Oral tumor patient pharmaceutical care management method and system based on multi-modal data

Abstract

The application discloses a multi-modal data-based oral tumor patient pharmaceutical care management method and system, belongs to the field of clinical pharmaceutical care and medical information processing, and aims at solving the problems of difficult unified correlation of multi-source data inside and outside the hospital, late discovery of drug risk and adjustment of drug administration. The application constructs a patient time sequence multi-modal data base, extracts features through double channels, and forms a drug sensitivity digital representation through cross-modal gating fusion. In combination with knowledge graph risk reasoning, follow-up semantic and physiological feature matching, and incremental Bayesian updating, the application generates a dose and drug administration interval recommendation under safety constraints, and forms a closed-loop care through verification, back transmission and monitoring cycle adjustment, so that the technical effect of continuously updating the drug state of the patient and timely outputting individualized care suggestions is realized.

Description

Technical Field

[0001] This invention relates to the field of clinical pharmaceutical monitoring and medical information processing, and in particular to a method and system for pharmaceutical monitoring and management of oral cancer patients based on multimodal data. Background Technology

[0002] Oral cancer patients typically involve the combined use of anti-tumor drugs, analgesics, anti-infective drugs, and supportive care drugs during hospitalization, discharge transition, and home rehabilitation. Current pharmaceutical monitoring mainly relies on electronic medical records, pathology reports, laboratory results, and medication records in the hospital information system for periodic assessments. Outside the hospital, it depends on regular follow-ups and proactive patient feedback. The continuous physiological signals collected by wearable devices have not yet formed a unified patient time-series data object with in-hospital diagnosis and treatment information and follow-up semantics.

[0003] Because structured fields, text, waveforms, and follow-up audio have different sampling frequencies and temporal semantics, existing systems struggle to establish stable associations within the same patient and monitoring period. Drug interaction and adverse reaction identification largely rely on fixed rules or pharmacist experience, making it difficult to adjust risk boundaries as liver and kidney function and physiological state change. Outpatient data is typically collected intermittently, and individualized pharmacokinetic parameters cannot be continuously updated using new observations. Consequently, dosage, dosing intervals, and monitoring frequencies cannot respond promptly to changes in patient status.

[0004] Therefore, there is a need for a pharmaceutical monitoring and management method and system for oral cancer patients based on multimodal data that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for pharmaceutical monitoring and management of oral cancer patients based on multimodal data. Addressing the problems in existing technologies such as difficulty in linking heterogeneous data from inside and outside the hospital, reliance on experience for medication risk identification, and the inability to continuously update individualized medication models, this invention proposes a method to construct a patient temporal multimodal data base using the initial diagnosis time as the origin. It generates a digital representation of medication sensitivity through temporal convolution and self-attention dual-channel extraction, followed by cross-modal gating fusion. This representation is used for knowledge graph graph attention risk inference, and the risk trigger threshold is adjusted based on liver and kidney function. Simultaneously, follow-up medical entities are matched with concurrent physiological statistical characteristics to form a state summary. The individualized medication model is sequentially updated using historical posterior distribution as the prior. Dosage and dosing interval recommendations are calculated under risk constraints. A closed-loop technical solution is formed through encrypted push notifications, identity verification, medication record feedback, and monitoring cycle adjustment. This invention achieves the technical effect of maintaining patient state correlation and dynamically outputting medication recommendations under safety constraints during continuous monitoring inside and outside the hospital.

[0006] This invention provides a method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data, comprising:

[0007] S1. Acquire in-hospital diagnosis and treatment data, medication data, continuous physiological signals and follow-up data of oral tumor patients, and perform time alignment and correlation of data based on timestamps and patient identifiers to generate a patient time-series multimodal data base;

[0008] S2. Input the patient’s temporal multimodal data into a dual-channel feature extraction network to extract the temporal variation features of continuous physiological signals and the long-range temporal correlation features of discrete medication events. Perform cross-modal gating fusion on the outputs of the two channels to generate a digital representation of medication sensitivity.

[0009] S3. Input the digital representation of drug sensitivity and the knowledge graph of drug interactions into the graph attention network, perform relational reasoning on the drug, disease and metabolic associations involved in the candidate drug combination, and correct the risk trigger threshold according to the patient's liver and kidney function indicators to generate the risk level label of the candidate drug combination.

[0010] S4. Perform text conversion and medical entity recognition on the speech in the follow-up data, standardize the identified symptom entities, and match the symptom entities with the statistical features of continuous physiological signals in the corresponding time period to generate a structured follow-up status summary vector.

[0011] S5. The follow-up status summary vector and the digital representation of medication sensitivity are used as new observation data to input the incremental Bayesian inference framework. The parameters of the individualized medication model are used as latent variables and the historical posterior distribution is used as the current prior distribution for sequential updates. Based on the risk level label, the parameter subset that meets the medication safety constraints is selected from the updated parameter posterior distribution. Based on the parameter subset, the conditional expectation of the dosage and dosing interval is calculated to generate the dosing recommendation result.

[0012] S6. Convert the medication recommendation results into pharmaceutical monitoring suggestions and send them to the patient's mobile terminal. Receive confirmation instructions associated with the patient's identity, transmit the confirmation information and current medication records back to the hospital information system to update the medication records, and adjust the monitoring cycle of the wearable device based on the transmitted results.

[0013] Optionally, S1 includes: acquiring electronic medical records, pathology reports and medication records from the hospital information system; acquiring photoplethysmography signals, triaxial acceleration signals and skin conductance response signals from the wearable device; and acquiring follow-up voice with acquisition time.

[0014] The time of the first pathological diagnosis recorded in the electronic medical record is determined as the patient's first diagnosis time, and the timestamps of each data are converted to a unified timeline with the patient's first diagnosis time as the origin.

[0015] Associate structured fields, text fragments, and waveform fragments within the same time window based on patient identifiers, and create indexes for the time dimension, data source dimension, and feature dimension respectively;

[0016] When an abnormal data acquisition is detected, a missing status marker is written at the corresponding index position without replacing it with data values ​​from other modalities, thus forming the patient's temporal multimodal data base.

[0017] Optionally, S2 includes: performing multi-scale causal convolution and dilated convolution on the continuous physiological signal using the temporal convolution sub-channel to obtain local fluctuation features and cross-time window change features;

[0018] The long-range temporal correlation features are obtained by calculating the attention weights between events using the self-attention sub-channel according to the drug identifier, dosing time, and dose code of the medication event;

[0019] The features of the two sub-channels are normalized and fully connected. Gating weights are generated based on the missing state labels and feature validity of each modality. The mapped features are then weighted, fused, and pooled using the gating weights to generate the digital representation of drug sensitivity.

[0020] When neither of the two sub-channels has a valid feature, stop generating a new drug sensitivity digital representation and write a data supplementation mark. If there is a drug sensitivity digital representation from the previous monitoring round, use it as the input for steps S3 and S5. If there is no drug sensitivity digital representation from the previous monitoring round, stop subsequent automatic recommendations and generate a pharmacist review alarm.

[0021] Optionally, S3 includes: establishing drug nodes, disease nodes, and metabolism-related nodes in the drug interaction knowledge graph, and configuring drug antagonistic relationships, drug synergistic relationships, drug-disease contraindication relationships, and drug-gene metabolism association relationships;

[0022] The node features involved in the candidate drug combination are fused with the drug sensitivity digital representation and then input into the graph attention network. The adjacent node information is aggregated according to the attention weight of each relation edge, and the drug combination conflict probability and adverse reaction type probability are output.

[0023] The risk trigger thresholds corresponding to each adverse reaction type are corrected based on the liver and kidney function index vector formed by estimated glomerular filtration rate, serum creatinine, alanine aminotransferase and total bilirubin. The probability is compared with the corrected risk trigger threshold to generate the risk level label.

[0024] Furthermore, the correction of the risk trigger threshold includes: calculating the distance of each indicator in the liver and kidney function indicator vector from the boundary of the corresponding reference interval according to the hospital test reference interval;

[0025] The threshold correction amount is determined based on the probability distribution of each adverse reaction type in the historical monitoring samples under the distance condition;

[0026] The change direction and magnitude of the risk trigger threshold are constrained by the threshold correction amount, and the indicators involved in the correction, the reference interval version, the probability distribution version of occurrence, and the reason for the correction are recorded to form a risk reasoning evidence record.

[0027] Optionally, S4 includes: performing word segmentation, part-of-speech tagging, dependency parsing, and relation tree pruning on the text obtained from the follow-up speech conversion;

[0028] The drug name, adverse reaction symptoms, symptom severity, and medication time are identified using a medical language model, and the adverse reaction symptoms are converted into standard medical term codes according to a medical terminology coding table.

[0029] A time-matching window is established with the time of symptom onset as the center. The length of the time-matching window is determined based on the elimination half-life of the corresponding drug and the sampling interval of the wearable device.

[0030] Extract statistical features of heart rate, blood oxygen saturation, and skin conductance within the time matching window, match the statistical features with the encoding vector of the medical entity according to the acquisition time and perform attention weighting to generate the follow-up status summary vector;

[0031] If no symptom entity is identified or no valid physiological signal fragment is found within the time matching window, a missing status marker is written to the corresponding component of the follow-up status summary vector. If both types of inputs are missing or cannot be matched, the follow-up status summary vector of this round is not used as quantitative observation data, and supplementary follow-up and physiological signal supplementation instructions are generated.

[0032] Optionally, S5 includes: setting the drug clearance rate, volume of distribution, and absorption rate constant as parameters of the individualized medication model; calculating the observation likelihood based on the newly added observation data; and combining the observation likelihood with the historical posterior distribution to obtain the parameter posterior distribution for the current round.

[0033] The single dose range and dosing interval range, therapeutic drug monitoring reference concentration range, and the risk level markers in the approved drug instructions are converted into the drug safety constraints.

[0034] From the posterior distribution of the parameters, a subset of parameters is selected that ensures the predicted peak blood concentration, the predicted trough blood concentration, and the cumulative exposure all meet the medication safety constraints.

[0035] The posterior mean and posterior quantile of the drug dosage and dosing interval under the parameter subset conditions are calculated respectively, and used as the drug dosing recommendation result and its confidence interval;

[0036] Furthermore, when the subset of parameters that satisfy the medication safety constraints is an empty set or the confidence interval of the posterior distribution of the parameters exceeds the credible range determined by the quantile boundaries of historical monitoring samples, the generation of new recommended values ​​for dosage and dosing interval is stopped.

[0037] If a current valid medication record exists, the system will revert to the dosing regimen corresponding to that record. If no current valid medication record exists, the case will be transferred to a pharmacist manual review status, and the existing prescription instructions confirmed by the physician will be retained. If no existing prescription instructions exist, the dosage and dosing interval values ​​will not be output.

[0038] Generate a pharmacist review alarm and determine the next monitoring cycle of the wearable device according to the state-cycle mapping table;

[0039] After receiving the pharmacist's review and confirmation and the new observation data, S5 is re-executed, and the cause of the abnormality, the rollback plan, the review result, and the re-inference result are written into the safety verification record.

[0040] Optionally, S6 includes: writing the risk level marker, drug name, single dose, daily dosing frequency, safety interval, and confidence interval into a formatted pharmaceutical monitoring recommendation, encrypting and signing the formatted pharmaceutical monitoring recommendation for integrity, and then sending it to the patient's mobile terminal;

[0041] The confirmation command is generated only when the biometric verification result of the patient's mobile terminal matches the patient's identifier;

[0042] After receiving the confirmation information and the current medication record, the hospital information system marks the previous medication record as expired and inserts a new medication record with the confirmation time and suggested version number;

[0043] A status-cycle mapping table is established based on historical monitoring samples. The status-cycle mapping table records the candidate monitoring cycles corresponding to each risk level mark, confirmation status, and data collection status. The candidate monitoring cycles are obtained from the status-cycle mapping table according to the current risk level mark, the returned confirmation status, and the data collection status. The candidate monitoring cycle with the smallest value is determined as the next monitoring cycle of the wearable device.

[0044] The status of sending, verification, updating, and periodic adjustment is written into the closed-loop pharmaceutical monitoring execution log.

[0045] On the other hand, the present invention also provides a pharmaceutical monitoring and management system for oral cancer patients based on multimodal data, comprising:

[0046] The data base construction module is used to execute step S1; the digital representation generation module is used to execute step S2; the medication risk reasoning module is used to execute step S3; the follow-up status parsing module is used to execute step S4; the medication model update module is used to execute step S5; and the monitoring closed-loop execution module is used to execute step S6.

[0047] The beneficial effects of this invention are:

[0048] 1. By using the time of the first diagnosis as the time origin and associating structured fields, text, and waveform fragments according to patient identification, in-hospital diagnosis and treatment data, continuous physiological signals, and follow-up data form a unified time-series multimodal data base, providing consistent data input for subsequent risk inference and model updates.

[0049] 2. By using dual-channel feature extraction, cross-modal gating fusion, and knowledge graph graph attention reasoning, and by adjusting the risk trigger threshold based on the patient's liver and kidney function, the risk of drug combination conflict and adverse reaction can be judged by combining the patient's status with the correlation between drugs.

[0050] 3. By using the follow-up status summary as a new observation to recursively update the individualized medication model, and calculating the recommended dosage and dosing interval on a subset of safety parameters, the confirmation and feedback results are used for record updates and monitoring cycle adjustments, so that new out-of-hospital statuses can enter the next round of pharmaceutical care. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of the pharmaceutical monitoring and management method for oral tumor patients based on multimodal data according to the present invention.

