An ai-based cerebral vascular disease risk dynamic prediction method and system

Abstract

The application discloses an AI-based cerebrovascular disease risk dynamic prediction method and system, relates to the technical field of intelligent medical data processing, and comprises the following steps: based on an annotated event stream, taking a corresponding time point as a time window boundary when an event trigger condition is met, dynamically adjusting a time window span with an irregular sampling density, generating a dynamic time window set, and mapping the dynamic time window set into a time window sample package; inputting the time window sample package into a risk prediction model based on a time interval attenuation gate cycle, obtaining an original risk probability value, and synchronously acquiring drift evidence deviating from a previous state of an individual; performing online recalibration on the original risk probability value according to the drift evidence, updating an individualized baseline calibration factor, forming a calibrated risk probability value, and simultaneously performing effectiveness verification to obtain a final effective risk probability value. The application realizes stable and dynamic prediction of cerebrovascular disease risk.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data processing technology, and in particular to an AI-based method and system for dynamic prediction of cerebrovascular disease risk. Background Technology

[0002] In recent years, cerebrovascular disease risk assessment technology has gradually evolved from traditional static assessments based on risk factor regression and scoring scales to intelligent predictions that integrate multi-source medical and health data such as electronic medical records, laboratory tests, vital sign monitoring, and brain imaging. With the improvement of medical information systems and data availability, artificial intelligence methods have been applied in feature representation learning, temporal modeling, and multimodal fusion, enabling the output of risk probabilities that are updated as the disease progresses. Simultaneously, addressing the objective existence of irregular time intervals and missing observations, the industry has begun to employ mechanisms such as time interval decay and gated loops to characterize information staleness, attempting to improve adaptability to the rhythm of real clinical data feedback and driving risk assessment from offline analysis to continuous management scenarios.

[0003] However, existing technologies still have two shortcomings closely related to dynamic prediction. First, the time organization method often relies on fixed-length time windows or fixed-frequency updates. When data arrives asynchronously and is sampled irregularly, it is difficult to adaptively adjust the time window span according to changes in information density. This leads to insufficient evidence in the sparse phase and dilution of key changes in the dense phase, resulting in fluctuations or lags in risk output. Second, there are insufficient means to control the stability of probability output during operation. There are few closed-loop mechanisms for evidence of drift "deviating from the individual's previous state," online recalibration, and validity verification. Changes in individual treatment plans, changes in follow-up criteria, or long-term baseline drift can easily cause systematic shifts in probability values, affecting the continuous availability and consistency in clinical applications. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI-based dynamic prediction method for cerebrovascular disease risk to address the problems of difficulty in adapting time window evidence organization under irregular sampling conditions and unstable risk probability output caused by individual baseline drift.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, this invention provides an AI-based method for dynamic prediction of cerebrovascular disease risk, comprising: connecting multi-source medical and health data of a target individual to the same acquisition channel and writing a unified time stamp; sorting the multi-source medical and health data according to the time stamp and writing it into a health event sequence carrier; establishing event triggering conditions based on the health event sequence carrier, and writing newly arriving multi-source medical and health data into an labeled event stream after triggering labeling; based on the labeled event stream, using the corresponding time point as the boundary of the time window when the event triggering conditions are met, and dynamically adjusting the time window span according to the irregular sampling density to generate a dynamic time window set, which is mapped as a time window sample package; inputting the time window sample package into a risk prediction model based on time interval decay gating loop to obtain the original risk probability value, and simultaneously acquiring drift evidence deviating from the individual's previous state; performing online recalibration on the original risk probability value based on the drift evidence, and updating the individualized baseline calibration factor to form a calibrated risk probability value, while performing validity verification to obtain the final effective risk probability value; binding the final effective risk probability value with the corresponding time window identifier to generate a patient status record, and writing it back to the health event sequence carrier for closed-loop optimization.

[0008] As a preferred embodiment of the AI-based dynamic prediction method for cerebrovascular disease risk described in this invention, the specific steps for writing a unified time stamp are as follows:

[0009] The multi-source medical and health data of the target individual are written one by one into the receiving buffer of the same acquisition channel, deduplication verification is performed, and multi-source medical and health data that has passed the deduplication verification is generated.

[0010] Read the record time field of the multi-source medical and health data that has passed the deduplication verification, convert it into millisecond-level timestamps and write it into a unified timestamp field to generate multi-source medical and health data with unified timestamps.

