a strict "time causality switch": when evaluating the target time point... Earlier than the Timing of the next dose Time (i.e.) ), Force output as At this point, the formula term is cleared to zero, meaning that no premature drug administration event will occur. When evaluating the target time point... Later than or equal to the Timing of the next dose Time (i.e.) ), The output is The formula is activated normally, and the calculation of the residual concentration decay resulting from this medication administration begins. This is achieved by embedding the formula within the summation formula. The system can automatically filter out all invalid forward interference on the timeline at the mathematical level, allowing the processing module to uniformly substitute all prescription records throughout the patient's lifespan into the matrix for parallel superposition operations without the need to write complex time truncation judgment statements. This optimizes the computational logic and improves the throughput of batch data processing.

[0037] Historical digital prescription information refers to standardized electronic records stored in hospital information systems or electronic medical record systems. These records adhere to medical information exchange standards such as HL7 or CDA and include patients' medication orders and administration records. Drug type identifiers are unique codes assigned to specific drug formulations according to national or international standards, such as the National Medical Products Administration's "National Drug Approval Number," used to accurately distinguish drugs from different manufacturers, dosage forms, and specifications. The medication timestamp set is a data list consisting of date and time strings conforming to ISO 8601 standards, accurately recording the time of each drug administration. The distributed pharmacokinetic database is a professional database service accessible via a network interface, storing and providing pharmacokinetic parameters for thousands of drugs in standard human models, such as data related to absorption, distribution, metabolism, and excretion. Half-life parameters... This is a core parameter in pharmacokinetics, defined as the time required for the drug concentration in the body to decrease by half. Its definition is based on extensive clinical trial statistics and is included in authoritative pharmaceutical databases. In this embodiment, the targeted concentration decay model specifically refers to the one-compartment open model. This model assumes that the drug is instantaneously and uniformly distributed in a single compartment in the body and eliminated at a first-order kinetic rate. It is a classic mathematical model describing the process of most drugs in the body. The theoretical decay time axis of individual drug concentration is a two-dimensional data structure. Its first dimension is a discrete time sequence over a period of time from the current time point, and the second dimension is the predicted blood drug concentration value corresponding to each of these time points, together forming a continuously changing theoretical curve.

[0038] The physiological feature extraction module is used to receive triaxial accelerometer signals and gyroscope angular velocity signals collected by the target wearable device, extract frequency domain and time domain features, and output micro-motion sequence feature stream.

[0039] In a specific embodiment of the present invention, receiving triaxial accelerometer signals and gyroscope angular velocity signals collected by the target wearable device, extracting frequency domain and time domain features, and outputting a micro-motion sequence feature stream includes: performing sliding window segmentation and filtering noise reduction processing on the triaxial accelerometer signals and gyroscope angular velocity signals to obtain continuous time window signals.

[0040] The frequency domain energy concentration is calculated by performing a Fast Fourier Transform on the continuous time window signal, and the time domain variance numerical features are extracted simultaneously.

[0041] By splicing the frequency domain energy concentration and time domain variance numerical characteristics, a micro-motion sequence feature stream is output.

[0042] Specifically, this method is executed by a data acquisition and preprocessing module running on a patient-side smart device, such as a smartphone or a proprietary gateway. First, the module establishes a stable real-time communication link with a pre-paired target wearable device via Bluetooth Low Energy. The target wearable device has a built-in microelectromechanical system (MEMS) inertial measurement unit (IMU), to which the module sends a subscription command, continuously receiving triaxial accelerometer and gyroscope angular velocity signals acquired by the device at a preset sampling frequency, e.g., 100Hz. These signals are transmitted to the patient-side smart device in real-time as a byte stream. Subsequently, the preprocessing module performs sliding window segmentation and filtering noise reduction on the received continuous raw signal stream. It defines a fixed-length time window, e.g., 2.56 seconds, and an overlap rate, e.g., 50%, i.e., a step size of 1.28 seconds. For each data slice within the time window, the module first applies a fourth-order Butterworth low-pass filter with a cutoff frequency of 25Hz to filter out high-frequency noise; then it applies a high-pass filter with a cutoff frequency of 0.5Hz to remove the gravity component from the triaxial accelerometer signal. After the above processing, the original signal stream is transformed into a series of clean and length-normalized continuous time window signals. Each continuous time window signal is a numerical matrix containing 256 sampling points and 6 channels. Finally, for each continuous time window signal, the feature extraction unit performs two calculations in parallel. First, it applies the Fast Fourier Transform algorithm to the signals of each of the 6 channels to calculate their power spectral density, and calculates the ratio of energy in a specific frequency band, such as 3Hz to 8Hz, to the total energy of the entire frequency band through integration, obtaining a scalar value as the frequency domain energy concentration. Second, it simultaneously calculates the statistical variance of the 256 sampling points in each channel of the continuous time window signal in the time domain, obtaining a six-dimensional vector as the time domain variance numerical feature. The feature extraction unit concatenates this scalar value and the six-dimensional vector along the feature dimension to form a seven-dimensional feature vector. As the sliding window continues to move, this process is repeated continuously, thereby outputting a time-continuous micro-motion sequence feature stream.