[0053] Figure 2 This is a flowchart of the incremental Bayesian inference and security constraint screening in the S5 method of the present invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0055] refer to Figures 1-2 A multimodal data-based approach to pharmaceutical monitoring and management of oral cancer patients includes:

[0056] S1. Acquire in-hospital diagnosis and treatment data, medication data, continuous physiological signals and follow-up data of oral tumor patients, and perform time alignment and correlation of data based on timestamps and patient identifiers to generate a patient time-series multimodal data base;

[0057] S2. Input the patient’s temporal multimodal data into a dual-channel feature extraction network to extract the temporal variation features of continuous physiological signals and the long-range temporal correlation features of discrete medication events. Perform cross-modal gating fusion on the outputs of the two channels to generate a digital representation of medication sensitivity.

[0058] S3. Input the digital representation of drug sensitivity and the knowledge graph of drug interactions into the graph attention network, perform relational reasoning on the drug, disease and metabolic associations involved in the candidate drug combination, and correct the risk trigger threshold according to the patient's liver and kidney function indicators to generate the risk level label of the candidate drug combination.

[0059] S4. Perform text conversion and medical entity recognition on the speech in the follow-up data, standardize the identified symptom entities, and match the symptom entities with the statistical features of continuous physiological signals in the corresponding time period to generate a structured follow-up status summary vector.

[0060] S5. The follow-up status summary vector and the digital representation of medication sensitivity are used as new observation data to input the incremental Bayesian inference framework. The parameters of the individualized medication model are used as latent variables and the historical posterior distribution is used as the current prior distribution for sequential updates. Based on the risk level label, the parameter subset that meets the medication safety constraints is selected from the updated parameter posterior distribution. Based on the parameter subset, the conditional expectation of the dosage and dosing interval is calculated to generate the dosing recommendation result.

[0061] S6. Convert the medication recommendation results into pharmaceutical monitoring suggestions and send them to the patient's mobile terminal. Receive confirmation instructions associated with the patient's identity, transmit the confirmation information and current medication records back to the hospital information system to update the medication records, and adjust the monitoring cycle of the wearable device based on the transmitted results.

[0062] In this specific embodiment, S1 includes:

[0063] In this specific embodiment, the data base construction module uses one round of patient pharmaceutical monitoring as the processing unit. It reads patient identifier, medical record document identifier, pathology report version, test item code, test value, test unit, drug generic name code, single dose, route of administration, administration time, and prescription status from the hospital information system, and reads follow-up task identifier, voice file identifier, and collection time from the mobile follow-up service. It reads photoplethysmography signal, triaxial acceleration signal, skin conductance response signal, and blood oxygen saturation field directly output by the device's blood oxygen measurement channel from the wearable device bound to the patient identifier. In this specific embodiment, the sampling frequency of the photoplethysmography signal is set to 64 Hz, the sampling frequency of the triaxial acceleration signal is set to 50 Hz, and the sampling frequency of the skin conductance response signal is set to 4 Hz. Blood oxygen saturation is expressed as a percentage and updated every 5 seconds. The blood oxygen saturation sampling record also includes the dual-wavelength measurement algorithm version, device calibration version, and blood oxygen quality status bit. Each sampling package also carries the device clock, receiving clock, device serial number, and quality status bit.

[0064] The data base construction module retrieves the earliest record from the electronic medical records that has a valid pathology report identifier and a confirmed diagnosis status, and determines the issuance time of that record as the patient's first diagnosis time; for the first... Each data object converts the original acquisition time into a unified time coordinate. , Indicates the first Each data object represents the number of minutes relative to the time of the first confirmed diagnosis, in minutes. This indicates the index of the data objects generated in the order they were received. This represents the Unix timestamp of the data object after clock correction, in seconds. This represents the Unix timestamp of the patient's first diagnosis, in seconds. The value is 60 and the unit is s / min. The time difference in seconds is divided by this scale to obtain a unified time coordinate in minutes. The deviation between the equipment clock and the hospital standard clock is given by the time synchronization record once a day. Sampling packets with a deviation of more than 5 seconds are first corrected according to the drift coefficient in the synchronization record. When there is no valid synchronization record, the received clock is retained and written with a clock uncertainty mark.

[0065] The data base construction module converts the timestamps of each data point to a unified timeline with the patient's first diagnosis time as the origin. Specifically, it calls the aforementioned unified time coordinate relationship to convert each object and writes the conversion results into a unified start time field. Clock calibration does not use a learning model. The drift coefficient is obtained by dividing the device time difference between the two most recent valid synchronization points by the hospital's standard time difference. It is updated every 24 hours and retains the synchronization point identifier, update time, and applicable interval. If the interval between two synchronization points is less than 1 hour or any synchronization point fails to verify, the coefficient is not updated and the previous valid calibration version is used.

[0066] The data infrastructure construction module uses 30 minutes as the basic association window for structured data, text fragments, and waveform summaries, and based on... Determine the window number. Indicates the first The index of the unified time window in which each data object resides is an integer. This indicates that the first time coordinate calculated according to the unified time coordinate formula in step S1 above... Each data object represents the number of minutes relative to the time of the first confirmed diagnosis, in minutes. This represents the floor function. The basic association window length is set to 30 min in this specific implementation. For data objects whose acquisition time is earlier than the first diagnosis time and originate from the same patient's verified in-hospital records, their negative values ​​are retained to form a unified time coordinate and written into the negative window index. The window index field of the monitoring window table uses a signed integer. The negative window is only used for baseline tracing and feature construction and does not trigger automatic drug recommendation before diagnosis. In this specific implementation, verified data within the earliest 180 days before the first diagnosis are retained. Objects outside this range only retain the original reference and out-of-range marker. Objects under the same patient identifier and the same window number are written into the same monitoring window record. Continuous waveforms that cross the window boundary are sliced ​​according to the sampling time. The text fragments corresponding to follow-up voice retain the original file identifier and start and end offsets so that the original data can be traced without being rewritten during slicing.

[0067] The patient time-series multimodal data substrate adopts a three-level carrier: patient table, monitoring window table, and modality fragment table. The patient table uses the irreversible hash value of the patient identifier as the primary key and records the time of first diagnosis and the current data version. The monitoring window table uses the patient hash value, window number, and monitoring round number as the composite key and records the window start and end times, data source bitmap, and integrity status. The modality fragment table records the fragment identifier, modality type, original object reference, unified start and end time, feature field address, quality status, missing status, and verification summary. The time dimension index is sorted according to the unified start time. The data source dimension index is bucketed according to the source code of the hospital information system, wearable devices, and follow-up services. The feature dimension index is organized according to the feature group code of test, medication events, photoplethysmography, blood oxygen saturation, acceleration, skin conductance response, and speech text.

[0068] For continuous physiological signals, the data basis construction module first eliminates saturated segments and transmission verification failure segments based on the quality status bit. It then applies a 0.5 Hz to 8 Hz bandpass filter to the photoplethysmography (PPG) signal, calculates the root mean square value of each axis within a 5-second window for the triaxial acceleration signal, and applies a 0.05 Hz low-pass filter to the skin conductance response signal, separating the slow-varying horizontal component and the fast-response component. For each 5-minute PPG time slice, it first searches for local maxima in the filtered waveform whose amplitudes are simultaneously greater than the neighboring 0.15-second sampling values ​​on both sides. For adjacent candidate points with an interval of less than 0.30 seconds, it retains only the candidate point with the higher peak amplitude. When the peak amplitudes are equal, it retains the earlier candidate point and then deletes the corresponding 5-minute PPG time slice. Candidate points whose synthesized root mean square acceleration exceeds the motion threshold listed in the equipment quality parameter table; the equipment quality parameter table uses equipment model, calibration version, and algorithm version as the lookup key, and the motion threshold is determined by the 95th percentile of the synthesized root mean square acceleration of the resting validation sample, reviewed and published by pharmacists and equipment engineers. If no valid version is found, the time slice is written to the pulsatility quality pending review status and the following features are not calculated; the times of the retained candidate points are recorded in ascending order as follows: and in accordance with Obtain the interval between adjacent pulses. Indicates the first A unified time coordinate for each effective pulsation peak, with the unit being seconds. Indicates the index of effective pulsation peaks. Indicates the first The first to the second The pulse interval between each effective beat peak is measured in seconds; only pulse intervals within the range of 0.30 s to 2.00 s are retained, and the effective beat ratio is the ratio of the retained intervals to the total number of candidate intervals. When the number of retained intervals is not less than 20 and the effective beat ratio is not less than 0.80, it is determined according to... and Generate the time slice features. This indicates heart rate, expressed in beats per minute. This indicates a second-to-minute conversion scale with a value of 60 and a unit of s / min. This represents the median of all effective pulse intervals. This represents the dimensionless coefficient of variation of the pulse interval. This indicates the standard deviation of the effective pulse interval calculated according to the definition of sample standard deviation. This represents the arithmetic mean of the effective pulse interval; when the quantity or proportion conditions are not met, both the heart rate and pulse interval coefficient of variation fields are written to the insufficient pulse missing status; the two values, effective pulse ratio, exercise threshold version, and pulse detection rule version are written to the photoplethysmography feature field address of the modal segment table, so that steps S2 and S4 can read the same version; for the blood oxygen saturation directly output by the device, only records with values ​​between 0% and 100%, valid blood oxygen quality status, and applicable calibration version are retained, and written to an independent blood oxygen saturation feature field according to a unified time coordinate, without inferring blood oxygen saturation from single-channel photoplethysmography waveform; the processed segment still retains the original sampling unit, and the waveform summary records the effective sample ratio, mean, standard deviation, minimum value, maximum value, and segment quality score. In this specific embodiment, segments with an effective sample ratio of less than 0.6 are not included in feature extraction, and only their index and missing status are retained. When the blood oxygen record does not meet the aforementioned value, quality, or calibration conditions, the blood oxygen missing status is written in the same field position;

[0069] When the total number of candidate intervals for a 5-minute photoplethysmography time slice is 0, the data basis construction module directly sets the effective pulsation ratio of the time slice to 0 and processes it as the aforementioned insufficient or missing pulsation state, thus not performing the proportional calculation with the number of candidate intervals as the denominator.

[0070] For structured fields within the hospital, the data base construction module uses the item codes and unit conversion rules in the hospital interface dictionary to merge items with the same name. Numerical verification items are only converted to the standard units specified in the interface dictionary when the units are convertible. Values ​​that cannot be converted are retained together with the original units and written into the unit anomaly flag. For electronic medical records, pathology reports, and follow-up texts, only character encoding unification, document paragraph positioning, and patient identification verification are performed. The original text is not replaced with automatically generated content. Duplicate objects are deduplicated according to patient identification, source object identification, collection time, and verification summary. Objects with different versions are saved side by side according to version number.

[0071] When the expected sampling package fails to arrive within three consecutive sampling periods, the effective sample ratio of waveform segments is less than 0.6, the hospital interface returns an empty field, or the follow-up voice verification summary is inconsistent, the data basis construction module writes a missing status marker at the corresponding time dimension, data source dimension, and feature dimension index position. The marker includes the missing modality code, the start and end time of the missing data, the detection rule code, the discovery time, and the supplementary sampling status. The mean, model output, or historical values ​​of other modalities are not used to fill this position, so that the subsequent gating network can distinguish between real observations and missing inputs.

[0072] After all objects in this round are associated, the data base construction module generates a data base version number. It writes the patient table primary key, monitoring window composite key, references to each modality segment, third-level index address, missing status marker, and object verification summary into the patient time-series multimodal data base and places this data base in the append-only monitoring data warehouse. Its data base version number, monitoring round number, and time window range are used as the dual-channel feature extraction input for step S2. At the same time, the follow-up voice segment reference and the corresponding unified time coordinate are provided to step S4.

[0073] In this specific embodiment, S2 includes:

[0074] The digital representation generation module reads the patient time-series multimodal data base written in step S1 and constructs continuous physiological signal tensors and discrete medication event sequences according to the monitoring window joint key. The continuous physiological signal tensor is divided into 5-minute time slices, with fields including heart rate, pulse interval variation coefficient, blood oxygen saturation, triaxial acceleration root mean square, skin conductance slow variation level, and skin conductance fast response count. Missing masks for each field are also added. The discrete medication event sequences are sorted according to the uniform dosing time. Each event records the drug code, route of administration code, dose normalized to the standard single dose in the instructions, time interval from the previous event, and prescription status.

[0075] The first continuous physiological signal The field in the th The observations of each time slice are generated by the digital representation generation module according to... Standardize, This represents the standardized dimensionless field value. This indicates the time slice index within the monitoring window. This indicates a continuous physiological field index. Indicates a validity mask. This indicates the dimensionless valid state value used in the mask. This indicates the dimensionless missing state value used in the mask, and in the normalized output context, it represents the dimensionless placeholder value used at the missing position. Represents the original field value while preserving the data collection unit for that field. Indicates the first training sample The average of each field, in the same unit as the original field. This indicates the corresponding standard deviation and has the same unit as the original field. The first in the feature standardization parameter table A continuous physiological field is used to prevent division by zero by strictly positive smoothing quantities of the same unit. This indicates taking the principal non-negative square root of the non-negative quantity within the radical sign; determining... First, compare the standard deviation of the field (one percent) with the minimum resolvable increment of the field. If the former is not less than the latter, take the former; otherwise, take the latter. This represents the smallest resolvable increment in the same unit, stored by field encoding and version in the device or interface dictionary, and must be strictly greater than zero; when hour, Pick ;when Time Read And calculate according to the previous formula, when Do not read missing data And directly order Meanwhile, the missing mask at that location is retained for subsequent network identification; the mean, standard deviation, minimum resolvable increment, and smoothing amount are stored in the feature standardization parameter table, retrieved by field encoding and model version, and updated when the training set is recalibrated. The parameter table rejects non-positive values. or ;

[0076] Multi-scale causal convolution and dilated convolution are performed on the continuous physiological signal by the temporal convolution sub-channel to obtain local fluctuation features and cross-time window change features. The temporal convolution sub-channel uses a four-layer causal convolutional network to process the normalized continuous physiological signal tensor. The convolution kernel widths are 3, 3, 5 and 5 respectively, and the dilation rates are 1, 2, 4 and 8 respectively. Each layer uses 64 output channels and performs layer normalization and gated linear units after convolution. Causal padding only references the current time slice and the time slices before it. The multi-scale output is concatenated in the time dimension and mapped to 128-dimensional local fluctuation features by one-dimensional convolution. The last layer covers the receptive field of no less than 32 time slices and outputs 128-dimensional cross-time window change features. The two features, along with the effective time slice ratio, are written into the continuous feature cache.