[0011] As a preferred embodiment of the AI-based dynamic prediction method for cerebrovascular disease risk described in this invention, the specific steps for writing the health event sequence into the carrier are as follows:

[0012] Sort the multi-source medical and health data with the same time stamp in ascending order according to the time stamp, and generate sorted multi-source medical and health data;

[0013] The sorted multi-source medical and health data are appended to the health event sequence carrier one by one according to the sorting order.

[0014] As a preferred embodiment of the AI-based dynamic prediction method for cerebrovascular disease risk described in this invention, the specific steps for writing newly arrived multi-source medical and health data into the labeled event stream after trigger annotation are as follows:

[0015] Extract the source identifier sequence and the unified time-stamp sequence of multi-source medical and health data from the health event sequence carrier according to the unified time stamp, and generate the arrival interval sequence by grouping according to the source identifier;

[0016] The arrival interval sequence is sorted in ascending order to obtain a sorted arrival interval sequence, and event triggering conditions are established by combining the boundary interval threshold.

[0017] Receive newly arrived multi-source medical and health data, read the source identifier field and the unified timestamp field, combine them with the event triggering conditions to perform trigger annotation, and write the annotated event stream.

[0018] As a preferred embodiment of the AI-based dynamic prediction method for cerebrovascular disease risk described in this invention, the step of using the corresponding time point as the time window boundary when the event triggering condition is met is as follows:

[0019] The trigger annotations are written into the multi-source medical and health data that has not yet been triggered, according to the labeled event flow, and recorded as the right boundary record of the time window;

[0020] The unified time marker recorded at the right boundary of the time window is used as the right boundary time marker of the time window, and a backtracking sampling method based on the source identifier is performed in the labeled event stream to form the time window boundary.

[0021] As a preferred embodiment of the AI-based dynamic prediction method for cerebrovascular disease risk described in this invention, the mapping is a time window sample package, and the specific steps are as follows:

[0022] By selecting multi-source medical and health data from the labeled event stream according to a unified time mark, which fall within the range of the time mark from the left boundary of the time window to the right boundary of the time window, a dynamic time window set is obtained.

[0023] Based on the dynamic time window set, the multi-source medical and health data corresponding to the time window boundary are grouped according to the record type, and the grouping results, along with the time markers of the left and right boundaries of the time window, are written into the time window sample package.

[0024] As a preferred embodiment of the AI-based dynamic prediction method for cerebrovascular disease risk described in this invention, the specific steps for obtaining the original risk probability value and simultaneously acquiring drift evidence deviating from the individual's past state are as follows.

[0025] Based on time window sample packages, multi-source medical and health data in the grouping results are organized sequentially according to a unified time stamp to obtain a risk prediction sequence.

[0026] A risk prediction model based on time interval decay gated recurrent units (GRU), residual convolutional neural networks, and fully connected neural networks is constructed.

[0027] The historical risk prediction sequence is labeled and bound to the cerebrovascular disease outcome events that are adjacent in time in the health event sequence carrier to form a training sample set, and then the validation sample set is obtained.

[0028] The training sample set is input into the risk prediction model to perform supervised training, and the effectiveness of the training process is determined by the validation sample set to obtain the trained risk prediction model.

[0029] Input the risk prediction sequence into the trained risk prediction model, perform irregular sampling time series probability inference, and output the original risk probability value;

[0030] Field values ​​are extracted from the grouping results of the time window sample package and the previous state time window sample package, and item-by-item comparison processing is performed to generate drift evidence.

[0031] As a preferred embodiment of the AI-based dynamic prediction method for cerebrovascular disease risk described in this invention, the specific steps for obtaining the final effective risk probability value are as follows:

[0032] The deviation direction, deviation magnitude level, and deviation persistence are extracted from drift evidence, and individualized baseline calibration factor values ​​are extracted from health event sequence carriers;

[0033] The deviation direction, deviation magnitude level, deviation persistence, and individualized baseline calibration factor values ​​are bound to the online recalibration context;

[0034] Based on the online recalibration context, probability mapping calibration is performed on the original risk probability value to obtain the calibrated risk probability value, and the individualized baseline calibration factor is rolled over with the deviation persistence as the update condition to obtain the updated individualized baseline calibration factor.

[0035] The validity of the calibrated risk probability value is verified under three types of constraints: consistency of probability range, consistency of calibration magnitude, and continuity of time, and the final effective risk probability value is obtained.