[0043] In this embodiment, frequency domain energy concentration The calculation method is as follows:

[0044]

[0045] in, It is the frequency domain energy concentration of a single channel signal. In practical applications, the average value of 6 channels can be calculated or it can be used as a multidimensional feature. It is the power spectral density of the signal at frequency f, calculated using the Welch method. and These represent the lower and upper limits of the target frequency band, respectively. This refers to the sensor's sampling frequency. The time-domain variance numerical eigenvector. Each component in The calculation is as follows:

[0046]

[0047] In the above formula, N is the number of sampling points in each time window. This represents the amplitude at the k-th sampling point in the j-th channel. It is the arithmetic mean of all N sampling points in the j-th channel.

[0048] The dimensional consistency of the formula is guaranteed. In the calculations, both the numerator and denominator are in units of power (e.g., ...). Multiplying this by the frequency unit (Hz), which is the energy unit, and dividing by the frequency unit, yields a dimensionless ratio. In the calculation, its dimension is the square of the original signal amplitude dimension. For example, for an acceleration signal, it is... For angular velocity signals, .

[0049] Wearable devices, such as smartwatches or medical-grade wearable sensors, are electronic devices integrating a three-axis accelerometer and a three-axis gyroscope, capable of acquiring and wirelessly transmitting data at a frequency of at least 50Hz. A continuous time window signal refers to a fixed-length segment extracted from a continuous sensor data stream, containing local motion characteristics of the patient's limbs within that time period. Fast Fourier Transform (FFT) is a computational algorithm that converts time-domain signals to the frequency domain, used to analyze the frequency composition of the signal. Frequency domain energy concentration is an indicator used to quantify the proportion of energy in a specific physiological activity frequency band. For example, the resting tremor energy in Parkinson's disease is mainly concentrated in the 4Hz to 6Hz frequency band; this parameter is based on existing clinical research on movement disorders in specific brain diseases. The time-domain variance numerical characteristic is a quantitative indicator reflecting the intensity of signal fluctuations within a time window; a larger value usually indicates greater amplitude or intensity of movement. A micro-movement sequence feature stream is a time-series data structure composed of a series of multi-dimensional feature vectors arranged in chronological order, where each vector is a quantitative description of the patient's micro-movements within a small time window.

[0050] The anomaly detection module is used to compare a preset database of early signs of brain diseases with a micro-movement sequence feature stream to generate an anomaly feature matrix and anomaly classification labels.

[0051] In a specific embodiment of the present invention, an abnormal feature matrix and an abnormal classification label are generated by comparing a preset brain disease precursor feature library with a micro-movement sequence feature stream. This includes: using a sequence pattern recognition algorithm to calculate the clustering distance between the micro-movement sequence feature stream and the baseline classification plane in the preset brain disease precursor feature library.

[0052] When the shortest comparison distance is less than the preset anomaly detection threshold, the continuous time feature vectors are extracted and stacked to generate an anomaly feature matrix.

[0053] The abnormal feature matrix is ​​labeled with abnormal classification labels that represent physiological feature patterns.

[0054] Specifically, this method is executed by an anomaly detection module deployed on the server. First, during the initialization phase, this module loads a preset feature library of early signs of brain diseases from the local file system or configuration database. This library is stored in key-value pairs, where the key is a string identifier for the pathological type, and the value is the corresponding multidimensional feature reference vector, such as a reference vector for epilepsy minor absence seizures and a reference vector for Parkinson's resting tremor. Subsequently, the anomaly detection module receives and processes the micro-movement sequence feature stream generated above in real time. For each newly arriving feature vector in this sequence stream, the module executes a sequence pattern recognition algorithm. The core of this algorithm is to traverse all reference vectors in the preset feature library of early signs of brain diseases and calculate the clustering distance between the current feature vector and each reference vector using a multidimensional spatial distance calculation formula. After calculating the distances for all references, the module determines the shortest alignment distance and its corresponding reference vector. Next, the system compares this shortest alignment distance with a preset anomaly judgment threshold. The system determines a potential pathological micro-movement event has occurred only when the detected shortest alignment distance is numerically less than a preset anomaly threshold. At this point, the system immediately performs a forced truncation operation, extracting the N most temporally consecutive feature vectors, including the current trigger vector, from its internally cached time-series data. These N feature vectors are then vertically stacked in chronological order, forming an N-row, 7-column anomaly feature matrix. Finally, the module uses the pathological type string identifier associated with the baseline vector generating the shortest alignment distance as the anomaly classification label, and binds this label to the newly generated anomaly feature matrix, forming a composite data structure output to subsequent processing steps.

[0055] In this embodiment, the clustering distance comparison between the current feature vector and the preset benchmark is calculated using the Euclidean distance formula. It is assumed that the current feature vector obtained from the micro-motion sequence feature stream is... The j-th baseline vector in the pre-defined brain disease precursor feature database is... Then the distance between them The calculation is as follows:

[0056]

[0057] The triggering conditions for an abnormal event can be represented as:

[0058]

[0059] In the above formula, It is the current D-dimensional feature vector obtained from the feature stream of the micro-action sequence. It is the j-th D-dimensional baseline vector stored in the pre-defined brain disease precursor feature database. D is the dimension of the feature vector, and according to the aforementioned output, D equals 7. and They are vectors and The kth component. It is the calculated scalar distance value. This indicates that the minimum value is taken among all j benchmarks. It is a preset anomaly detection threshold.