[0077] The self-attention subchannel concatenates drug encoding embedding, route of administration embedding, normalized dose, and time interval encoding into a 128-dimensional event vector, and extracts long-range temporal correlations using a six-layer, eight-head self-attention encoder; The first medication event on the first Attention weighting for each medication event according to calculate, Represents dimensionless attention weights and for index The sum is 1. This indicates a query for the event index. Indicates the index of the events being followed. This represents the index of the valid historical events traversed in the denominator. Indicates the first A dimensionless query vector for each event. Indicates the first A dimensionless key vector of events In the summation of denominators, the first... A dimensionless key vector of a valid historical event This indicates the transpose operation. The dimension of the positive integer vector representing a single attention head is 16 in this specific implementation. This indicates taking the principal non-negative square root of the non-negative quantity inside the radical sign. It represents the natural exponential function and both the input and output are dimensionless; events that are masked or whose time is later than the query event are not included in the denominator, and the encoder output is pooled by valid events to obtain 128-dimensional long-range temporal correlation features;

[0078] The dual-channel outputs are processed through layer normalization and fully connected mapping to obtain two 128-dimensional continuous feature vectors and medication event feature vectors, respectively. The digital representation generation module inputs the two vectors, the proportion of continuous effective time slices, the proportion of effective medication events, and two types of missing state bits into the gating layer; the gating weights are determined according to... generate, This represents a 128-dimensional dimensionless gated vector with values ​​ranging from 0 to 1. This represents a 128-dimensional dimensionless continuous eigenvector. This represents a 128-dimensional dimensionless feature vector of medication events. This represents a 4-dimensional dimensionless quality vector consisting of the proportion of consecutive effective time slices, the proportion of effective medication events, the missing state bits of consecutive channels, and the missing state bits of medication event channels, in sequence. This represents a dimensionless gated weight matrix with 128 rows and 260 columns, where the 260 columns correspond one-to-one with the concatenation order of the two 128-dimensional eigenvectors and the 4-dimensional mass vector. This represents the 128-dimensional dimensionless gated layer bias vector. and All vectors are determined from the historical monitoring training set and stored together in the model version file. The semicolon indicates vector concatenation. Represents an element-wise logical function;

[0079] Cross-modal fusion vectors according to calculate, This represents a 128-dimensional dimensionless cross-modal fusion vector. , and Using the complete definition from the aforementioned gating weight formula, let's represent the 128-dimensional dimensionless gating vector, the continuous feature vector, and the medication event feature vector, respectively. This represents element-wise multiplication. The dimensionless vector of all 1s with the same dimension as the gate vector is represented. The gate network is trained using historical monitoring samples. The training labels are formed by the adverse reaction status after pharmacist review and the subsequent treatment drug monitoring results. The loss function consists of classification cross-entropy, regression error of treatment drug monitoring value and gate sparsity regularization term. The training set and validation set are divided according to patient identifier to avoid leakage of the same patient across sets. The model version is jointly accepted by the macro-average comprehensive accuracy index and calibration error of the validation set.

[0080] The digital representation generation module uses effective time slices and effective events as pooling objects. It first performs attention pooling on the cross-modal fusion vector, then concatenates it with the patient's basic diagnosis and treatment vector and maps it to a 256-dimensional drug sensitivity digital representation through a two-layer fully connected network. The basic diagnosis and treatment vector only includes the currently effective age group, body surface area integral group, pathology type code, treatment stage code, and the missing status of liver and kidney function items in the data base. It does not use the risk results that have not yet been calculated in step S3 as input for this step. The representation record simultaneously saves the model version, data base version, effective modality bitmap, and generation time.

[0081] When there is no valid time slice in the continuous physiological channel but the medication event channel is valid, the gating layer sets the corresponding gating component of the continuous channel to 0 and generates the current representation only from the medication event features; when there is no valid event in the medication event channel but the continuous physiological channel is valid, the corresponding gating component is set to 1 and the current representation is generated only from the continuous features. The record generated by any single channel is written into the single modality state and the confidence quality field is downgraded by one level for steps S3 and S5 to identify its input quality.

[0082] When neither of the two sub-channels has a valid feature, the digital representation generation module does not perform fusion and pooling, does not generate a new drug sensitivity digital representation, and writes a data supplementation mark in the monitoring task table. If there is a valid drug sensitivity digital representation in the patient table from the previous monitoring round with the same model main version, its object reference is copied but the representation value is not copied. It is marked as a historical input and sent to steps S3 and S5. If the historical representation does not exist, the automatic new recommendation calculation from steps S3 to S5 in this round is stopped, a pharmacist review alarm containing the missing modality, missing time period and monitoring round number is generated, and the monitoring task orchestrator calls the unified blocking convergence interface of step S5 to form a structured drug administration recommendation result before continuing to execute step S6 safety closed loop.

[0083] The digital representation generation module ultimately writes the 256-dimensional drug sensitivity digital representation, quality level, effective modal bitmap, gating weight summary, model version, data base version, and historical usage status into the drug sensitivity representation library, and simultaneously registers the object identifier in the current round of monitoring task record; step S3 performs drug combination relationship inference based on the object identifier, and step S5 reads the same object identifier and the follow-up status summary from step S4 to form the new observation data.

[0084] In this specific embodiment, S3 includes:

[0085] The medication risk reasoning module reads the medication sensitivity digital representation from step S2 and the current patient's candidate drug combinations. The representation is obtained by performing multi-scale causal convolution and dilated convolution on continuous physiological signals through the temporal convolution subchannel of step S2 to obtain local fluctuation features and cross-time window change features, and is fused with the long-range temporal correlation features of the self-attention subchannel. The candidate combinations consist of drugs in the current effective prescription and comparison schemes pre-configured by the hospital pharmacist. The drug interaction knowledge graph is mainly composed of drug nodes, disease nodes, and metabolism-related nodes. Drug nodes record generic name codes, routes of administration, and targets. Disease nodes record standard diagnostic codes and contraindication levels. Metabolic-related nodes record metabolic enzyme or transporter codes and functional states, and configure drug antagonistic relationships, drug synergistic relationships, drug-disease contraindication relationships, and drug-gene metabolism association relationships.

[0086] The knowledge graph construction service reads node and relationship records from the hospital's drug knowledge base, the hospital-confirmed structured instruction manual database, and the pharmacist review rule base. Node keys are composed of node type and standard code, and relationship keys are composed of start key, end key, relationship type, and source version. The construction service matches node pairs according to relationship type and filters duplicate relationships according to evidence order, retaining records with the highest order and most recent effective time among the established order of evidence levels one, two, and three. When the start or end point is invalid, the candidate qualification for that edge in this round is deleted. If the candidate subgraph is empty or the drug or disease cannot be mapped to the standard code, the graph reasoning for that combination is stopped, and the candidate combination is removed. The identifier, reason for knowledge deficiency, code to be reviewed, and map version are written into the queue to be reviewed, and a pharmacist review risk status with a priority higher than the ordinary risk level is formed. Combinations in this status are prohibited from entering the automatic new recommendation in step S5. Step S6 selects a conservative monitoring period according to the unconfirmed review status. After the pharmacist completes the code or map relationship review and publishes the new version, the construction service re-executes this step with the original candidate combination identifier. The blocking status is only lifted when the new version can form a non-empty candidate sub-map and all codes pass the verification. The reviewer identifier, lifting time, and new map version are appended to the original knowledge deficiency record, thereby generating a candidate drug combination sub-map that can be traced by version.

[0087] The medication risk reasoning module combines the type embedding, standard encoding embedding, relation statistics, and medication sensitivity digital representation from step S2 of each node into a 128-dimensional node input feature after fully connected mapping, and employs a two-layer, eight-head relation-aware graph attention network; for nodes... Pointing to node And the relation type is The attention weights of the edges are determined according to... Normalization Representing relation type Next node Aggregation Node The dimensionless weight of information Indicates the index of the central node. Indicates the index of adjacent nodes. Indicates a relation type index. This represents the dimensionless correlation value calculated from the features of the central node, the features of neighboring nodes, and the relationship embedding. This represents the index of adjacent nodes in normalized summation. Indicates the relationship With nodes The set of valid adjacent nodes for connection. Following the complete definition of the natural exponential function in step S2, both its input and output are dimensionless;

[0088] The graph attention network aggregates the information of adjacent nodes under each relation and then performs relation-gated fusion. The graph representation of all drug nodes in the candidate combination is input into two output heads. The first output head gives the probability of drug combination conflict, and the second output head gives the probability of occurrence of each type in the preset adverse reaction type set. The model is trained using pharmacist-confirmed interaction cases and non-conflict control cases. Conflict labels are from closed-loop pharmaceutical monitoring execution records, and adverse reaction labels are jointly generated by structured follow-up status and pharmacist review records. The validation set is isolated according to patient identification, and the output probability is calibrated by temperature scaling and the calibrated version is saved.

[0089] The medication risk reasoning module reads the most recently valid estimated glomerular filtration rate, serum creatinine, alanine aminotransferase, and total bilirubin from the patient's time-series multimodal data base, and reads the lower limit, upper limit, and interval versions for the same test method, gender group, and age group from the hospital's laboratory reference interval table; for the first... Each liver and kidney function indicator, according to Calculate the boundary deviation distance. This represents the non-negative dimensionless deviation distance. An index representing four liver and kidney function indicators. This indicates the patient's current test value. Indicates the lower limit of the reference under the same unit. Indicates the upper limit of the reference range for the same unit. Indicates the relationship with the first A scale quantity whose test value is of the same unit and strictly greater than zero. This indicates that the maximum value among the candidate values ​​of the same dimension within the parentheses is taken. A dimensionless constant whose deviation is zero when located within the reference interval; scale quantities are defined according to... Sure, This represents the interquartile range of historical normal samples under the same test method, gender grouping, and age grouping, and is in the same unit as the test value. The risk parameter library is arranged according to the first... The code and test method version for each liver and kidney function test item are saved; a strictly positive smoothed quantity of the same unit is used only for risk scale calculation; in the second formula... Using the complete definition of the aforementioned maximum value function; determine First, read the minimum resolvable increment from the interface dictionary and one-hundredth of the historical non-zero interquartile range. If the minimum resolvable increment is not less than one-hundredth, take the minimum resolvable increment; otherwise, take one-hundredth. When the interquartile range is zero, use... And write the scale rollback flag, the risk parameter library will not publish non-positive values. ;

[0090] Historical monitoring samples are binned according to adverse reaction type and deviation distance to establish an occurrence probability distribution table. The table fields include adverse reaction type, indicator index, deviation distance interval, sample number, occurrence probability, confidence lower bound, source time range, and distribution version. In this specific implementation, each bin contains at least 100 reviewed monitoring rounds. When the sample is insufficient, it is merged with adjacent bins. An execution record can be represented as nephrotoxicity type, estimated glomerular filtration rate deviation distance between 1 and 2, occurrence probability of 0.18, and distribution version of 2026Q2. The table is reconstructed by newly added reviewed samples each quarter and stored in the risk parameter database.

[0091] For risk item types The revised risk trigger threshold is based on Sure, This represents the risk item type index, whose value set is encoded by a drug combination conflict-specific type. Together with the preset adverse reaction type code, it forms a whole. This represents the traversal index for estimating four indicators: glomerular filtration rate, serum creatinine, alanine aminotransferase, and total bilirubin. Using the non-negative dimensionless deviation distance from the aforementioned boundary deviation distance formula, the summation range is the range of the four indicators that have valid values ​​or have generated conservative correction amounts for missing states. This represents the corrected dimensionless risk trigger threshold, with a value ranging from 0 to 1. This represents the dimensionless baseline threshold when liver and kidney indicators are within the reference range. This represents the non-negative dimensionless index weight obtained by fitting historical samples. This represents the dimensionless threshold correction amount, obtained from the probability distribution table by deviation distance and truncated to between 0 and 1. This indicates the dimensionless lower bound of the threshold defined by the hospital risk rules. The hospital risk rule table only allows the publication of records that satisfy the condition that zero is not greater than the lower bound of the threshold, the lower bound of the threshold is not greater than the basic threshold, and the basic threshold is not greater than 1. This indicates that the result will be truncated between the given lower and upper bounds; the dedicated records for drug combination conflicts and adverse reaction records use the same fields, correction rules, and missing rollback rules, and are uniquely queried by type code, indicator index, route of administration, and rule version. No match found. Automatic recommendation of the combination is blocked during recording; the correction amount does not decrease when the deviation or historical probability increases; when the weighted correction exceeds the difference between the basic threshold and the lower threshold, the lower threshold is taken by the truncation function. Therefore, the risk threshold only changes in the direction of decreasing value and is not lower than the lower threshold.