[0036] As a preferred embodiment of the AI-based dynamic prediction method for cerebrovascular disease risk described in this invention, the specific steps for generating patient status records are as follows:

[0037] Read the left and right boundary time markers of the time window corresponding to the final effective risk probability value from the time window sample package, and generate the time window identifier using a fixed splicing rule;

[0038] A unified time stamp, source identifier, and record type identifier are written into the time window identifier and the final effective risk probability value to generate a patient status record.

[0039] Secondly, this invention provides an AI-based dynamic prediction system for cerebrovascular disease risk, comprising: a time-stamping module for connecting multi-source medical and health data of a target individual to the same acquisition channel and writing a unified time stamp, sorting the multi-source medical and health data according to the time stamp, and writing it into a health event sequence carrier; a triggering and labeling module for establishing event triggering conditions based on the health event sequence carrier, and writing newly arriving multi-source medical and health data into the labeled event stream after triggering and labeling; and a time window sampling module for using the corresponding time point as the time window boundary when the event triggering conditions are met, based on the labeled event stream, and dynamically adjusting the time window according to irregular sampling density. The system comprises four modules: a time span module, a dynamic time window set, and a time window sample package; a risk drift module, which inputs the time window sample package into a risk prediction model based on time interval decay gating loop to obtain the original risk probability value and simultaneously acquire drift evidence deviating from the individual's past state; a calibration and verification module, which performs online recalibration on the original risk probability value based on the drift evidence and updates the individualized baseline calibration factor to form a calibrated risk probability value, while performing validity verification to obtain the final effective risk probability value; and a write-back closed-loop module, which binds the final effective risk probability value with the corresponding time window identifier, generates a patient status record, and writes it back to the health event sequence carrier for closed-loop optimization.

[0040] The beneficial effects of this invention are as follows: by constructing a dynamic time window driven by event triggering, the time window span is adaptively adjusted according to the irregular sampling density to form a time window sample package; then, the original risk probability is output by a time interval decay gated loop model, and online recalibration and validity verification are driven by drift evidence. Combined with the write-back closed loop, stable dynamic prediction of cerebrovascular disease risk is achieved. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of an AI-based dynamic prediction method for cerebrovascular disease risk.

[0043] Figure 2 This is a schematic diagram of an AI-based dynamic prediction system for cerebrovascular disease risk.

[0044] Figure 3A flowchart for constructing event triggering conditions and trigger annotations.

[0045] Figure 4 This is a diagram showing the comparison and magnified view of risk probability trajectories. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Reference Figures 1-4 This is one embodiment of the present invention, which provides an AI-based method for dynamic prediction of cerebrovascular disease risk, comprising the following steps:

[0050] S1: Connect the multi-source medical and health data of the target individual to the same acquisition channel, write a unified time stamp, sort the multi-source medical and health data according to the time stamp, and write it into the health event sequence carrier;

[0051] S1.1: Write the multi-source medical and health data of the target individual into the receiving buffer of the same acquisition channel one by one, perform deduplication verification, and generate multi-source medical and health data that has passed the deduplication verification;

[0052] Furthermore, after multi-source medical and health data arrive at the same acquisition channel, the multi-source medical and health data are written into the receiving buffer of the same acquisition channel one by one in the order of arrival. The same acquisition channel reads the source identifier field, record time field, and record type identifier for each multi-source medical and health data in the receiving buffer, and performs consistency verification on the source identifier field and record type identifier before performing deduplication verification. Deduplication verification adopts a combination matching method of source identifier field, record time field, and record type identifier. If the same combination matching result already exists in the receiving buffer, writing stops and duplicate records are discarded. If the same combination matching result does not exist in the receiving buffer, the record is retained and the multi-source medical and health data that has passed the deduplication verification is output.

[0053] S1.2: Read the record time field of the multi-source medical and health data that has passed the deduplication verification, convert it into a millisecond-level time stamp and write it into a unified time stamp field to generate multi-source medical and health data with unified time stamps;

[0054] Furthermore, for the multi-source medical and health data that has passed the deduplication verification, the record time field is read one by one and time parsing processing is performed. The time parsing processing converts the record time field into a millisecond-level time stamp under the same time base. The millisecond-level time stamp is written into a unified time stamp field. After the writing is completed, the multi-source medical and health data with unified time stamp is output.