[0060] The consistency of the dimensions of the formula is guaranteed. Since the components of the feature vector may have different physical dimensions, before constructing the preset brain disease precursor feature library and performing real-time comparison, all feature components are usually standardized, such as Z-score standardization, to make them dimensionless values, thereby ensuring that the calculation of Euclidean distance is physically reasonable.

[0061] The preset brain disease prodromal feature library is a knowledge base generated through machine learning training on wearable device data from a large number of diagnosed patients during specific symptom episodes. Each benchmark in the library, such as the epileptic minor absence benchmark, is a centroid vector obtained by clustering the feature vectors of hundreds of minor absence seizure events, representing the most typical digital feature expression of this type of event. In this embodiment, the sequence pattern recognition algorithm specifically refers to an online, vector-wise pattern matching process. Unlike the analysis of the entire long sequence, it focuses on the similarity between instantaneous features and standard patterns. A preset anomaly detection threshold is also included. This is a key control parameter, its value set based on receiver operating characteristic curve analysis performed on a validation dataset, aiming to balance detection sensitivity and specificity. Assuming the application scenario requires high sensitivity to avoid false negatives, this threshold might be set in the range of 0.8 to 1.2 after all features have been normalized. The anomaly feature matrix is ​​a two-dimensional array that contains not only the key feature vector that triggered the decision but also its immediately adjacent context feature vectors, forming a complete snapshot of the micro-event. Its number of rows, N, is typically set to 5 to 10 to capture a complete micro-movement process lasting several seconds. The anomaly classification label is a standardized string used for subsequent logical branching in the system.

[0062] In a specific embodiment of the present invention, after the step of labeling the abnormal feature matrix with abnormal classification labels that refer to physiological feature patterns, the method further includes: performing principal component analysis dimensionality reduction on the abnormal feature matrix and extracting principal component feature vectors with a cumulative contribution rate reaching a preset ratio.

[0063] The principal component feature vectors and anomaly classification labels are input into a pre-defined graph neural network model to reconstruct the topological relationship, and the output is a graph structure anomaly feature matrix with spatial correlation weights.

[0064] When generating micro-anomaly beacon data packets, the abnormal feature matrix is ​​replaced with a graph structure abnormal feature matrix before entering the encapsulation process.

[0065] Specifically, after obtaining the labeled anomaly feature matrix, the engineering implementation mechanism and technical objectives of performing principal component analysis (PCA) dimensionality reduction and graph neural network topology reconstruction are as follows: PCA dimensionality reduction: The micro-motion sequence feature stream captured by wearable devices often contains severe information redundancy. For example, the triaxial acceleration of the left wrist and the gyroscope angular velocity are highly linearly correlated in certain tremor modes. The system uses singular value decomposition (SVD) to solve for the eigenvalues ​​and eigenvectors of the covariance matrix of the anomaly feature matrix. After sorting the eigenvalues ​​in descending order, the system extracts the top few eigenvectors whose cumulative contribution rate reaches a preset proportion (e.g., set to 90% or 95%), combining them into new "principal component eigenvectors". This step compresses the originally high-dimensional and collinear feature data into a low-dimensional and orthogonal optimal feature combination, not only eliminating white noise interference but also significantly reducing the computational complexity of the subsequent graph network.

[0066] Graph Neural Network Model Topology Reconstruction: Traditional feature matrices treat micro-actions at different times as isolated points, ignoring their spatiotemporal evolution relationships. To address this issue, the system uses principal component feature vectors, reduced in dimensionality through principal component analysis, as nodes in the graph structure. "Anomaly classification labels (representing specific physiological characteristic pattern rules)" are used as initial edge attributes connecting each node, and input into a pre-trained graph convolutional network or graph attention network model. The graph neural network calculates and learns the intrinsic correlations and spatiotemporal influence weights between nodes through a multi-layer message passing mechanism. The final output graph structure anomaly feature matrix is ​​no longer a simple two-dimensional array, but a complex topological data structure containing node features and their "spatial correlation weight matrix."

[0067] Technical Effects and Replacement: When generating micro-anomaly beacon data packets, the original anomaly feature matrix is ​​replaced with this graph-structured anomaly feature matrix. Its core advantage lies in the fact that the graph-structured matrix elevates the simple "action record" to a "pathological action evolution network" with contextual relationships. This allows the server to not only see "when the patient had an attack" during subsequent cross-validation or when doctors review diagnoses, but also to intuitively analyze "how micro-movements spread and are transmitted across different physiological dimensions" through graph network weights. This provides highly in-depth graph-based criteria for accurate follow-up diagnoses and dosage adjustments.

[0068] The beacon encapsulation module is used to encapsulate the anomaly feature matrix, anomaly classification labels, and extracted anomaly occurrence timestamps to generate micro-anomaly beacon data packets.

[0069] In a specific embodiment of the present invention, an abnormal feature matrix, an abnormal classification label, and an extracted abnormal occurrence timestamp are encapsulated to generate a micro-anomaly beacon data packet, including: reading system clock state parameters and converting them into non-decreasing abnormal occurrence timestamps.

[0070] The timestamps of anomalies, anomaly classification labels, and anomaly feature matrices are compiled into the core data payload.

[0071] Add a data check bit and a communication identifier header to the outer end of the core data payload, and perform multi-layer structured encapsulation and compression to generate micro-anomaly beacon data packets.