[0092] When a test value is missing or the unit cannot be converted, the medication risk reasoning module does not infer the value based on other indicators. Instead, it sets the corresponding threshold correction to the conservative quantile value of the adverse reaction type in the missing state, as determined by historical samples, and writes the reason for the missing value in the evidence record. When the query key of the probability distribution table is not matched, the module uses 90% of the correction amount of all valid samples of the same adverse reaction type. If there are still no valid samples, it stops outputting an automatic low-risk conclusion for the type and marks it as the pharmacist review level.

[0093] The attention layer is first arranged according to Calculate the correlation of relationships. , and Using the complete definition from the aforementioned relational attention weight formula, let's represent the center node index, adjacent node index, and relation type index, respectively. This indicates that the aforementioned node pair is in relation type The dimensionless correlation value under the following conditions Represents a node 128-dimensional dimensionless input features Represents a node 128-dimensional dimensionless input features Representing relation type The corresponding 128-row, 128-column dimensionless node mapping matrix, Multiplying the input feature by any node outputs a 128-dimensional dimensionless vector. Representing relation type The 32-dimensional dimensionless embedding vector, the total dimension of the three vectors concatenated within the square brackets is 288. This represents a dimensionless attention parameter vector, which, like the concatenated vector, has 288 dimensions. The semicolon within the square brackets indicates the transpose operation, and it signifies vector concatenation. This represents an activation function with a slope of 0.2 on the negative half-axis. , and All were trained using verified graph relation samples and stored with the graph model version;

[0094] Central node updates according to implement, Represents a node The aggregated 128-dimensional dimensionless graph features This represents the exponential linear activation function, in the summation... Indicates a valid relation type index, in the summation. This represents the node index within the aforementioned set of adjacent nodes. The attention weight, the set of adjacent nodes, the mapping matrix, and the node features follow the previous definitions. The empty set only retains the node's own mapping features and writes them into the isolated node state. The aggregated drug node features are pooled according to the candidate combination identifier and then input into the conflict output header and the adverse reaction output header.

[0095] The medication risk reasoning module compares the probability of drug combination conflicts and the probability of each adverse reaction type with corresponding correction thresholds. When any probability reaches or exceeds the threshold, the corresponding risk item is marked as triggered. The risk level is determined sequentially based on the highest preset severity among the triggers, the proportion of probability exceeding the threshold, and the existence of contraindications. Valid contraindications are directly marked as the highest risk. The risk level uses four ordered enumerations: low risk, medium risk, high risk, and highest risk, with ranking numbers of 0, 1, 2, and 3 respectively. Candidate combinations are arranged in ascending order of ranking number. Combinations with knowledge gaps, contraindications, or those under pharmacist review are not included in the automatic selection sequence. When multiple automatically selectable combinations have the same risk level, the one with the lowest conflict probability is prioritized. If both the risk level and conflict probability are the same, the stable order is determined by ascending candidate combination identifiers. The risk grading rule table is reviewed and published by pharmacists and is categorized by rule version, risk item type, preset severity code, and probability exceeding the threshold proportion. The segment composition follows a unique rule; the preset severity codes are 0, 1, 2, and 3, corresponding to low, medium, high, and highest severity, respectively. The probability exceeding the threshold ratio is calculated by subtracting the correction threshold from the risk probability and then dividing by the maximum value between the correction threshold and 0.01. When the threshold is not reached, the ratio is recorded as 0; when the threshold is reached and the ratio is less than 0.25, the segment code is 1; when the ratio is not less than 0.25 and less than 1, it is 2; and when the ratio is not less than 1, it is 3. The level ordinal number of each trigger item is the maximum value between the preset severity code and the ratio segment code. When there is no trigger item, the combined level ordinal number is fixed at 0; when there is a valid taboo relationship, it is fixed at 3. When there are multiple trigger items, the one with the largest level ordinal number is selected. When the largest ones are tied, the evidence recording order is determined only by the ascending order of the risk item standard code without changing the level. When the severity code, threshold, or rule item is missing, exceeds the limit, or the version is invalid, the automatic risk level is not output and the case is transferred to the pharmacist for review. The risk reasoning evidence record simultaneously saves the rule version, the hit item, the intermediate codes of each item, and the final level ordinal number.

[0096] The medication risk reasoning module writes a candidate combination identifier, a set of drug codes within the combination, a risk level marker, a conflict probability, an adverse reaction type probability, thresholds before and after correction, a graph version, a graph model version, and a characterization object identifier for each candidate drug combination. When any probability field fails the 0-1 range validation, the automatic classification of the combination is stopped and the status is written to the pharmacist review status. For combinations that are not blocked, the first item is selected as the selected combination for this round according to the stable order of the aforementioned risk level, conflict probability, and candidate combination identifier, and the status is written to the selected combination. The adjustment priority table configured by the pharmacist determines the target adjustment drug code according to the candidate combination identifier, risk type, and drug code. If no match is found, the blocking is automatically recommended and transferred to the pharmacist review. The module writes the selected combination identifier, the set of drug codes within the combination, the target adjustment drug code, the liver and kidney function indicators involved in the correction, the test reference interval version, the probability distribution version, the threshold change direction, the change magnitude, and the correction reason into the risk reasoning evidence record. The selected combination risk level marker and evidence record are provided to step S5 to construct medication safety constraints. The risk level marker is also provided to step S6 to generate formatted pharmaceutical monitoring recommendations and determine the monitoring cycle.

[0097] In this specific embodiment, S4 includes:

[0098] The follow-up status analysis module reads the follow-up speech segment references and unified time coordinates saved in step S1, performs endpoint detection, spectral subtraction and noise reduction, and speaker channel verification on the 16kHz mono audio, and then converts it into text with character-level start and end times by a Chinese speech recognition model enhanced with medical vocabulary. The speech recognition model is trained with patient-authorized and de-identified in-hospital follow-up speech. The vocabulary includes generic drug names, adverse reaction terms, and expressions related to oral tumor treatment. The decoding results simultaneously save the text, character-level time, and recognition confidence. When the average character confidence is below 0.75, it is marked as text to be reviewed and not included in quantitative observation.

[0099] The follow-up status parsing module performs word segmentation, part-of-speech tagging, and dependency parsing on valid text. It prunes the dependency tree by retaining drug names, symptom words, and time words, deleting modified branches that do not connect to any retained nodes. The pruned word sequence input is a medical entity recognition model based on a Chinese Transformer encoder and a conditional random field output layer. The output is the entity of drug name, adverse reaction symptoms, symptom severity, and medication time, along with the corresponding confidence level. The training labels are independently labeled by two pharmacists and verified against divergent samples. The model version is released after achieving an entity-level comprehensive accuracy of 0.90 on an independent patient validation set.

[0100] The identified drug names are first normalized to hospital drug codes using a generic drug name dictionary and an alias table. Adverse reaction symptoms are normalized to standard medical terminology codes using a medical terminology coding table. Severity is mapped to ordered values ​​from 0 to 3 according to four levels in the terminology table: none, mild, moderate, and severe. Medication time entities are converted to a unified time coordinate. The parsing module establishes a symptom occurrence time state for each symptom entity. Only when the symptom is in the same positive dependency subtree, the entity confidence is not lower than 0.80, and there are no two parsing results that differ by more than 30, is the symptom corrected. Only time entities with a time value of min are written to the valid symptom occurrence time with their unified time coordinates. When no time entity that meets the conditions is identified, the time expression cannot be converted, or multiple time entities conflict as mentioned above, the symptom occurrence time status is fixed as "time to be supplemented," and is not replaced by the voice collection time, medication time, or historical symptom time. No time matching window is established, no quantitative physiological summary is generated, and the symptom record is not allowed to be used as a quantitative observation in step S5. At the same time, a supplementary time inquiry and pharmacist review task containing the original sentence, symptom code, candidate time expression, conflict reason, and follow-up task identifier are generated. Only after the supplementary answer is verified by the patient and confirmed by the pharmacist is a new follow-up round generated and this step is re-executed. The original time to be supplemented record is only appended with the review result and is not overwritten. The parsing module matches the standard terminology code according to the source of the original word and synonym. If no match is found within the version validity period, the original text is retained and written to the terminology pending review status. The medical terminology code table records the original word, standard terminology code, synonym source, validity period, and version number and is reviewed and updated by the pharmacist quarterly. It is not automatically replaced by the closest terminology.

[0101] The follow-up status analysis module performs deterministic symptom-drug association for each symptom entity: First, it selects standardized drug entities within the same sentence that form a positive dependency path with the symptom through the relational words "taken," "used," "infused," "discontinued," or "increased dosage," and are not modified by negative words. If multiple drugs meet the conditions in the same sentence, an association record and matching window are created for each drug separately, without merging their evidence. If no dependency path is formed, the drug with the most recent administration time within 24 hours before the symptom onset is selected from the patient's current effective medication record. If the administration times are the same, the first item is selected in ascending order of hospital drug codes, and the fallback type is written into the association evidence. Drug entities are missing, drug codes are pending review, and drugs with the most recent administration time within 24 hours before the symptom onset are also considered. If there are no effective dosing events within h or if there is a conflict between positive and negative relationships, a quantitative physiological summary is not generated. Instead, the association is written to an unresolved state and a pharmacist review task containing the original sentence, candidate drugs, dependency paths, and monitoring round numbers is generated. For drugs with confirmed associations but missing half-lives, non-positive or incomparable units in the package insert's structured library, the conservative matching half-window parameter table reviewed by the pharmacy committee is queried first according to the route of administration and dosage form. If a match is found, the half-window length within the range of 30 min to 720 min in the table is directly used and the parameter table version is recorded. If no match is found, the quantitative summary is also stopped and the process proceeds to pharmacist review.

[0102] For the corresponding drug At the time of symptom onset, the follow-up status analysis module determines the time-matching half-window based on the drug's elimination half-life and the wearable device sampling interval, according to... calculate, This indicates the length of the matched half-window used before and after the onset of symptoms, in minutes. Represents a standardized drug coding index. This indicates the current effective sampling interval of the wearable device, in minutes. This represents the floor function. This indicates that the dimensionless half-life ratio is set to 0.25 in this specific embodiment. This indicates the drugs in the structured library of the instruction manual. The elimination half-life is converted to min; the calculation results are cut off between 30 min and 720 min to avoid too few samples or crossing irrelevant monitoring phases.

[0103] The follow-up status analysis module reads the effective photoplethysmography signal and skin conductance response signal from the matching window, and combines the triaxial acceleration to synthesize the root mean square to exclude periods of intense exercise; extracts the mean heart rate, standard deviation of heart rate, coefficient of variation of pulse interval, mean blood oxygen saturation, proportion of patients with blood oxygen saturation below 95% of the patient's resting baseline, median slow change level of skin conductance, and number of rapid responses from the effective samples. Each statistic retains its original unit or is converted to a dimensionless quantity by the ratio of the baseline of the same unit. The resting baseline is determined by the median of the samples in the same time period that are symptom-free and have effective quality status in the most recent seven days.

[0104] The medical entity encoding vector consists of standard symptom encoding embedding, severity encoding embedding, drug encoding embedding, and relative medication time encoding. Physiological statistical features are first standardized field-by-field according to the standardization parameter table in step S2, then mapped to a vector of the same dimension. Missing physiological components are standardized by writing the value 0 and concatenating it with the corresponding binary missing mask. Then, a fixed linear mapping published by the scoring parameter table is used to form a 64-dimensional dimensionless physiological vector. The 64-dimensional dimensionless entity vector of each symptom entity is denoted as . , will the The 64-dimensional dimensionless physiological vector of each physiological time slice is denoted as . Correlation values ​​are based on calculate, Indicates the first The symptom entity and the first Dimensionless correlation values ​​of individual physiological time slices Represents the symptom entity index. Indicates the physiological time slice index. This represents the transpose of a vector. This represents a dimensionless bilinear rating matrix with 64 rows and 64 columns. This indicates a dimensionless scoring bias. This represents the non-negative dimensionless time penalty coefficient. Indicates the first Subtract the first physiological time slice center time from the first The time difference obtained from the occurrence of each effective symptom, in minutes. This indicates taking the absolute value. The aforementioned drugs The corresponding strictly positive matching half-window length is in min, therefore the time penalty term is dimensionless; , and The training process involves matching symptom-physiological time slices confirmed by pharmacists. Relevant time slices are marked as 1, and irrelevant time slices within the same window are marked as 0. The objective function is the sum of binary cross-entropy and matrix L2 regularization terms, and the training and validation sets are divided according to patient identifiers. The scoring parameter table uses symptom standard codes, drug codes, feature standardized versions, and scoring model versions as query keys, recording the matrix, bias, time penalty coefficient, training sample time range, effective time, expiration time, and pharmacist review identifier. Quantitative summarization stops and pharmacist review begins when the query fails, the dimension is mismatched, the validation summary fails, or the version is invalid. When the scoring parameters are valid and the set of valid physiological time slices is not empty, the first... The symptom entity and the first The matching weight of each physiological time slice is based on calculate, This indicates that the dimensionless attention weights sum to 1 within the corresponding matching window. This represents the physiological time slice index in normalized summation. Indicates the first Each entity corresponds to a set of valid physiological time slices within the matching window. Following the complete definition of the natural exponential function in step S2, both its input and output are dimensionless;

[0105] No. Physiological summaries corresponding to each symptom entity are categorized according to calculate, Indicates the relationship with the first Physiological statistical feature vectors of each symptom entity time-matched, maintaining the dimensionless nature of each standardized component. Indicates the first The standardized statistical feature vector of each physiological time slice, the weight, entity index, time slice index and effective time slice set follow the definition of the aforementioned matching weight formula; if the effective time slice set is empty, the formula is not calculated and the branch with missing physiological components is entered.