[0055] S1.3: Sort the multi-source medical and health data with the same time stamp in ascending order according to the time stamp, and generate sorted multi-source medical and health data;

[0056] Furthermore, using a unified timestamp as the timestamp key value, the multi-source medical and health data with the same timestamp are sorted in ascending order. During the sorting process, the writing order of the receiving buffer remains unchanged when the timestamp key values ​​are the same, so as to ensure the stability of the record order under the same timestamp. After the sorting is completed, the sorted multi-source medical and health data is output.

[0057] It should be noted that the deduplication check is performed at the level of the combination of the source identifier field, the record time field, and the record type identifier. If the combination matching results are consistent, it is determined to be a duplicate. The millisecond-level time stamp conversion takes the parsable time information of the record time field as input, unifies the record time fields from different sources into the same time base expression, and writes them into the unified time stamp field, thereby ensuring that the subsequent sorting by time stamp has a consistent basis.

[0058] S1.4: The sorted multi-source medical and health data are appended to the health event sequence carrier one by one according to the sorting order;

[0059] Furthermore, the data is written into the health event sequence carrier one by one in ascending order within the sorted multi-source medical and health data. During writing, the unified time stamp field of the multi-source medical and health data is written together with the other fields of the multi-source medical and health data into the same record position of the health event sequence carrier. After the writing is confirmed, this step ends and the health event sequence carrier is obtained.

[0060] S2: Based on the health event sequence carrier, establish event triggering conditions, and write newly arriving multi-source medical and health data into the labeled event stream after triggering and labeling;

[0061] S2.1: Extract the source identifier sequence and the unified time-stamp sequence of multi-source medical and health data from the health event sequence carrier according to the unified time stamp, and generate the arrival interval sequence by grouping according to the source identifier;

[0062] Furthermore, multi-source medical and health data are sequentially read from the health event sequence carrier according to a unified time stamp, and the source identifier field is extracted and written into the source identifier sequence in turn. At the same time, the unified time stamp field is extracted and written into the unified time stamp sequence. After the source identifier sequence and the unified time stamp sequence are extracted, the unified time stamp sequence is grouped and organized according to the source identifier sequence to form a unified time stamp group sequence corresponding to each source identifier. In each unified time stamp group sequence, adjacent records are arranged in ascending order according to the unified time stamp field. Differential processing is performed on the unified time stamp fields of two adjacent records to obtain the arrival interval, and the arrival interval sequence is written in chronological order.

[0063] S2.2: Sort the arrival interval sequence in ascending order to obtain the sorted arrival interval sequence, and establish event triggering conditions based on the boundary interval threshold;

[0064] Furthermore, the arrival interval sequence is sorted in ascending order to obtain a sorted arrival interval sequence. The first quartile value and the third quartile value are extracted from the sorted arrival interval sequence, and the difference between the third quartile value and the first quartile value is taken as the interquartile range. The sum of the third quartile value and the interquartile range is written into the boundary interval threshold. The boundary interval threshold is associated with the source identifier field and written into the event triggering condition. After the event triggering condition is established, the searchable event triggering condition is obtained.

[0065] It should be noted that the boundary interval threshold is derived from the quantile statistics and quantile distance calculation of the sorted arrival interval sequence. The boundary interval threshold is used to transform the data return rhythm corresponding to the source identifier into a comparable trigger criterion, thereby supporting subsequent trigger annotation of newly arriving multi-source medical and health data.

[0066] S2.3: Receive newly arrived multi-source medical and health data, read the source identifier field and the unified time stamp field, perform trigger annotation based on the event trigger conditions, and write the annotated event stream;

[0067] Furthermore, upon receiving newly arrived multi-source medical and health data, the source identifier field and unified timestamp field of the newly arrived multi-source medical and health data are read, and the boundary interval threshold corresponding to the source identifier field is retrieved in the event triggering conditions. In the health event sequence carrier, multi-source medical and health data records with the same source identifier field and the largest unified timestamp field are retrieved. The unified timestamp field of the multi-source medical and health data record is read and differentially processed with the unified timestamp field of the newly arrived multi-source medical and health data to obtain the new arrival interval. The new arrival interval is compared with the boundary interval threshold. If the new arrival interval is greater than the boundary interval threshold, a trigger label is written as triggered; if the new arrival interval is not greater than the boundary interval threshold, a trigger label is written as non-triggered. The trigger label is bound to the newly arrived multi-source medical and health data and written to the labeled event stream, thus completing the trigger labeling of the newly arrived multi-source medical and health data and writing it to the labeled event stream.