[0072] Specifically, this method is executed by a data encapsulation module running on a server or edge computing node. Upon receiving the aforementioned anomaly feature matrix and anomaly classification labels, the module immediately triggers the encapsulation process. First, it calls the operating system kernel interface to read the state of the underlying system clock. This system clock is synchronized with an authoritative time server via a network time protocol to ensure high accuracy and consistency. The module formats the read time value, such as a combination of a Unix timestamp and milliseconds, converting it into a string with millisecond precision conforming to the ISO 8601 standard. This string is then established as the non-decreasing anomaly occurrence timestamp. Next, the module performs structured assembly on the anomaly occurrence timestamp, anomaly classification labels, and anomaly feature matrix as a whole. Based on a predefined JSON data pattern, it creates a data object containing three key-value pairs, with each key corresponding to one of the three data items mentioned above, thereby generating a text-formatted core data payload. Subsequently, the module performs additional encapsulation operations on the core data payload. It queries the system's device management table for a unique device identifier associated with the current data source, using this as the terminal device communication header. Simultaneously, it applies a cyclic redundancy check algorithm, such as CRC32, to the serialized core data payload byte stream to generate a 32-bit data check bit. Finally, the module inputs the core data payload byte stream into a general data compression library, such as zlib, and performs compression using the DEFLATE algorithm. It concatenates the terminal device communication header, data check bit, and compressed data body into a continuous binary data block according to a predetermined data frame format, i.e., [header][check bit][compressed data body], thereby generating a compact micro-anomaly beacon data packet with complete metadata, and pushes it into the network message queue to be sent.

[0073] The anomaly timestamp is a millisecond-accurate timestamp used to uniquely pinpoint the moment a micro-anomaly event was objectively detected on the global timeline of the entire system. The core data payload is the primary information carrier before encapsulation; in this embodiment, it is a JSON-formatted string that organizes the event's time, type, and original data in a readable manner for easy subsequent parsing and debugging. The data checksum is a check code calculated from the core data payload, used at the data receiving end to verify whether bit errors occurred in the micro-anomaly beacon data packet during transmission, ensuring communication reliability. The terminal device communication identifier header is a globally unique identifier, such as a UUID, assigned when the patient device first registers with the system. It ensures that downlink commands can be accurately routed back to the specific patient device that generated the beacon. The micro-anomaly beacon data packet is the final binary data unit generated in this step, used for transmission in the network. Its multi-layered structured design balances route identification, data verification, and transmission efficiency.

[0074] The questionnaire distribution module is used to parse micro-anomaly beacon data packets to extract anomaly classification tags, match them with a pre-set subjective question and answer questionnaire library, and assemble and issue instant question and answer confirmation instructions.

[0075] In a specific embodiment of the present invention, parsing micro-anomaly beacon data packets to extract anomaly classification tags, matching them with a preset subjective question and answer questionnaire library, and assembling and issuing an instant question and answer confirmation instruction includes: disassembling micro-anomaly beacon data packets to extract anomaly classification tags as method index pointers.

[0076] Traverse the pre-defined subjective question and answer questionnaire tree nodes and retrieve multiple subjective feeling confirmation questions based on the method index pointer.

[0077] Multiple subjective feeling confirmation questions are assembled and merged into an immediate question confirmation instruction with a high-priority immediate response identifier, and then executed and issued.

[0078] Specifically, this method is executed by a server-side rhetoric generation module. This module continuously monitors a network message queue, and the processing flow is activated when a micro-anomaly beacon data packet generated earlier is retrieved from the queue. First, the module performs decomposition and analysis on the received micro-anomaly beacon data packet. According to the predetermined data frame format, it sequentially reads and temporarily stores the terminal device communication identifier header, and then reads the data checksum. The module applies the same cyclic redundancy check algorithm as during encapsulation to the remaining compressed data body in the data packet, comparing the calculated new checksum with the read data checksum. If they do not match, the data packet is discarded; if they match, the decompression library is called to decompress the compressed data body, restoring the original core data payload. Subsequently, the module uses a JSON parser to analyze the core data payload, extracting anomaly classification labels with string values. This module directly uses the anomaly classification labels as method index pointers for targeted correlation queries. Next, the module uses the method index pointers to traverse and query a server-side globally pre-built subjective rhetoric questionnaire library. The subjective question and answer questionnaire database is constructed as a hierarchical tree-like data structure, where the root node represents the disease category and the child nodes represent specific symptoms or subtypes. The module parses the pathological path based on the semantic content of the method index pointer, such as from "basic criteria for epilepsy and mild absence seizures," strictly matches and locates the corresponding leaf node in the tree structure, and retrieves pre-stored multiple subjective feeling confirmation questions that have the highest pathological logical relevance to the abnormality classification label from that node. Finally, the module assembles the extracted multiple subjective feeling confirmation questions into a structured instruction data body, such as a JSON array containing a list of questions, and attaches a high-priority immediate response identifier to this instruction data body. This assembly constitutes the immediate question and answer confirmation instruction, which the module precisely pushes to the user interface of the patient end, specified by the previously temporarily stored terminal device communication identifier header, through a priority downlink communication link, such as calling a high-priority message push service interface.