[0106] The follow-up status parsing module concatenates the physiological summary with entity encoding vectors, speech recognition confidence, entity recognition confidence, and data quality bits, and generates a 128-dimensional follow-up status summary vector through a two-layer fully connected network. This network is trained using historical follow-up rounds and symptom statuses confirmed by pharmacists within the following 24 hours. The loss function consists of symptom status classification cross-entropy and physiological feature reconstruction error. The output records store the symptom standard code, occurrence time, associated drug code, matching window, proportion of effective physiological segments, model version, and glossary version.

[0107] When no symptom entity is identified, the follow-up status parsing module writes a missing status marker into the entity status component of the summary vector; when there is no valid physiological signal fragment in the matching window, a missing status marker is written into the physiological status component and the identified entity code is retained. When both types of input are missing, the average word confidence is less than 0.75 and manual review cannot be completed, or the set of valid physiological time slices is empty and there is no valid entity, the summary vector of this round is not used as the quantitative observation in step S5, and a supplementary follow-up and physiological signal supplementary sampling instruction containing follow-up task identifier, missing type, suggested supplementary sampling time window and device identifier is generated.

[0108] The follow-up status parsing module writes the valid 128-dimensional follow-up status summary vector, missing status of each component, entity recognition results, matching weight summary, data base version and model version into the follow-up status database, and forms a joint key with the patient identifier, monitoring round number and follow-up task identifier; this joint key and summary vector object identifier are sent to step S5, and together with the drug sensitivity digital representation generated in step S2, they form new observation data in the same monitoring round.

[0109] In this specific embodiment, S5 includes:

[0110] The medication model update module reads the follow-up status summary vector from step S4, the medication sensitivity digital representation from step S2, and the selected combination risk level label from step S3. After verifying the consistency of these three elements—patient identifier, monitoring round number, and data base version—it constructs a new observation record. The medication sensitivity digital representation in step S2 is generated by performing multi-scale causal convolution and dilated convolution on continuous physiological signals using a temporal convolution sub-channel to obtain local fluctuation features and cross-time window change features. This is then fused with the long-range temporal correlation features of the medication event. Record fields include the selected combination identifier, the set of drug codes within the combination, the target adjustment drug code, and the given... The steps include: route of administration, characterization vector reference, follow-up summary vector reference, risk level, most recent dosing event sequence, therapeutic drug monitoring concentration and sampling time, body surface area, weight, liver and kidney function indicators, and the quality status of each input. This step only establishes parameter objects and candidate dosing regimens for the target adjustment drug. Non-target drugs within the combination use the current effective dosing records and their dosages and intervals must not be changed. Drug interaction limits are used as combination-level safety constraints in the screening. If a quantitative follow-up summary is missing, updates will only continue if the drug sensitivity numerical characterization is effective and there is a therapeutic drug monitoring concentration for the target adjustment drug; otherwise, the abnormal branch of this step will be entered.

[0111] The personalized medicine model employs a one-compartment absorption and elimination structure branched by route of administration and mode of administration. For the target modifier in the selected combination, parameter objects are established separately for drug code, route of administration, and mode of administration. Mode of administration codes are retrieved from confirmed dosing events or physician prescriptions; for intravenous routes, only three published codes are allowed: intravenous bolus, intermittent intravenous infusion, and continuous intravenous infusion. For oral administration and other routes of administration with identifiable absorption processes, the personalized medicine model parameters are set to... , This represents the patient parameter vector representing the current target drug for adjustment. This indicates the drug clearance rate, expressed in L / h. This indicates the apparent volume of distribution of the drug, expressed in liters (L). This represents the drug absorption rate constant, expressed in units of 1 / h; The overall representation is a vector of patient parameters with mixed dimensions, consisting of clearance rate, apparent volume of distribution, and absorption rate constant. The overall unit is not applicable because the units of its three components are L / h, L, and 1 / h, respectively, and cannot be assigned a single physical unit; for intravenous administration, the parameter vector is set to , This represents a vector of patient parameters for the intravenous route. The overall representation is a vector of patient parameters consisting of clearance rate and apparent volume of distribution. The overall unit is not applicable because the units of its two components are L / h and L, respectively, and cannot be assigned a single physical unit. The absorption rate constant field in the intravenous parameter object is written to an inapplicable state and is not included in the prior fitting, posterior inference, particle perturbation, or confidence range determination. The model applicability table reviewed and issued by the pharmacist uses the drug code, route of administration, administration method, and model version as a unique key and records the allowed state and required input fields. If the allowed state is not hit, the administration method cannot be identified, or the required input field is invalid, automatic new recommendations are blocked and the model enters the pharmacist review process. The population mean and covariance of the parameter prior are determined by the same drug code, the same treatment stage, and the same administration method. For historical monitoring samples with the same route of administration and method of administration, when patients already have a historical posterior distribution with the same drug code, route of administration, and method of administration, the model does not return to the overall prior, but directly reads the posterior parameters and their versions from the previous effective monitoring round; the model applicability table fixes the automatic adjustment status of continuous intravenous infusion as manual review only; the confirmed rate and duration of this method are only used for replaying past actual events, current exposure assessment, and pharmacist monitoring, and do not enter candidate protocol generation, dose-interval inference, or daily frequency derivation. When this method is hit, it directly enters the unified blocking convergence rule and retains the physician's prescription or switches to manual review only;

[0112] The medication model update module performs numerical integration on the absorption chamber and central chamber drug quantities based on the dosing event sequence. The starting state of the actual event integration is determined according to a unique priority order: first, the previous effective pharmacokinetic state object whose patient identifier, target drug code, route of administration, method of administration, and model version are consistent and whose state time is no later than the earliest dosing event in this round is read. Only when the dosing records from that state object to the earliest event in this round are continuous and complete and the verification summary is consistent, the absorption chamber and central chamber drug quantities are restored for oral and other absorption routes, while only the central chamber drug quantity is restored for intravenous routes. When a state object meets the conditions, it is reconstructed from the most recent discontinuation time confirmed by the pharmacist. Only if the dosing-free period between that discontinuation time and the next dosing is at least 5 times the conservative terminal half-life published in the package insert's structured library, will the absorption chamber and central chamber drug amounts for the applicable route be initialized to 0 at that discontinuation time. If neither condition is met, the dosing record after discontinuation is incomplete, or the state version is invalid, integration is stopped, automatic new recommendations are blocked, and the system enters pharmacist review; the initial state is not replaced with 0. For oral administration and other routes of administration involving absorption, the continuous state relationship is written as... , and , This indicates the amount of drug in the absorption chamber, in mg. This represents the time elapsed since the initial state of integration, expressed in hours (h). This indicates the dosage in the central compartment, in mg. This represents the predicted blood drug concentration, in mg / L; the bioavailability correction factor for the route of administration is denoted as... , Indicates drug Route of administration and administration method The dimensionless bioavailability coefficient with a value strictly greater than 0 and not greater than 1. Indicates the target drug coding index. Indicates the route of administration coding index, This indicates the administration route code index; this coefficient is read from the bioavailability parameter table reviewed and issued by the pharmacist. The parameter table uses a unique key composed of drug code, route of administration, administration method, approved package insert version, and parameter version. The numerical value is derived from the absolute bioavailability stated in the approved package insert or the fitted results of drug monitoring of the same drug and route of administration group, reviewed by the pharmacist, and records the sample range, effective time, expiration time, and review identifier; each time an oral or other absorption route of administration is administered... Add to the absorption chamber, This indicates the dose of the confirmed dosing event, and the unit is mg; therefore, the amount of drug added to the absorption chamber is in mg. If the parameter table is not hit, the coefficient is out of bounds, the version is invalid, or the review identifier is invalid, 1 or other default values ​​will not be used; instead, the simulation will stop, automatic new recommendations will be blocked, and the process will proceed to pharmacist review. For intravenous bolus injections, Write the inapplicable status and add the entire dose to the central chamber at the confirmed injection time; for intermittent or continuous intravenous infusions, do not read the bioavailability coefficient, but read the confirmed total infusion volume, infusion rate, start time, and end time, verify that the infusion duration is strictly greater than zero and that the relative error between the total infusion volume and the infusion rate multiplied by the duration does not exceed the tolerance of the published interface. In each integration step, only the dose obtained by multiplying the overlap time between the step and the infusion interval by the infusion rate is added to the central chamber, and the dose in the central chamber is updated according to the clearing item; if any infusion field is missing, verification fails, or the model applicability table does not allow it, stop the simulation, block automatic new recommendations, and write the reason for pharmacist review. The candidate regimen generation and subsequent safety recalculation both use the same dosing method branch and the same input record version; the absorption chamber status is not established or read for each intravenous branch. In this specific embodiment, an integration step size of no more than 5 minutes is used and does not cross the start or end boundary of drug administration or infusion. The predicted concentration is output at each time of therapeutic drug monitoring sampling. At the same time, the source of the integration start state, the state object version or drug discontinuation record identifier, the bioavailability parameter version and the range of drug administration events are written into the simulation summary.

[0113] For the One round of monitoring, below This refers generally to the aforementioned parameter vector corresponding to the current route of administration; the posterior parameter is based on... Sequential update, Indicates including the 1st to the 1st The posterior density of the current parameters after each round of new observations. Indicates the first to the last New observation sets added in the round This indicates that, given the parameter vector, the first... The likelihood of the new observations in the round, Indicates the first New observation records added in the round This indicates the prior density prior to this update. and The Lebesgue reference measure is defined as a product of the parameter space of the current route of administration. The reference measure for absorption routes such as oral administration is written as... Since the unit is L² / h², the unit for both parameter densities is h² / L². The reference measure for the venous route is written as... And the unit is L² / h, so the unit of both parameter densities is h / L²; This represents the dimensionless likelihood weight obtained by multiplying the effective observations in this step after normalization by variance or calibration temperature. It is not used as the probability density relative to the original observed physical quantities. This represents a mixed dimensional parameter vector selected according to the current route of administration. The overall unit is not applicable, and each component uses the aforementioned L / h, L, or 1 / h units. This indicates a newly added set of observation records with mixed dimensions; the overall unit is not applicable. Indicates the first For newly added observation records with mixed dimensions, the overall unit is not applicable. The therapeutic drug monitoring concentration, sampling time, body surface area, weight, and liver and kidney function indicators for both types of data retain their respective source field units. The category, code, reference, and quality status fields do not apply physical units. If a previous effective patient with the same drug code, route of administration, method of administration, and version compatibility exists in the posterior, that posterior is used; otherwise, the approved overall prior with the same drug code, treatment stage, route of administration, and method of administration is used. This represents the monitoring round index, which increments by confirmation time. The proportional symbol indicates that after the update, it still needs to be weighted and normalized by the likelihood of all parameter samples. Particle initialization uses a unique existence branch: if there is a previous valid patient posterior compatible with the aforementioned version, 2000 particles are extracted from that posterior and its normalized weights are inherited; if not, 2000 particles are extracted from the aforementioned reviewed overall prior and each particle is assigned the same normalized initial weight. Any monitoring round and review retry selects the prior according to this existence condition, without presuming the prior source based on the round sequence number. Only when the prior object itself is missing necessary parameters or the version is invalid will it be transferred to the pharmacist for review. New observation records save the prior source type, source object identifier, overall prior or patient posterior version, and particle initialization branch code.