[0068] S3: Based on the labeled event stream, when the event triggering conditions are met, the corresponding time point is used as the boundary of the time window, and the time window span is dynamically adjusted according to the irregular sampling density to generate a dynamic time window set, which is mapped to the time window sample bag;

[0069] It should be noted that the time window span is dynamically adjusted according to the irregular sampling density through the source identifier backtracking time mark. The source identifier backtracking time mark is obtained by comparing the backtracking arrival interval with the boundary interval threshold. When the backtracking arrival interval is too large, the source identifier backtracking time mark moves forward; when the backtracking arrival interval is too small, the source identifier backtracking time mark moves backward. The time mark of the left boundary of the time window moves forward and backward along with the earliest source identifier backtracking time mark, so that the span from the time mark of the left boundary of the time window to the time mark of the right boundary of the time window changes with the irregular sampling density.

[0070] S3.1: Write the trigger annotations into the multi-source medical and health data that has not been triggered according to the labeled event flow, and record them as the right boundary record of the time window;

[0071] Furthermore, multi-source medical and health data are sequentially retrieved in the labeled event stream according to a unified time stamp. Multi-source medical and health data with trigger annotations are filtered and selected from the filtered results. The multi-source medical and health data with the largest unified time stamp field is recorded as the right boundary record of the time window. The right boundary record of the time window is output for subsequent generation of the right boundary time stamp of the time window, thus obtaining the right boundary record of the time window.

[0072] S3.2: Use the unified time stamp recorded at the right boundary of the time window as the right boundary time stamp of the time window, and perform backtracking sampling based on the source identifier in the labeled event stream to form the time window boundary;

[0073] Furthermore, the unified timestamp field of the record at the right boundary of the time window is read and written to the right boundary timestamp of the time window. The source identifier and boundary interval threshold registered in the event triggering conditions are read. For each source identifier in the event triggering conditions, multi-source medical and health data with the same source identifier field and a unified timestamp field not greater than the right boundary timestamp of the time window are retrieved from the labeled event stream. The unified timestamp with the largest unified timestamp field is extracted as the most recent unified timestamp, and the second largest unified timestamp field is extracted as the previous unified timestamp. If the previous unified timestamp is not found in the labeled event stream, the most recent unified timestamp will be used. A unified time stamp is written to the previous unified time stamp; the difference between the most recent unified time stamp and the previous unified time stamp is calculated to obtain the backtracking arrival interval, and the backtracking arrival interval is compared with the boundary interval threshold. If the backtracking arrival interval is greater than the boundary interval threshold, the previous unified time stamp is written to the source identifier backtracking time stamp; if the backtracking arrival interval is not greater than the boundary interval threshold, the most recent unified time stamp is written to the source identifier backtracking time stamp; all source identifier backtracking time stamps are summarized and the earliest unified time stamp is selected and written to the left boundary time stamp of the time window. The left boundary time stamp of the time window and the right boundary time stamp of the time window together form the time window boundary.

[0074] S3.3: Select multi-source medical and health data from the labeled event stream according to a unified time mark, where the unified time mark falls within the range of the time mark from the left boundary of the time window to the right boundary of the time window, to obtain a dynamic time window set;

[0075] Furthermore, the left and right boundary timestamps of the time window are read. Multi-source medical and health data whose unified timestamp fields fall within the range of the left to right boundary timestamps of the time window are filtered in the labeled event stream according to the unified timestamp. The filtered results are organized in ascending order according to the unified timestamp fields and written into the dynamic time window set to obtain the dynamic time window set.

[0076] S3.4: Based on the dynamic time window set, the multi-source medical and health data corresponding to the time window boundary are grouped according to the record type, and the grouping results, together with the time mark of the left boundary of the time window and the time mark of the right boundary of the time window, are written into the time window sample package;

[0077] Furthermore, the dynamic time window set is read and the multi-source medical and health data within the dynamic time window set is traversed. The record type identifier is read one by one and grouped according to the record type identifier and written into the grouping result. The grouping result, along with the time mark of the left boundary of the time window and the time mark of the right boundary of the time window, are written into the time window sample package to obtain the time window sample package.