[0079] The method index pointer is a specific designation for the anomaly classification label function in this step, clarifying its purpose as a query key for the database or knowledge base. The subjective question-and-answer questionnaire database is a pre-built knowledge base by medical experts based on clinical guidelines and diagnostic criteria. It maps objectively measurable micro-movement patterns to subjective feelings that patients may experience during an attack. For example, for objective patterns of absence seizures in epilepsy, the database stores highly relevant confirmation questions such as "Do you feel a sudden interruption of thought?" or "Do you unconsciously blink or smack your lips?" Its tree-like node structure makes queries efficient and scalable. Subjective feeling confirmation questions are specific question texts stored in this questionnaire database, designed to guide patients to confirm or deny specific physiological or psychological feelings at a specific moment in a simple and unambiguous manner. The immediate question-and-answer confirmation instruction is the final output of this step. It is a data package containing specific question content and metadata. Its high-priority immediate response identifier is a Boolean or enumerated field used to inform the receiving terminal that this message needs to be processed immediately and presented to the user, rather than entering a regular notification queue.

[0080] The cross-validation module receives subjective feedback confirmation signals and supplementary symptom description text in response to immediate rebuttal confirmation instructions, fuses the abnormal feature matrix, and extracts the timestamp of the abnormal occurrence to establish the effective confirmation time point of the abnormal event.

[0081] In a specific embodiment of the present invention, receiving a subjective feedback confirmation signal and supplementary symptom description text in response to an immediate rebuttal confirmation command, fusing an abnormal feature matrix, and extracting an abnormal occurrence timestamp to establish an effective confirmation time point for the abnormal event includes: activating a hardware timer to establish a countdown listening thread, and listening for the subjective feedback confirmation signal and supplementary symptom description text.

[0082] When capturing subjective feedback confirmation signals within a preset tolerance time window, the supplementary symptom description text and the abnormal feature matrix are fused and associated with a database table.

[0083] The timestamp of the anomaly occurrence extracted from the micro-anomaly beacon data packet is established as the valid confirmation time point of the anomaly event.

[0084] Specifically, this method is initiated and executed immediately by the server-side cross-validation control module after the immediate feedback confirmation command is issued. First, this module calls a system bus hardware timer via a low-level API and establishes a countdown listening thread with a limited duration based on this timer. The duration of the countdown listening thread is set to a preset tolerance time window. After starting, this thread continuously listens for a network socket bound to a specific patient's device to capture the subjective feedback confirmation signal returned by the patient's click in response to the immediate feedback confirmation command, along with any accompanying supplementary symptom description text. The subjective feedback confirmation signal is a data packet encapsulating a Boolean confirmation flag and text content. Second, if the countdown listening thread successfully captures the subjective feedback confirmation signal before the preset tolerance time window expires, the system's core logic parses the signal content. If the Boolean confirmation flag in the signal is true, the system determines that the subjective feedback fully confirms the previously detected minor objective anomaly. At this point, the system executes a database transaction, fusing and associating the received supplementary symptom description text with the abnormal feature matrix that triggered this interaction in the database tables. For example, a new entry is created in the event log table, storing both of the above data items and establishing a relationship through a foreign key. The instant this operation completes, the system records and marks a true positive abnormal event detected in this entry. Finally, the verification control module extracts the original record's anomaly occurrence timestamp from the payload of the micro-anomaly beacon data packet cached in memory and associated with the current transaction. This timestamp is the physical time stamp of the objective event. Confirmed by the core processor, this timestamp is formally established as the valid confirmation time point of the abnormal event, verified through a two-way interlock between subjective and objective mechanisms, and persistently stored as the final result in the event log table for subsequent steps. If the listening times out or a negative confirmation signal is received, the event is marked as unconfirmed and archived.

[0085] Hardware timers, provided by the central processing unit or motherboard chipset, offer higher precision and lower system load compared to application-layer software timers. A countdown monitoring thread is a specially designed concurrent execution unit dedicated to network I / O monitoring within its defined lifecycle, automatically terminating upon completion of the countdown, thus achieving efficient resource management. The preset tolerance time window is a key time parameter, typically ranging from 60 to 180 seconds. This range is based on a trade-off between human-computer interaction research and clinical feedback effectiveness, aiming to ensure patients have sufficient time to respond while maintaining peak memory of the event. A subjective feedback confirmation signal is a structured data packet containing at least a Boolean field indicating user confirmation and a string field carrying supplementary symptom descriptions input by the user. A true positive abnormal event is a system-level defined status flag or record type, representing a highly clinically reliable fluctuation in a patient's condition that is supported by objective data and confirmed subjectively and immediately by the patient. Database table fusion and association is a data persistence strategy that associates primary keys of records derived from objective monitoring with primary keys of records derived from subjective feedback in a relational database, thus solidifying the one-to-one correspondence between objective and subjective data in a structured manner. The effective confirmation time of the anomaly event is the key timestamp output in this step; its value is strictly equal to the generated anomaly occurrence timestamp, representing the moment when the verified anomaly event actually occurred in the physical world.

[0086] The drug efficacy correlation analysis module is used to map the effective confirmation time point of abnormal events to the theoretical decay time axis of individual drug concentration to obtain the current predicted blood drug concentration value, calculate the time difference from the single most recent target drug administration time point, and generate efficacy intervention early warning indicators when the time difference falls into the preset drug efficacy exhaustion period segment and the current predicted blood drug concentration value is lower than the preset lower limit blocking edge threshold.