[0114] The observational likelihood is obtained by multiplying the therapeutic drug monitoring concentration error term, the follow-up status classification term, and the drug sensitivity characterization consistency term. Each term is first converted to a dimensionless conditional probability of 0 to 1 according to its respective training set variance or calibration temperature. For each parameter particle, the drug model update module reads the parameter subset and the sequence of most recent dosing events according to the route of administration. For oral and other absorption routes, clearance rate, volume of distribution, and absorption rate constant are used, while for intravenous routes, only clearance rate and volume of distribution are used. Then, the concentration, peak, trough, and cumulative exposure at the sampling time are simulated. The therapeutic drug monitoring concentration error term is obtained by normalizing the Gaussian density of the difference between the simulated and measured concentrations under the error variance of the training set of the same drug detection method. The follow-up status classification term uses a multi-class logical logic trained with pharmacist-verified samples. The logical model uses three logarithmic parameters of the particles, simulated peak, trough, cumulative exposure, time since most recent administration, 128-dimensional follow-up status summary, and corresponding missing mask as input for the absorption pathway. For the intravenous pathway, a dedicated model version is used, with only two logarithmic parameters (clearance rate and volume of distribution) and other identical features input. After temperature calibration and normalization, the output probability of the actual follow-up category is taken. The consistency term of drug sensitivity characterization uses a binary classification model jointly calibrated by therapeutic drug monitoring results and pharmacist-confirmed adverse reactions. Three or two logarithmic parameters, simulated exposure, 256-dimensional drug sensitivity digital characterization, target adjusted drug code, and quality status bits are input according to the same pathway rules. The output probability of consistency with the current observation is obtained through a logical function. The feature order of the two conditional models... The coefficients, calibration temperature, training set time range, and version are stored in the observation model parameter table. The drug code, route of administration, and model version are adjusted according to the target for retrieval. If no match is found, the data is transferred to a pharmacist for review. When a certain type of observation is missing, it is excluded from the likelihood product in this round using a missing mask, and the reason for exclusion is recorded. If there are no valid items in the therapeutic drug monitoring concentration, follow-up summary, and digital representation, the posterior is not updated. In this specific implementation, a sequential Monte Carlo algorithm with 2000 particles is used for updating. The initial particles are drawn from the posterior of the previous valid patient or the reviewed overall prior according to the previously recorded prior existence branch. The weights are updated according to the observation likelihood corresponding to each particle. When the effective sample size is less than half the number of particles, system resampling is performed, and only the parameters in the current route of administration parameter subset are resampled. Log-normal perturbation is applied to positive parameters. Particles, weights, routes of administration, discontinuation parameter status, observation versions, and random seeds are written into the posterior object of the current parameter. The peak, trough, and cumulative exposure of the actual event likelihood are calculated only within a single evaluation window of the actual event with a version: when there is an effective therapeutic drug monitoring sampling point, the most recent dosing interval containing the last effective sampling point of this round and with a complete start and end dosing event is selected; when there is no sampling point, the last complete dosing interval that has ended by the data cutoff time of this round is selected; when there are multiple dosing events in the same interval, all of them are used as integral inputs but only one evaluation window is formed. When the evaluation window spans monitoring rounds, the starting event of the previous round is used only when the event chain is continuous, the data base version is traceable, and the verification summary is consistent.If the window has not yet ended, the start and end events are missing, or the cross-cycle event chain is discontinuous, the likelihood terms corresponding to the peak, trough, and cumulative exposure will not be calculated and the case will be transferred to the pharmacist for review. The steady-state evaluation window of the candidate scheme shall not be used to replace the evaluation window of the actual event. The newly added observation record shall save the evaluation window code, start and end event identifier, start and end time, selected branch, cross-cycle status, and the version of the observation model parameter table.

[0115] The medication model update module reads the single dose lower limit, single dose upper limit, minimum dose step, dosage form splitting rules, allowed dosing interval set, and applicable route of administration from the approved drug instructions structured library; it reads the peak concentration reference upper limit, trough concentration reference lower and upper limits, and cumulative exposure reference range from the therapeutic drug monitoring rule library; and it reads the dose upper limit multiplier, peak concentration upper limit multiplier, and minimum interval increment from the risk constraint parameter table. The query key for the risk constraint parameter table consists of the selected combination identifier, target adjustment drug code, route of administration, risk level, adverse reaction type, and parameter table version. Fields also include three correction parameters, effective time, expiration time, and pharmacist review identifier. The instructions structured library is searched by drug code, route of administration, and approved version; the monitoring rule library is searched by drug code, indication group, and detection method version. The dose upper limit multiplier and peak concentration upper limit multiplier are both limited to between one-half and one, multiplied by the single dose upper limit in the instructions and the peak concentration reference upper limit in the monitoring rules, respectively, and truncated towards a more conservative published dose step or detection accuracy. The candidate dose set is... generate, Represents a finite set of candidate doses. This indicates the positive minimum dose step size for the approved release, expressed in mg. This indicates the lower limit of a single dose via the same route, expressed in mg. This indicates the upper limit of a single dose after risk factor correction and truncation towards a conservative dose step, and the unit is mg. Indicates an integer dose range index. This represents the floor function. This represents a floor function; it only retains dose ranges that satisfy the dosage form divisibility rule and are neither lower than the lower limit nor higher than the upper correction limit. When the upper and lower endpoints are not divisible by the step size, they contract inwards respectively; when the set is empty, new recommendations are blocked; candidate dosing regimens are... The Cartesian combination of the allowed dosing interval set and applicable routes and methods of administration is used; intermittent intravenous infusion candidate protocols are read from the pharmacist-approved infusion parameter table, including the allowed duration set, allowed infusion rate set, infusion rate step size, minimum and maximum infusion pump rates, total volume consistency interface tolerance, and parameter table version; for each candidate single dose, the allowed duration set is traversed in ascending order of duration, the candidate dose is divided by the candidate duration to obtain the original infusion rate, and then mapped to the nearest rate level in the allowed infusion rate set, with the rate level having the smallest rate value if the distance is the same; only the infusion rate is retained. The dose-rate-duration triplets are selected based on the following conditions: the rate is within the infusion pump's capacity boundary, the relative error between the candidate dose and the mapped infusion rate multiplied by the candidate duration does not exceed the interface tolerance, and the instruction manual's path constraints are valid. If multiple qualified triplets exist for the same dose, all of them will proceed to the subsequent safety simulation. If there is no qualified triplet, the rate level is not matched, or the parameter table version is invalid, automatic new recommendations will be blocked. The safety verification record saves the source version, traversal order, mapping rules, total error, and rollback status of the candidate rate and duration. Continuous intravenous infusion does not generate automatic candidate schemes. The minimum interval increment is a non-negative integer. After sorting the dosing intervals allowed by the instructions in ascending order, the interval is shifted from the minimum allowable interval to a longer interval by the corresponding number of increments. If the interval exceeds the set, the longest interval in the set is used. When the same combination triggers multiple risk items, the dose and peak concentration are the minimum values ​​among the corrected results of each risk item, and the minimum interval is the maximum value among the corrected results of each risk item, so that the upper limit of dose and peak concentration do not increase and the minimum dosing interval does not shorten when the risk level increases. No new recommendations are generated and the process is transferred to pharmacist manual review when any query key is not hit, the corrected parameter is out of bounds, or the version is invalid. When the selected combination is low risk and no risk item reaches the trigger threshold, the adverse reaction type query key is fixed to the BASELINE code approved and released by the pharmacist, and the baseline constraints are read according to the other query keys. Automatic new recommendations are prohibited when the BASELINE record is not hit or the version is invalid. The safety verification record saves the query key, parameter table version, hit status, and rollback reason. The complete candidate regimen for intermittent intravenous infusion is composed of a combination tuple based on single dose, dosing interval, infusion rate, and infusion duration. Each combination tuple is assigned a stable regimen identifier and serves as the minimum regimen object for subsequent safety sets, combination weights, discrete recalculation, and result selection.

[0116] For each weighted parameter particle and candidate dosing regimen, the drug model update module performs a deterministic periodic steady-state simulation using the aforementioned pharmacokinetic state relationships consistent with the route and manner of administration of that regimen; the candidate regimen simulation does not inherit the residual state of the patient's actual event integrals, and the absorption routes such as oral administration are set at the beginning of cycle 0. and The intravenous route uses the aforementioned zero value for the central compartment dosage as initialization and sets... Write an inapplicable state. and Represent the initial drug volume in the absorption chamber and the central chamber of the candidate regimen, respectively, both in mg; oral, intravenous bolus, and intermittent intravenous infusions were administered repeatedly at the candidate dosing intervals, with the length of one evaluation cycle being the candidate dosing interval; the first... The particle and the first The first scheme is in the The peak and trough values ​​within each evaluation period are respectively denoted as and Both are in mg / L. Indicates the particle index of the parameter. This represents an index of candidate dosing regimens. This indicates an evaluation period index that increments from 1; starting from the 2nd period, it follows... Calculate the steady-state change. It represents a non-negative dimensionless steady-state change. This indicates that the maximum value of the candidate values ​​with the same dimension within the parentheses is selected. This indicates taking the absolute value. This indicates that the therapeutic drug monitoring rule base publishes a strict concentration resolution lower limit greater than zero, with units of mg / L, based on drug code, detection method version, and steady-state evaluation window code; when two consecutive evaluation periods... All are no greater than the dimensionless steady-state tolerance published by the rule base. When the second satisfying cycle ends, the steady-state arrival time is determined. This represents the steady-state tolerance, which is strictly greater than 0 and less than 1; the maximum number of simulations is 100 evaluation periods, and the total time domain from the 0th period to the steady state must not exceed 30 days. Any particle-scheme pair that fails to converge even after reaching any upper limit... or If a query fails to find a match or exceeds the limit, stop all automatic recommendations for the current round, write down the reason for the lack of steady-state convergence or invalid steady-state parameters, and proceed to pharmacist review. Once steady-state is reached, use the immediately following complete evaluation cycle as the safety evaluation window. The evaluation window length for oral, intravenous bolus, and intermittent intravenous infusion is the dosing interval of the regimen. The peak value is the maximum predicted concentration within the window, and the trough value is the predicted concentration of the integration step before the next dosing. Cumulative exposure is calculated according to... calculate, Indicates the first The particle and the first The cumulative exposure of each scheme within the steady-state safety evaluation window, expressed in mg·h / L. Indicates the first The safety evaluation window length for each protocol is given in hours (h), and in non-continuous infusion cases, it is equal to the candidate dosing interval. This indicates that the particle-scheme pair is within the safety evaluation window at time... The predicted concentration is given in mg / L. Indicates the range from 0 to The concentration-time numerical integration is calculated using the trapezoidal method within an integral grid of no more than 5 minutes. The peak upper limit, trough range, and cumulative exposure range in the therapeutic drug monitoring rule base must be recorded with completely consistent evaluation window codes, administration methods, units, and integration calibers. For non-continuous infusions, interval steady-state window codes are used. Automatic recommendation is stopped and pharmacist review is initiated if any code, unit, or integration caliber is inconsistent, or if the integration data within the evaluation window is incomplete. The safety recalculation after discrete mapping uses the same initial state, convergence criterion, evaluation window, and integration caliber. The set of safe protocols comprises all protocols that simultaneously meet the following conditions: single dose within the approved range, dosing interval within the permissible set, peak concentration not exceeding the risk-corrected upper limit of peak concentration, trough concentration within the reference range, and cumulative exposure within the reference range. and with express The number of schemes in the list; and All are positive integer indices. For a finite set, The parameter vectors are non-negative integers; a subset of parameters satisfying the drug safety constraints is retained, consisting of particle parameter vectors with at least one safe scheme, and the constraint code that is first violated for each deleted particle-scheme pair is recorded.

[0117] for Strictly greater than zero Each particle, first according to To each Assign weights to dimensionless conditional schemes, and then according to... Determine the joint weights of the particle and the scheme; Indicates the first Within the first particle The dimensionless conditional weights of each security scheme and for All The sum is 1. Indicates the 1st in the current posterior. The non-negative dimensionless weights of each particle This represents the non-negative dimensionless joint weight of the corresponding particle-scheme pair; the medication model update module follows... and Calculate the recommended value. This indicates the post-conditional single-dose dose, in mg. Indicates the first The single-dose dosage for each safety regimen is in mg. This indicates the conditional post-test dosing interval, in hours. Indicates the first The dosing interval for each safety regimen, in hours; both denominators are only for particles with at least one safety regimen and their... Summation is performed, and if the sum of all joint weights is zero, the subset of safety parameters is treated as empty.

[0118] The confidence intervals for the dosage and dosing interval are taken from the 5% and 95% quantiles of the safety subset weighted empirical distribution, respectively, and the recommended dose is mapped to the aforementioned candidate dose set. The nearest value of the conditional posterior mean in the medium distance is selected; if the distances are the same, the smallest value among two candidate doses is chosen. The recommended interval is mapped to the nearest value of the conditional posterior mean in the set of allowed dosing intervals; if the distances are the same, the largest value among two candidate intervals is chosen. The medication model update module re-executes the concentration simulation on the discretely mapped scheme using a model branch consistent with the target drug's route and method of administration. Dosing recommendations are only generated if the peak, trough, and cumulative exposure all still meet the safety constraints. If not all are met, the abnormal reason is uniquely encoded as a failure of discrete mapping safety recalculation, and the failure constraint code, discrete scheme, and simulation summary are recorded. No new dose or interval is generated. The safety interval is defined as the minimum allowed dosing interval after the risk constraint parameter table is corrected, in hours. It represents the safety boundary, not the actual recommended dosing interval, and the actual recommended dosing interval must not be shorter than the safety interval. The daily dosing frequency is based on... derived, This indicates the average number of doses administered per day, expressed in doses / day. This indicates the number of hours per day, expressed in hours (h / d). Represents the recommended dosing interval obtained through discrete mapping and recalculation, with units in hours; when When the value is not an integer, it is recommended to display the field using per... The interval is expressed as once per hour, and the numerical field stores the average daily frequency to two decimal places without rounding to an integer. The module records the recommended dosing interval, safety interval, daily dosing frequency, frequency expression, conditional posterior mean, quantile interval, constraint version, and simulation summary in the results. For intermittent intravenous infusion, after the dose and interval are discretized, only the joint tuples with the same single dose and recommended dose, the same dosing interval and the same discrete safety recalculation are retained. The particle-scheme joint weight of each retained tuple in all safety particles is summarized, and the tuple with the highest summarized weight value is selected as the unique execution scheme. When the summarized weights are the same, the tuple with the shortest infusion duration, the smallest infusion rate value, and the first stable scheme identifier in ascending character order is selected. If there are no retained tuples that have passed the recalculation, the abnormal reason is set to the intermittent infusion joint tuple unavailable and automatic new recommendations are blocked. The infusion rate and infusion duration of the selected tuple are used as the unique execution values ​​of the dosing recommendation result, and no fields are supplemented from other tuples.