[0078] S4: Input the time window sample package into the risk prediction model based on the time interval decay gating cycle to obtain the original risk probability value, and simultaneously obtain drift evidence that deviates from the individual's past state;

[0079] S4.1: Based on the time window sample package, the multi-source medical and health data in the grouping results are organized sequentially according to a unified time stamp to obtain a risk prediction sequence;

[0080] Furthermore, the grouping results in the time window sample package are read, and the multi-source medical and health data in the grouping results are organized in ascending order according to the unified time mark. The writing order of multi-source medical and health data with the same unified time mark is kept unchanged in the time window sample package. The record type identifier is read one by one in the multi-source medical and health data after ascending order organization, and the record type identifier, unified time mark field and corresponding field value are written into the risk prediction sequence in chronological order to obtain the risk prediction sequence.

[0081] S4.2: Construct a risk prediction model based on time interval decay gated loops using gated recurrent units (GRU), residual convolutional neural networks, and fully connected neural networks;

[0082] Furthermore, the risk prediction model based on time-lapse decay gated loops consists of a gated recurrent unit (GRU), exponential decay, missing item labeling, a residual convolutional neural network (RNN), a fully connected neural network (WNN), and a sigmoid mapping. The residual convolutional neural network extracts features from the field values ​​corresponding to the brain image class identified by the record type identifier and outputs image feature values. The missing item labeling writes missing item labels into the missing fields in the risk prediction sequence and organizes the missing item labels and field values ​​together as a temporal input. The exponential decay generates time decay weights based on the difference between adjacent unified time labels in the risk prediction sequence and performs time-lapse decay on the temporal update process of the gated recurrent unit (GRU). The gated recurrent unit (GRU) performs state recursion on the temporal input and outputs temporal state values. The fully connected neural network performs probability mapping on the temporal state values ​​and outputs probability logarithmic values. The sigmoid mapping performs probabilistic mapping on the probability logarithmic values ​​and outputs the original risk probability values, thus completing the construction of the risk prediction model based on time-lapse decay gated loops.

[0083] The Sigmoid mapping performs a probabilistic mapping on the logarithmic values ​​of the probability and outputs the original risk probability value, expressed as:

[0084] ;

[0085] in, This represents the original risk probability value. The positive class probability logarithm value of the output of the fully connected neural network corresponds to the positive class component in the probability logarithm value obtained by the fully connected neural network performing probability mapping on the temporal state values; The negative class probability logarithm of the output of the fully connected neural network corresponds to the negative class component in the probability logarithm obtained by performing probability mapping on the temporal state values ​​of the fully connected neural network.

[0086] It should be noted that the missing item marking process uses the absence of a value written into a field within the risk prediction sequence as the missing criterion and writes it into the missing item mark. The exponential decay uses the difference between adjacent unified time marks within the risk prediction sequence as input to generate the time decay weight. The time decay weight is used to decay the time interval of the time-series update of the gated recurrent unit (GRU), thereby supporting the inference of irregular sampling time-series probability and outputting the original risk probability value.

[0087] S4.3: Bind the historical risk prediction sequence to the cerebrovascular disease outcome events that are adjacent in time in the health event sequence carrier to form a training sample set, and divide it to obtain a validation sample set;

[0088] Furthermore, time window sample packages corresponding to historical intervals are extracted from the health event sequence carrier, and historical risk prediction sequences are generated. For each historical risk prediction sequence, a unified time marker at the end of the historical risk prediction sequence is extracted as an anchor unified time marker. In the health event sequence carrier, the occurrence marker of cerebrovascular disease outcome events with a unified time marker greater than the anchor unified time marker and the smallest time difference is retrieved. The occurrence marker of cerebrovascular disease outcome events is bound to the historical risk prediction sequence to form a training sample set. The training sample set is divided into a training sample set and a validation sample set according to the unified time marker to obtain the training sample set and the validation sample set.

[0089] S4.4: Input the training sample set into the risk prediction model to perform supervised training, and use the verification sample set to determine the effectiveness of the training process, and obtain the trained risk prediction model;

[0090] Furthermore, the training sample set is input into the risk prediction model based on time-decay gated loops for supervised training. During the supervised training process, the error of the risk prediction model based on time-decay gated loops is calculated based on the occurrence markers of cerebrovascular disease outcome events in the training sample set, and the parameters of the gated recurrent unit (GRU), residual convolutional neural network, and fully connected neural network are iteratively updated. During the iterative update process, the validation sample set is input into the risk prediction model based on time-decay gated loops to output the validation results, and the validity determination is performed based on the validation results. After the validity determination is passed, the trained risk prediction model based on time-decay gated loops is obtained.