[0087] In a specific embodiment of the present invention, the effective confirmation time point of the abnormal event is mapped to the theoretical decay time axis of individual drug concentration to obtain the current predicted blood drug concentration value, and the time difference value from the single most recent target drug administration time point is calculated. When the time difference value falls into the preset drug efficacy exhaustion period segment and the current predicted blood drug concentration value is lower than the preset lower limit blocking edge threshold, an efficacy intervention warning indicator is generated, including: aligning and mapping the effective confirmation time point of the abnormal event to the theoretical decay time axis of individual drug concentration to extract the current predicted blood drug concentration value.

[0088] Retrieve historical digital prescription information to obtain the most recent target medication time point, and calculate the time difference between the effective confirmation time point of the abnormal event and the most recent target medication time point.

[0089] When the time difference falls into the preset period of drug efficacy depletion and the current predicted blood drug concentration is lower than the preset lower limit of the blocking edge threshold, an efficacy intervention early warning indicator is generated.

[0090] Specifically, this method is executed by the efficacy evaluation and early warning module deployed on the server. First, this module obtains the valid confirmation time point of the abnormal event and simultaneously retrieves the theoretical decay time axis of the individual drug concentration. It uses the valid confirmation time point of the abnormal event as input and performs precise alignment mapping along the time-domain scale dimension of the theoretical decay time axis of the individual drug concentration. Specifically, the module finds the time point closest to the valid confirmation time point of the abnormal event in the time axis data sequence and calculates the current predicted blood drug concentration value corresponding to that exact moment using linear interpolation. Next, the module globally traverses and retrieves historical digital prescription information stored in memory or a database, especially the set of medication timestamps. It compares all timestamps in this set with the valid confirmation time point of the abnormal event, filters out all timestamps earlier than that confirmation time point, and selects the one with the closest time distance as the single nearest target medication time point. Subsequently, the module performs time arithmetic operations to calculate the time difference between the valid confirmation time point of the abnormal event and the single nearest target medication time point. Finally, a core verification system performs a dual-condition judgment. First, it determines whether the calculated time difference falls within the drug's efficacy depletion period, pre-defined in the pharmacology database and associated with the current drug. Simultaneously, it verifies whether the current predicted blood drug concentration obtained from the timeline mapping is already below a preset lower limit threshold for drug inhibition. Only when both conditions are met simultaneously does the system automatically extract a set of key parameters, including the time difference and the current predicted blood drug concentration, and encapsulate them into a structured data object, generating an efficacy intervention early warning indicator to trigger subsequent clinical intervention logic.

[0091] In this embodiment, the formula for calculating the time difference is:

[0092]

[0093] Efficacy intervention early warning indicators The generation logic can be defined by the following conditional expression:

[0094]

[0095] In the above formula, This represents the calculated time difference. It is the effective confirmation time point of the acquired abnormal event. It is the single most recent target medication time point found from the set of medication timestamps. It is a Boolean flag; a value of 1 indicates the generation of an efficacy intervention early warning indicator, while a value of 0 indicates that no indicator is generated. and These are the lower and upper limits of the preset drug efficacy depletion period, respectively. Is The current predicted blood drug concentration value is obtained by querying or interpolating at any time. It is a preset lower limit blocking edge threshold.

[0096] The predicted blood drug concentration refers to the theoretical drug concentration in the patient's body calculated based on a pharmacokinetic model at the time of the confirmed adverse event. The most recent target dosing time is the precise time record of the last medication administration performed by the patient before the adverse event occurred. The time difference is a quantitative indicator representing the length of time elapsed from the most recent medication administration to the occurrence of the adverse event. The drug exhaustion period is a pre-defined time window based on the pharmacodynamic characteristics of a specific drug (such as onset time, peak time, and duration of action). It defines the time range within which the drug's effect is expected to begin to significantly weaken; this data is derived from the drug's package insert or a clinical pharmacology database. The lower limit of blockade threshold, also known as the minimum effective concentration, refers to the lowest concentration of a drug required in plasma to produce a therapeutic effect. Below this threshold, the drug is considered unable to effectively control symptoms. This value is also an inherent pharmacological property of the drug and is determined by authoritative clinical data. The efficacy intervention early warning indicator is the core output of this step. It is a composite data object containing the warning level, triggering reason, and relevant data snapshots, used to clearly indicate the need for clinical intervention.

[0097] The follow-up appointment scheduling module is used to respond to efficacy intervention early warning indicators. It injects the preset drug efficacy decay alarm code statement and supplementary symptom description text into the intercepted outpatient follow-up appointment data frame set, and generates follow-up reminder information with priority identifier.

[0098] In a specific embodiment of the present invention, in response to the efficacy intervention warning indicator, a preset drug efficacy decay alarm code statement and supplementary symptom description text are injected into the intercepted outpatient follow-up appointment data frame set to generate follow-up reminder information with priority identifier, including: intercepting the outpatient follow-up appointment data frame set output by the underlying follow-up cycle scheduling logic.

[0099] The command bit of the response efficacy intervention warning indicator is flipped, and the preset drug efficacy decay alarm code statement is merged with the supplementary symptom description text and dynamically injected into the abnormal pop-up field area of ​​the outpatient follow-up appointment data frame set.

[0100] Perform a package sealing operation on the data frame sequence to generate follow-up visit reminder information with priority indicators.