[0119] The monitoring task orchestrator uniformly inputs the blocking codes generated in any of steps S2, S3, and S5 into the convergence rules of this segment. When any blocking code is hit, the subsequent automatic calculations are no longer executed. All new recommended dose, infusion rate, infusion duration, recommended interval, and daily frequency fields are set to empty. At the same time, the blocking code, the obtained risk level or risk status, the source step, and the object version are retained. Then, according to the current effective medication record, the existing prescription confirmed by the physician, and the fixed priority of neither of the two, a rollback plan is generated, the physician prescription is retained, or only the manual review status is checked. Step S6 is always called to complete the alarm, periodic query, and closed-loop record keeping. The data blocking in step S2 only stops the automatic new recommendation calculation in steps S3 to S5, and does not stop the safety closed loop in step S6. When the safety parameter subset is empty, the posterior interval of any parameter in the current administration route parameter subset exceeds the boundary of the 1% to 99% quantile of the same drug, same route, and same administration method monitoring sample, the discrete mapping safety recalculation fails, or the administration method and model applicability are checked. When the test is blocked, the medication model update module stops generating new dosages and dosing intervals. For oral and other absorption routes, clearance rate, volume of distribution, and absorption rate constant are verified; for intravenous routes, only clearance rate and volume of distribution are verified, without reading or verifying the absorption rate constant. The module determines the rollback status according to a unique priority order: if a current valid medication record exists, the recommended status is set to rollback to the existing protocol, referencing only the valid record without generating replacement values; otherwise, if a physician-confirmed existing prescription instruction exists, the recommended status is set to retain the physician's prescription and referencing it; if neither exists, the recommended status is set to manual review only, with the dosage, recommended interval, and safety interval fields all left empty and the reason for stopping recorded. Any of the above anomalies is written with the corresponding anomaly reason, failure constraint code, and simulation summary, generating a pharmacist review alarm. Simultaneously, step S6 is called to determine the next monitoring cycle based on this mutually exclusive recommended status and the current data collection status. The four statuses—valid new recommendation, rollback to the existing protocol, retaining the physician's prescription, and manual review only—are mutually exclusive; only the valid new recommendation status allows step S6 to generate a medication change confirmation instruction.

[0120] An abnormal branch generates a pharmacist review alarm and calls step S6 to use the status-cycle mapping table to determine the next monitoring cycle. The safety verification record writes the abnormal cause, trusted range version, violated constraints, current valid medication record identifier, and rollback plan. All parameters resulting from the abnormal execution are marked as unsubmitted intermediate objects and cannot be used as priors in subsequent review attempts. Upon receiving pharmacist review confirmation and new valid observation objects, a new review execution attempt number is generated at the confirmation time, while maintaining the original Bayesian update round index. The prior objects recorded at the start of the original abnormal attempt are reread. If the original prior was the posterior of the previous valid patient, that object is still read; if the original prior was the reviewed overall prior, particles are initialized according to the same version of the overall prior. The monitoring task orchestrator... The original round observations and the supplementary observations are merged according to the observation object identifier and the observation version. Duplicate objects are rejected idempotently. The merged results are then sorted stably by observation time and object identifier and used as the complete round of observations to recalculate all valid likelihood terms. Only after the a priori update, safety screening, discrete recalculation and successful generation of dosing recommendation results are completed are new valid posterior objects submitted and the Bayesian update round index incremented. The safety verification record saves the prior object identifier before execution, the complete consumed observation object identifier and version, the supplementary observation object identifier, the likelihood term list, the deduplication results, the intermediate posterior unsubmitted status and the final submission status. Then, the verification results, the re-inferred posterior object identifier and the new recommendation status are appended to the same safety verification record without overwriting the original abnormal record.

[0121] The medication model update module writes the recommended status, selected combination identifier, drug code set within the combination, target adjustment drug code, route of administration, administration method, single dose, infusion rate, infusion duration, recommended dosing interval, safety interval, daily dosing frequency, frequency expression, risk level, dose confidence interval, interval confidence interval, parameter posterior version, model suitability table version, instruction manual version, therapeutic drug monitoring rule version, risk constraint parameter table version, and safety verification record identifier into the dosing recommendation results; for non-infusion methods, the infusion rate and infusion duration fields are written as inapplicable status, and abnormal branches are simultaneously written with the abnormal reason, failure constraint code, and simulation summary; the daily dosing frequency and frequency expression are derived only from the recommended dosing interval that has been recalculated in this step according to the aforementioned unique rule, and externally filled values ​​without a source are not accepted; valid new recommendations are based on the selected group. The combination identifier and target adjustment drug code are recorded item by item, and non-target drugs within the combination continue to reference the current effective medication record; this result is only used as the structured input for generating formatted pharmaceutical care recommendations in step S6, and is sent to the closed-loop execution module along with the risk reasoning evidence record and the current effective medication record identifier; for continuous intravenous infusion, the dosing recommendation result is only allowed to be written to the existing physician prescription or only manually reviewed status, and no new single dose, infusion rate, infusion duration, recommended dosing interval or daily dosing frequency is generated; when referencing an existing physician prescription, the prescription object identifier is saved separately and the existing value is not marked as a new recommendation of the model; the intermittent intravenous infusion result simultaneously saves the stable protocol identifier of the uniquely selected combination tuple, single dose, recommended interval, infusion rate, infusion duration, summary combination weight, parallel decision path, infusion parameter table version and discrete safety recalculation summary.

[0122] In this specific embodiment, S6 includes:

[0123] The closed-loop monitoring module reads the dosing recommendation results from step S5 and the risk level markers from step S3, and generates formatted pharmaceutical monitoring recommendations according to the recommendation pattern file. The recommendation object fields include patient identifier hash value, monitoring round number, recommendation version number, recommendation status, selected combination identifier, drug code and name, route of administration, administration method, single dose, daily dosing frequency, recommended dosing interval, safety interval, dose and interval confidence interval, intermittent intravenous infusion stable protocol identifier, infusion rate, infusion duration, infusion parameter table version, risk level, generation time, expiration date, and pharmacist review status. The safety interval is directly read from the dosing recommendation results in step S5, determined by the risk constraint parameter table, in hours. The recommended dosing interval represents the proposed interval and must not be shorter than the safety interval. Valid new recommendations... The recommended status generates a medication change confirmation entry. For rollback of existing protocols and retention of physician prescriptions, only informed consent and review entry points are generated. For manual review only, no medication change confirmation entry point is generated. When there are no new recommended values ​​in abnormal branches, null values ​​are written to the dosage, recommended interval, and safety interval fields, along with the rollback protocol or manual review status; null values ​​are not converted to zero. For non-infusion methods, the stable protocol identifier, infusion rate, infusion duration, and infusion parameter table version fields for intermittent intravenous infusion are written with an "inapplicable" status. When the administration method is intermittent intravenous infusion and the recommended status is a valid new recommendation, the aforementioned four fields are read item by item from the uniquely selected combination tuple and its infusion parameter table version in step S5, and none of them can be null. When the administration method is continuous intravenous infusion, the retention of physician prescriptions or manual review only status from step S5 is used, and no new recommended infusion parameters are generated by the model.

[0124] The monitoring closed-loop execution module maintains the device public key and session key version for each patient's mobile terminal. It first encrypts the suggested object using a symmetric encryption algorithm approved by the hospital, and then uses the hospital server's private key to perform an integrity signature on the ciphertext digest, suggested version number, and generation time. The message record includes the patient hash value, terminal identifier, ciphertext, signature, key version, serial number, and replay protection random number. The mobile terminal verifies the signature and serial number, decrypts and displays the result. If the verification fails or the serial number is duplicated, the module refuses to generate a confirmation entry and sends the reason for the failure back to the closed-loop execution record.

[0125] When a patient initiates confirmation on their mobile terminal, the terminal calls the local biometric verification interface. The verification template is stored only in the terminal's secure area. The closed-loop execution module only receives the verification success status, terminal signature, patient hash value, suggested version number, confirmation time, and a one-time random number. A medication change confirmation instruction is generated only when the recommendation status is a valid new recommendation, the patient hash value matches the suggested object, the terminal signature is valid, the suggested version is still valid, and the one-time random number has not been used. If any of these conditions are not met, the confirmation status is set to rejected, and the medication record in the hospital information system is not updated. When the sending service does not receive a terminal response within 30 minutes after the suggestion expires or after successful sending, a timeout event is generated. The confirmation status is atomically updated from sent to unconfirmed using the suggested version number as the concurrency control key, keeping the current valid medication record unchanged. The status-cycle mapping table is queried immediately with the unconfirmed status, and a cycle adjustment instruction is generated. Late confirmations arriving after the expiration of the validity period are rejected. Only newly generated suggested versions can re-enter the confirmation process. The timeout event, the number of resends, and the final confirmation status are appended to the closed-loop execution record.

[0126] Confirmation instructions and current medication records are transmitted back to the hospital information system via the hospital interface gateway. Interface transactions use an idempotent key composed of the patient hash value, recommendation version number, recommendation status, and confirmation time, and first execute mutual exclusion gating based on the recommendation status. Only when the recommendation status is a valid new recommendation, and the dosage, recommended dosing interval, and safety interval are all non-empty, the recommended dosing interval is not shorter than the safety interval, and the stable protocol identifier, infusion rate, infusion duration, and infusion parameter table version are all non-empty when the dosing method is intermittent intravenous infusion, are all valid. Within the same database transaction, the previously valid medication record is marked as expired and the expiration time is written. Subsequently, a new drug code, route of administration, dosing method, single dose, dosing interval, intermittent intravenous infusion stable protocol identifier, infusion rate, infusion duration, infusion parameter table version, confirmation time, recommendation version number, source monitoring round, and treatment are inserted. For new medication records in the associated state, the four intermittent intravenous infusion fields (excluding infusion methods) are written to the "inapplicable" state. After successful transaction commit, a new record identifier is returned and associated with the aforementioned field values ​​using this identifier. When the recommended state is to roll back an existing plan or retain a physician's prescription, only confirmation and review records are added, and the original valid medication records are maintained. When the recommended state is to only perform manual review or when any value is empty, medication change transactions are rejected. If a transaction fails, all writes of this transaction are rolled back, and the original valid records are retained. The result codes of each branch and subsequent recovery conditions are written to the closed-loop execution record. When the recommended state is to only perform manual review, the closed-loop execution module adds pharmacist review queue records and safety alarm records without creating hospital medication change transactions. For the three states of rolling back an existing plan, retaining a physician's prescription, and only performing manual review, the non-change reason, referenced object, or empty value state, as well as the record appending result, are all written to the closed-loop execution record.

[0127] After step S4 generates the follow-up status summary, the data acquisition status aggregator reads the monitoring window integrity, modality-by-modality missing status, and re-collection status from step S1, the effective modality bitmap, quality level, single modality status, and historical status from step S2, and the symptom entity component and physiological component missing status from step S4, using the monitoring round number and data base version. It then generates a unique data acquisition status from high to low according to a fixed priority order: acquisition blocked, acquisition pending, acquisition downgraded, and acquisition complete. Acquisition is blocked when there is no currently valid representation and no historical representation to use in step S2, or when the integrity check of a key object fails. Otherwise, acquisition is blocked if three consecutive sampling cycles have not been reached, the re-collection status is pending or failed, step S2 uses historical representation, or any part of the follow-up status summary is missing. When a quantity is missing, it is considered as pending collection; otherwise, collection is downgraded when any modality is missing, the effective sample ratio is less than 0.6, or the single modality representation or quality level is downgraded; collection is considered complete only when all expected modalities are valid, step S2 uses the current round of dual-channel representation, and step S4 has both types of components complete and no pending supplementary collection; when multiple conditions are met in the same round, only the highest priority state is retained; the aggregator completes state freezing before outputting in step S5, and writes the state code, formation time, the field on which it is based and its value, the data base version, the representation object version, and the follow-up summary version into the current round of monitoring task record; after supplementary collection is completed, the original state must not be overwritten, but a new monitoring round is generated and re-aggregated, and step S6 only queries the state-cycle mapping table with the unique state and version of the current round;

[0128] The status-cycle mapping table is constructed from historical closed-loop pharmaceutical monitoring samples. Each record uses a unique query key composed of a status dimension type and a status code. The status dimension types are limited to three categories: risk level, confirmed status, and data collection status. Fields include status dimension type, status code, candidate monitoring cycle, cycle unit, sample size, anomaly detection rate, version, and effective time. The construction service statistically analyzes the time distribution of risk escalation before the next monitoring session according to the three status dimensions. The shortest sampling cycle that can cover 90% of risk escalation events is written into the corresponding single-dimensional record and reviewed and published by the pharmacist. In this specific implementation, a risk dimension record can be represented as a status dimension type of risk level, a status code of high risk, and a candidate monitoring cycle of 30. The minimum number of samples (420) and version 2026Q2 are used. Confirmation status and data collection status are recorded separately according to their respective codes. Before release, the monotonic non-incremental constraint of the candidate period is checked item by item according to the ordered status: the corresponding period shall not increase when the risk level rises from low risk to the highest risk; the corresponding period shall not increase when the confirmation status changes from confirmed to unconfirmed, rejected, or timed out; the corresponding period shall not increase when the data collection status deteriorates from complete collection to collection downgrade, collection pending, and collection blocked. If any record violates the constraint, the entire version shall not be released and shall be rolled back to the previous valid mapping table version. If there is no previous valid version, a 15-minute safe default period shall be used for the corresponding dimension, and the conflict status pair, rejection reason, and reviewer identifier shall be recorded. When the cumulative number of newly added review samples reaches 10% of the current sample number, the corresponding dimension record reconstruction shall be triggered.