[0091] S4.5: Input the risk prediction sequence into the trained risk prediction model, perform irregular sampling time series probability inference, and output the original risk probability value;

[0092] Furthermore, the risk prediction sequence is input into the trained risk prediction model based on time interval decay gating loop. The trained risk prediction model based on time interval decay gating loop performs irregular sampling time-series probability inference on the risk prediction sequence and outputs the original risk probability value.

[0093] S4.6: Extract field values ​​from the grouping results of the time window sample package and the previous state time window sample package respectively, and perform item-by-item comparison processing to generate drift evidence;

[0094] Furthermore, the grouping results of the time window sample package are read and the set of field values ​​are extracted. The grouping results of the past state time window sample package are read and the set of field values ​​is extracted. After aligning the set of field values ​​according to the record type identifier, item-by-item comparison processing is performed. The item-by-item comparison processing outputs the difference direction and difference magnitude for each field value, and writes the difference direction and difference magnitude into the drift evidence to obtain the drift evidence.

[0095] S5: Based on the drift evidence, perform online recalibration on the original risk probability value and update the individualized baseline calibration factor to form a calibrated risk probability value. At the same time, perform validity verification to obtain the final effective risk probability value.

[0096] S5.1: Extract deviation direction, deviation magnitude level, and deviation persistence from drift evidence, and extract individualized baseline calibration factor values ​​from health event sequence carriers;

[0097] Furthermore, the drift evidence is read and the difference direction and difference magnitude corresponding to each field value within the drift evidence are extracted. The difference direction is synthesized into a deviation direction according to a preset direction aggregation rule, and the difference magnitude is mapped into a deviation magnitude level according to a preset grading rule. The continuous occurrence of the difference direction in adjacent time windows is mapped into deviation persistence according to a preset persistence determination rule. Records with record type identifiers written as individualized baseline calibration factor values ​​are selected from the health event sequence carrier, and records with a unified time mark field that is not greater than the right boundary time mark of the time window and have the largest unified time mark field are selected. The individualized baseline calibration factor values ​​are obtained by reading the selected records.

[0098] It should be noted that the individualized baseline calibration factor values ​​are derived from the historical data and rolling write-back within the health event sequence carrier. Specifically, the source chain consists of two parts: initialization writing and runtime update writing. In the initialization writing phase, after supervised training of the risk prediction model based on time-lapse gating loops using the training and validation sample sets, the systematic deviation between the cerebrovascular disease outcome event occurrence markers corresponding to the training sample set and the original risk probability values ​​output by the historical risk prediction sequences is used to form a calibration magnitude table according to the deviation magnitude level and write it into the individualized baseline calibration factor values. Then, the individualized baseline calibration factor values ​​are written into the individualized baseline calibration factor value using the record type identifier. The records of the online calibration factor values ​​are appended to the health event sequence carrier. During the runtime update and write phase, after each probability mapping calibration is completed based on the online recalibration context, when the deviation persistence meets the rolling update condition, the individualized baseline calibration factor values ​​are subjected to a tiered migration to obtain the updated individualized baseline calibration factor. The updated individualized baseline calibration factor is then appended to the health event sequence carrier again in the form of a record of individualized baseline calibration factor values, with the record type identifier as the appended record. Therefore, the individualized baseline calibration factor values ​​extracted from the health event sequence carrier come from the latest individualized baseline calibration factor value record in the unified time stamp field of the health event sequence carrier.

[0099] The preset direction aggregation rule takes the difference direction and difference magnitude corresponding to the value of each field in the drift evidence as input. First, it aggregates the difference magnitude according to the record type identifier, and then performs weighted synthesis on the aggregated values ​​of the record type identifier. The direction synthesis result is written into the deviation direction according to the positive and negative direction dominance relationship of the weighted synthesis, thereby avoiding the occasional difference of a single field value from dominating the deviation direction.

[0100] The preset grading rule takes the sum of difference magnitude within the drift evidence as input, extracts the historical difference magnitude sum of value sequence of the same target individual from the health event sequence carrier, and calculates the value of the first quartile, the value of the third quartile, and the interquartile range. The quartiles and the interquartile range are combined to form the boundary point sequence, and the boundary interval into which the current difference magnitude sum of value falls is mapped to the deviation magnitude level, so that the deviation magnitude level adapts to the historical fluctuation range of the target individual.