[0101] Specifically, this method is triggered by the server-side follow-up appointment scheduling and information generation module in response to efficacy intervention warning indicators. First, this module uses a message subscription mechanism to monitor in real-time the data stream output by a low-privilege work queue from a low-level routine follow-up cycle scheduling logic. When this scheduling logic generates a set of outpatient follow-up appointment data frames according to a preset routine cycle (e.g., every three months), the module intercepts and temporarily stores it, suspending its direct push to downstream systems. Next, when the module receives the aforementioned generated efficacy intervention warning indicator, it checks whether the instruction bits in the indicator have been effectively flipped, confirming that the warning status is active. Once confirmed, the module performs a dynamic content injection operation. It concatenates or structurally merges a pre-compiled alarm code statement (used by developers and stored in the system resource library) indicating drug efficacy decay with the obtained supplementary symptom description text. Subsequently, the module forcibly injects this fused information body dynamically into a specific field area of ​​the previously intercepted outpatient follow-up appointment data frame set; this area is defined as the abnormal pop-up field with the highest response level. Finally, after completing the field injection modification, the module performs final wrapping and end-capping operations on the corresponding data frame sequence. This includes adding the necessary communication protocol headers and trailers and converting the entire sequence into a priority-based follow-up reminder message with a specific display hierarchy. Ultimately, the module pushes this message simultaneously to the relevant doctor's workstation diagnostic analysis and monitoring list and the patient's mobile application reminder display area via a two-way communication channel, thus publishing the priority-based follow-up reminder message. This process achieves intelligent judgment based on objective data and subjective feedback, upgrading routine periodic follow-ups to event-driven precision intervention, thereby achieving automated full-cycle follow-up control and an efficient follow-up visit closed loop for chronic disease management.

[0102] The outpatient follow-up appointment data frame set consists of standardized data packets generated by traditional follow-up systems based on fixed time intervals. Its content typically only includes suggested follow-up times and department information. The drug efficacy decay alarm code statement is a pre-defined text or code snippet with a specific format and highlighting, such as a piece of HTML code.<spanstyle='color:red; font-weight:bold;'> The "[Drug Efficacy Decay Alert]" is used to create a prominent visual effect on the doctor's and patient's interfaces. The "Abnormal Pop-up Field" area is a reserved data segment in the data frame structure used to carry urgent or non-standard information; the receiving application is designed to prioritize parsing and displaying the content of this field. Specific priority display hierarchy is a front-end interface display strategy that ensures that follow-up appointment reminders with priority indicators appear at the top, in a modal dialog box, or as a system-level notification, regardless of the application scenario, and cannot be easily ignored by the user. A two-way communication channel is a communication strategy that ensures that critical information can be delivered to both doctors and patients synchronously and without delay, avoiding intervention delays caused by information asymmetry.

[0103] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A smart follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, characterized in that, include: The pharmacokinetic modeling module is used to acquire historical digital prescription information and preset distributed pharmacokinetic database parameters to calculate and generate the theoretical decay time axis of individual drug concentrations. The physiological feature extraction module is used to receive triaxial accelerometer signals and gyroscope angular velocity signals collected by the target wearable device, extract frequency domain and time domain features, and output micro-motion sequence feature stream. An anomaly detection module is used to compare a preset brain disease precursor feature library with a micro-movement sequence feature stream to generate an anomaly feature matrix and anomaly classification labels. The beacon encapsulation module is used to encapsulate the anomaly feature matrix, anomaly classification labels, and extracted anomaly occurrence timestamps to generate micro-anomaly beacon data packets; The questionnaire distribution module is used to parse micro-anomaly beacon data packets to extract anomaly classification tags, match them with a preset subjective question and answer questionnaire library, and assemble and issue instant question and answer confirmation instructions. The cross-validation module is used to receive subjective feedback confirmation signals and supplementary symptom description text in response to immediate question confirmation instructions, fuse abnormal feature matrices, extract abnormal occurrence timestamps, and establish valid confirmation time points for abnormal events. The drug efficacy correlation analysis module is used to map the effective confirmation time point of abnormal events to the theoretical decay time axis of individual drug concentration to obtain the current predicted blood drug concentration value, calculate the time difference from the single most recent target drug administration time point, and generate efficacy intervention early warning indicators when the time difference falls into the preset drug efficacy exhaustion period segment and the current predicted blood drug concentration value is lower than the preset lower limit blocking edge threshold. The follow-up appointment scheduling module is used to respond to efficacy intervention early warning indicators. It injects the preset drug efficacy decay alarm code statement and supplementary symptom description text into the intercepted outpatient follow-up appointment data frame set, and generates follow-up reminder information with priority identifier.

2. The intelligent follow-up and revisit reminder system for outpatients throughout their entire treatment cycle, as described in claim 1, is characterized in that... The process of acquiring historical digital prescription information and preset distributed pharmacokinetic database parameters to calculate and generate a theoretical decay time axis for individual drug concentrations includes: Analyze historical digital prescription information to obtain drug type identifiers, single-dose dosages, and medication timestamp sets; Half-life parameters and target concentration decay models are obtained by querying a distributed pharmacokinetic database based on drug type identifiers. The set of single-dose dose and administration timestamp is substituted into the target concentration decay model, and the predicted blood drug concentration on the predicted sequence slice is calculated by combining the half-life parameter. The theoretical decay time axis of individual drug concentration is then fitted to generate the model.