[0129] The closed-loop execution module retrieves three candidate monitoring periods using the current risk level, the feedback confirmation status, and the current data collection status as query keys, and then follows the... Determine the next monitoring cycle. This indicates the next monitoring cycle to be sent to the wearable device, in minutes. This indicates the candidate monitoring period corresponding to the risk level, in minutes. This indicates the candidate monitoring period corresponding to the confirmed status, in minutes. This indicates the candidate monitoring period corresponding to the data acquisition status, in minutes. This means that the smallest value is taken; this rule ensures that any state of increased risk, unconfirmed, or incomplete data collection will not be covered by a longer period.

[0130] When any single-dimensional query key in risk level, confirmation status, or data collection status is not matched, the closed-loop execution module first uses the minimum period of published records under the same status dimension type for the missing dimension. If there are no valid published records for the dimension, the hospital's configured safe default period is used. In this specific implementation, the safe default period is set to 15 minutes. After all three dimensions have obtained candidate periods, the next monitoring period is determined according to the aforementioned minimum value formula. The determined next monitoring period must be within the range of 5 minutes to 240 minutes supported by the wearable device and be an integer multiple of 5 minutes. If the original period is less than 5 minutes, it is set to 5 minutes; if it is greater than 240 minutes, it is set to 240 minutes. If it is within this range but is not an integer multiple of 5 minutes, it is discretized downwards to the nearest integer multiple of 5 minutes. After discretization, the 5-minute to 240-minute boundary clamp is executed again, and the original period, adjustment reason, and final period are saved in the period adjustment record. When the device is offline, the period adjustment instruction is written to the pending queue and the current period is maintained. At the same time, the offline status is fed back to the quality status of the next monitoring round in step S5.

[0131] The cycle adjustment instruction includes the patient hash value, device identifier, next monitoring cycle, execution start time, validity period, monitoring round number, and instruction signature. After the wearable device verifies the signature, it takes effect at the next sampling boundary and returns the current cycle, target cycle, and execution result. The closed-loop execution module determines the mutually exclusive final state based on the recommendation status and transaction result: a valid new recommendation is only marked as completed when the hospital medication update transaction is successfully submitted and the device returns a successful adjustment. If the hospital update succeeds but the device adjustment fails, the new medication record is retained, a device maintenance alarm is generated, and it is marked as device adjustment pending retry; confirmed medication records are not rolled back due to device failure. Rollback of existing plans or retention of physician prescriptions occurs only after successful confirmation and review of the record addition and the device returns a successful adjustment. When a work is completed, it is marked as a safe closed loop completed. When manual review is successful and the pharmacist review queue and safety alarm record are successfully added, and the equipment returns a successful adjustment, it is marked as a state awaiting manual review. The above three states of hospital medication records that do not change do not require the submission of hospital medication update transactions as a final state condition. After a valid newly recommended hospital transaction fails and is rolled back, it is marked as a state awaiting retry after confirming that the original valid record is preserved, the failure alarm is added, and the equipment cycle adjustment result is recorded. The retry entry uses the suggested version number and idempotent key to call the hospital interface again. When the addition of any record in a state that does not change fails or the equipment adjustment fails, it is marked as a closed loop record or a device adjustment awaiting retry, and the result code, failure stage, retry entry, and current medication record status are saved.

[0132] The closed-loop monitoring execution module writes the start time, end time, input object identifier, result code, failure reason, and number of retries for each stage of suggestion generation, encrypted transmission, terminal signature verification, identity verification, confirmation generation, hospital record update, and cycle adjustment into the closed-loop pharmaceutical monitoring execution record. The record is saved in an append-only manner and associated with risk reasoning evidence records, security verification records, and hospital new medication record identifiers by monitoring round number and suggestion version number. This allows the next round of steps S1 to determine the data status, step S2 to identify the conditions for using historical representations, and step S5 to read the current valid medication records, thereby forming a continuous pharmaceutical monitoring closed loop.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0134] This invention integrates multi-source data time alignment, patient medication sensitivity characterization, drug-related risk inference, and follow-up status structure into the same monitoring data chain. This enables new outpatient observations to work together with previous inpatient information to update the parameters of the individualized medication model. Furthermore, the model update results are converted into dosage and dosing interval recommendations through safety constraint screening.

[0135] Compared to processing structures that rely solely on fixed rules or periodic manual review, this invention employs dual-channel feature extraction and cross-modal gating fusion, risk threshold correction driven by liver and kidney function, incremental Bayesian updates based on historical posterior and prior recursion, and monitoring cycle linkage after suggestion confirmation feedback. This enables risk identification, parameter updates, and monitoring execution to form a continuous closed loop throughout the patient's entire lifecycle.

Claims

1. A method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data, characterized in that, include: S1. Acquire in-hospital diagnosis and treatment data, medication data, continuous physiological signals and follow-up data of oral tumor patients, and perform time alignment and correlation of data based on timestamps and patient identifiers to generate a patient time-series multimodal data base; S2. Input the patient’s temporal multimodal data into a dual-channel feature extraction network to extract the temporal variation features of continuous physiological signals and the long-range temporal correlation features of discrete medication events. Perform cross-modal gating fusion on the outputs of the two channels to generate a digital representation of medication sensitivity. S3. Input the digital representation of drug sensitivity and the knowledge graph of drug interactions into the graph attention network, perform relational reasoning on the drug, disease and metabolic associations involved in the candidate drug combination, and correct the risk trigger threshold according to the patient's liver and kidney function indicators to generate the risk level label of the candidate drug combination. S4. Perform text conversion and medical entity recognition on the speech in the follow-up data, standardize the identified symptom entities, and match the symptom entities with the statistical features of continuous physiological signals in the corresponding time period to generate a structured follow-up status summary vector. S5. The follow-up status summary vector and the digital representation of medication sensitivity are used as new observation data to input the incremental Bayesian inference framework. The parameters of the individualized medication model are used as latent variables and the historical posterior distribution is used as the current prior distribution for sequential updates. Based on the risk level label, the parameter subset that meets the medication safety constraints is selected from the updated parameter posterior distribution. Based on the parameter subset, the conditional expectation of the dosage and dosing interval is calculated to generate the dosing recommendation result. S6. Convert the medication recommendation results into pharmaceutical monitoring suggestions and send them to the patient's mobile terminal. Receive confirmation instructions associated with the patient's identity, transmit the confirmation information and current medication records back to the hospital information system to update the medication records, and adjust the monitoring cycle of the wearable device based on the transmitted results.

2. The method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data according to claim 1, characterized in that, S1 includes: Electronic medical records, pathology reports, and medication records are obtained from the hospital information system. Photoplethysmography (PPG), triaxial acceleration (TAC), and electrodermal response (EDS) signals are obtained from wearable devices, along with follow-up audio recordings with the acquisition time. The initial pathological diagnosis time recorded in the electronic medical record is determined as the patient's initial diagnosis time, and the timestamps of each data are converted to a unified timeline with the patient's initial diagnosis time as the origin. Structured fields, text fragments, and waveform fragments within the same time window are associated according to the patient's identifier, and indexes are established for the time dimension, data source dimension, and feature dimension, respectively. When an abnormal data acquisition is detected, a missing status marker is written at the corresponding index position without replacing it with data values ​​from other modalities, forming the patient's temporal multimodal data basis.

3. The method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data according to claim 1, characterized in that, S2 includes: Multi-scale causal convolution and dilated convolution are performed on the continuous physiological signal by the temporal convolution sub-channel to obtain local fluctuation features and cross-time window change features; The long-range temporal correlation features are obtained by calculating the attention weights between events using the self-attention sub-channel according to the drug identifier, dosing time, and dose code of the medication event; The features of the two sub-channels are normalized and fully connected. Gating weights are generated based on the missing state markers and feature validity of each modality. The mapped features are then weighted, fused, and pooled using the gating weights to generate the drug sensitivity digital representation. When neither sub-channel has a valid feature, the generation of a new drug sensitivity digital representation is stopped and a data supplementation marker is written. If a drug sensitivity digital representation from the previous monitoring round exists, it is used as the input for steps S3 and S5. If a drug sensitivity digital representation from the previous monitoring round does not exist, subsequent automatic recommendations are stopped and a pharmacist review alarm is generated.

4. The method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data according to claim 1, characterized in that, S3 includes: establishing drug nodes, disease nodes, and metabolism-related nodes in the drug interaction knowledge graph, and configuring drug-drug antagonism, drug-drug synergistic effects, drug-disease contraindication, and drug-gene metabolism association; fusing the node features involved in the candidate drug combination with the drug sensitivity digital representation and inputting them into the graph attention network; aggregating adjacent node information according to the attention weight of each relation edge, and outputting the probability of drug combination conflict and the probability of adverse reaction type; correcting the risk trigger threshold corresponding to each adverse reaction type according to the liver and kidney function index vector formed by estimated glomerular filtration rate, serum creatinine, alanine aminotransferase, and total bilirubin; comparing the probability with the corrected risk trigger threshold to generate the risk level label.

5. The method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data according to claim 1, characterized in that, S4 includes: performing word segmentation, part-of-speech tagging, dependency parsing, and relation tree pruning on the text obtained from the follow-up speech conversion; using a medical language model to identify drug names, adverse reaction symptoms, symptom severity, and medication time entities, and converting the adverse reaction symptoms into standard medical terminology codes according to a medical terminology coding table; establishing a time matching window centered on the symptom occurrence time, the length of which is determined based on the elimination half-life of the corresponding drug and the sampling interval of the wearable device; extracting statistical features of heart rate, blood oxygen saturation, and skin conductance within the time matching window, matching the statistical features with the coding vectors of medical entities according to the acquisition time, and performing attention weighting to generate the follow-up status summary vector; when no symptom entity is identified or there is no effective physiological signal segment within the time matching window, writing a missing status marker into the corresponding component of the follow-up status summary vector; when both types of inputs are missing or cannot be matched, not using the follow-up status summary vector of this round as quantitative observation data, and generating supplementary follow-up and physiological signal supplementation instructions.

6. The method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data according to claim 1, characterized in that, S5 includes: setting drug clearance rate, volume of distribution, and absorption rate constant as parameters of the individualized medication model; calculating the observation likelihood based on the newly added observation data; and combining the observation likelihood with the historical posterior distribution to obtain the parameter posterior distribution for the current round; converting the single dose range and dosing interval range, the therapeutic drug monitoring reference concentration range, and the risk level marker in the approved medication instructions into the medication safety constraints; selecting a subset of parameters from the parameter posterior distribution that satisfy the medication safety constraints for the predicted peak blood drug concentration, the predicted trough blood drug concentration, and the cumulative exposure; and calculating the posterior mean and posterior quantile interval of the administered dose and dosing interval under the parameter subset conditions, respectively, as the dosing recommendation result and its confidence interval.

7. The method for pharmaceutical monitoring and management of oral tumor patients based on multimodal data according to claim 6, characterized in that, S6 include: The risk level marker, along with the drug name, single dose, daily dosing frequency, safety interval, and confidence interval, are written into a formatted pharmaceutical monitoring recommendation. This recommendation is then encrypted and signed for integrity before being sent to the patient's mobile terminal. A confirmation command is generated only when the biometric verification result on the patient's mobile terminal matches the patient's identifier. Upon receiving the confirmation information and the current medication record, the hospital information system marks past medication records as expired and inserts a new medication record with the confirmation time and recommendation version number. A status-cycle mapping table is established based on historical monitoring samples. This table records candidate monitoring cycles corresponding to each risk level marker, confirmation status, and data collection status. Candidate monitoring cycles are obtained from the status-cycle mapping table based on the current risk level marker, the returned confirmation status, and the data collection status. The candidate monitoring cycle with the smallest value is determined as the next monitoring cycle for the wearable device. The sending, verification, updating, and cycle adjustment statuses are written into the closed-loop pharmaceutical monitoring execution record.

8. The method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data according to claim 4, characterized in that, The correction of the risk trigger threshold includes: calculating the distance of each indicator in the liver and kidney function index vector from the boundary of the corresponding reference interval according to the hospital's test reference interval; determining the threshold correction amount according to the probability distribution of each adverse reaction type in historical monitoring samples under the distance condition; using the threshold correction amount to constrain the direction and magnitude of change of the risk trigger threshold, and recording the indicators involved in the correction, the reference interval version, the probability distribution version, and the reason for the correction, forming a risk reasoning evidence record.

9. The method for pharmaceutical monitoring and management of oral cancer patients based on multimodal data according to claim 7, characterized in that, When the subset of parameters satisfying the medication safety constraints is an empty set or the confidence interval of the posterior distribution of the parameters exceeds the credible range determined by the quantile boundaries of historical monitoring samples, the generation of new recommended values ​​for dosage and dosing interval is stopped; if a current valid medication record exists, the dosing regimen corresponding to the current valid medication record is rolled back; if no current valid medication record exists, the case is transferred to the pharmacist manual review state and the existing prescription instructions confirmed by the physician are retained; if no existing prescription instructions exist, the dosage and dosing interval values ​​are not output; a pharmacist review alarm is generated, and the next monitoring cycle of the wearable device is determined according to the state-cycle mapping table; after receiving pharmacist review confirmation and new observation data, S5 is re-executed, and the cause of the anomaly, the rollback plan, the review result, and the re-inference result are written into the safety verification record.

10. A pharmaceutical monitoring and management system for oral cancer patients based on multimodal data, used to execute the pharmaceutical monitoring and management method for oral cancer patients based on multimodal data as described in any one of claims 1 to 9, characterized in that, include: The data infrastructure construction module is used to perform step S1; The digital representation generation module is used to execute step S2; The medication risk reasoning module is used to execute step S3; The follow-up status parsing module is used to execute step S4; the medication model update module is used to execute step S5. The monitoring closed-loop execution module is used to execute step S6.

Images