3. The intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, as described in claim 1, is characterized in that... The receiver acquires triaxial accelerometer signals and gyroscope angular velocity signals from the target wearable device, extracts frequency and time domain features, and outputs a micro-motion sequence feature stream, including: A continuous time window signal is obtained by performing sliding window segmentation and filtering noise reduction on the triaxial accelerometer signal and the gyroscope angular velocity signal; The frequency domain energy concentration is calculated by performing a fast Fourier transform on the continuous time window signal, and the time domain variance numerical features are extracted simultaneously. By splicing the frequency domain energy concentration and time domain variance numerical characteristics, a micro-motion sequence feature stream is output.

4. The intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, as described in claim 1, is characterized in that... The comparison with a pre-defined database of early warning signs of brain diseases and a micro-movement sequence feature stream generates an abnormal feature matrix and abnormal classification labels, including: The clustering distance between the feature flow of micro-movement sequences and the baseline classification plane in the pre-set brain disease precursor feature database is calculated using a sequence pattern recognition algorithm. When the shortest comparison distance is less than the preset anomaly detection threshold, the continuous time feature vectors are extracted and stacked to generate an anomaly feature matrix. The abnormal feature matrix is ​​labeled with abnormal classification labels that represent physiological feature patterns.

5. The intelligent follow-up and revisit reminder system for outpatients throughout their entire treatment cycle, as described in claim 4, is characterized in that... After labeling the abnormal feature matrix with abnormal classification tags representing physiological feature patterns, the process also includes: Principal component analysis is performed on the abnormal feature matrix to reduce its dimensionality, and principal component feature vectors with a cumulative contribution rate reaching a preset ratio are extracted. The principal component feature vectors and anomaly classification labels are input into a pre-defined graph neural network model to reconstruct the topological relationship, and the output is a graph structure anomaly feature matrix with spatial correlation weights. When generating micro-anomaly beacon data packets, the abnormal feature matrix is ​​replaced with a graph structure abnormal feature matrix before entering the encapsulation process.

6. The intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, as described in claim 1, is characterized in that... The encapsulation of the anomaly feature matrix, anomaly classification labels, and extracted anomaly occurrence timestamps generates a micro-anomaly beacon data packet, including: Read the system clock status parameters and convert them into a non-decreasing timestamp of the exception occurrence; The timestamps of anomalies, anomaly classification labels, and anomaly feature matrices are compiled into the core data payload. Add a data check bit and a communication identifier header to the outer end of the core data payload, and perform multi-layer structured encapsulation and compression to generate micro-anomaly beacon data packets.

7. The intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, as described in claim 1, is characterized in that... The process of parsing micro-anomaly beacon data packets to extract anomaly classification tags, matching them with a pre-defined subjective question and answer database, and assembling and issuing an immediate question and answer confirmation command includes: Disassemble the micro-anomaly beacon data packets to extract anomaly classification labels as method index pointers; Traverse the preset subjective question and answer questionnaire tree nodes, and retrieve multiple subjective feeling confirmation questions based on the method index pointer; Multiple subjective feeling confirmation questions are assembled and merged into an immediate question confirmation instruction with a high-priority immediate response identifier, and then executed and issued.

8. The intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, as described in claim 1, is characterized in that... The process of receiving subjective feedback confirmation signals and supplementary symptom description text in response to immediate rebuttal confirmation commands, fusing an abnormal feature matrix, and extracting the timestamp of the abnormal occurrence to establish the effective confirmation time point of the abnormal event includes: A hardware timer is activated to establish a countdown listening thread, which listens for subjective feedback confirmation signals and supplementary symptom description text. When capturing subjective feedback confirmation signals within a preset tolerance time window, the supplementary symptom description text and the abnormal feature matrix are fused and associated with a database table. The timestamp of the anomaly occurrence extracted from the micro-anomaly beacon data packet is established as the valid confirmation time point of the anomaly event.

9. The intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, as described in claim 1, is characterized in that... The process involves mapping the effective confirmation time of an abnormal event to the theoretical decay time axis of individual drug concentration to obtain the current predicted blood drug concentration value, calculating the time difference from the most recent target dosing time, and generating efficacy intervention early warning indicators when the time difference falls within a preset drug efficacy exhaustion period and the current predicted blood drug concentration value is lower than a preset lower limit blocking threshold. These indicators include: Align and map the effective confirmation time point of the abnormal event to the theoretical decay time axis of the individual drug concentration to extract the current predicted blood drug concentration value; Retrieve historical digital prescription information to obtain the most recent target medication time point for a single dose, and calculate the time difference between the effective confirmation time point of the abnormal event and the most recent target medication time point for a single dose. When the time difference falls into the preset period of drug efficacy depletion and the current predicted blood drug concentration is lower than the preset lower limit of the blocking edge threshold, an efficacy intervention early warning indicator is generated.

10. The intelligent follow-up and re-visit reminder system for outpatients throughout their entire treatment cycle, as described in claim 1, is characterized in that... The aforementioned response efficacy intervention early warning indicator injects a preset drug efficacy attenuation alarm code statement and supplementary symptom description text into the intercepted outpatient follow-up appointment data frame set, generating follow-up reminder information with priority identifiers, including: Intercept the set of outpatient follow-up appointment data frames output by the underlying follow-up cycle scheduling logic; The command bit of the response efficacy intervention warning indicator is flipped, the preset drug efficacy decay alarm code statement is merged with the supplementary symptom description text, and dynamically injected into the abnormal pop-up field area of ​​the outpatient follow-up appointment data frame set; Perform a package sealing operation on the data frame sequence to generate follow-up visit reminder information with priority indicators.

Images