[0101] The preset persistence determination rule takes the deviation direction sequence of drift evidence corresponding to adjacent time windows in the health event sequence carrier as input, retrieves adjacent time windows in reverse order according to the unified time mark field, and counts the length of time windows with continuous and consistent deviation directions and the total number of retrieved time windows. When the degree of continuity and consistency falls within the preset proportion range, the deviation persistence is written, so that the deviation persistence reflects both the continuity of the deviation direction and the integrity of the time window coverage.

[0102] S5.2: Bind the deviation direction, deviation magnitude level, deviation persistence and individualized baseline calibration factor value into an online recalibration context;

[0103] Furthermore, the deviation direction, deviation magnitude level, deviation persistence, and individualized baseline calibration factor value are written into the same online recalibration context. Within the online recalibration context, the deviation direction field, deviation magnitude level field, deviation persistence field, and individualized baseline calibration factor value field are written respectively. After the online recalibration context is written, the online recalibration context is obtained.

[0104] S5.3: Based on the online recalibration context, perform probability mapping calibration on the original risk probability value to obtain the calibrated risk probability value, and perform rolling update on the individualized baseline calibration factor with deviation persistence as the update condition to obtain the updated individualized baseline calibration factor.

[0105] Furthermore, the online recalibration context is read, and the calibration magnitude corresponding to the deviation magnitude level is retrieved from the individualized baseline calibration factor values. When the deviation direction is written as upward deviation, the original risk probability value is adjusted upward according to the calibration magnitude to obtain the calibrated risk probability value. When the deviation direction is written as downward deviation, the original risk probability value is adjusted downward according to the calibration magnitude to obtain the calibrated risk probability value. When the deviation direction is written as no deviation, the original risk probability value is written into the calibrated risk probability value. When the deviation persistence meets the rolling update condition, the correspondence between the deviation magnitude level and the calibration magnitude in the individualized baseline calibration factor values ​​is performed in a graded migration according to the deviation direction to obtain the updated individualized baseline calibration factor. When the deviation persistence does not meet the rolling update condition, the individualized baseline calibration factor values ​​are kept unchanged and written into the updated individualized baseline calibration factor to obtain the calibrated risk probability value and the updated individualized baseline calibration factor.

[0106] It should be noted that the rolling update condition is set based solely on the persistence of deviation. The rolling update condition is met when the deviation is written as a persistent deviation, and not met when the deviation is written as a non-persistent deviation. When the deviation is written as a persistent deviation, the deviation direction is consistent within adjacent time windows and the deviation magnitude remains stable within adjacent time windows, indicating that the source of the deviation is more consistent with the baseline drift of the target individual's past state rather than short-term fluctuations. Therefore, the individualized baseline calibration factor value is triggered for graded migration update to absorb long-term offset. When the deviation is written as a non-persistent deviation, the continuity of the deviation direction is insufficient or the deviation magnitude fluctuates significantly, indicating that the source of the deviation is more likely to be a short-term anomaly or sampling perturbation. Therefore, the individualized baseline calibration factor value is not triggered to avoid solidifying short-term fluctuations into baseline drift.

[0107] S5.4: Perform validity verification on the calibrated risk probability value under three types of constraints: consistency of probability range, consistency of calibration magnitude, and continuity of time, and obtain the final effective risk probability value;

[0108] Furthermore, the calibrated risk probability value is read and a probability range consistency check is performed. The probability range consistency check performs boundary checks on the calibrated risk probability value based on the allowable range of the risk probability value and outputs the check result. The original risk probability value is read and a calibration amplitude consistency check is performed. The calibration amplitude consistency check performs amplitude checks on the difference between the calibrated risk probability value and the original risk probability value based on the calibration amplitude corresponding to the deviation amplitude level in the individualized baseline calibration factor value and outputs the check result. Records with record type identifiers written as patient status records are selected from the health event sequence carrier, and records with a unified timestamp field not greater than the right boundary timestamp of the time window and the largest unified timestamp field are selected. The selected records are read to obtain the previous final effective risk probability value. A time continuity check is performed based on the previous final effective risk probability value and the calibrated risk probability value and outputs the check result. When the probability range consistency check, calibration amplitude consistency check, and time continuity check all pass, the calibrated risk probability value is written into the final effective risk probability value. If any check fails, the original risk probability value is written into the final effective risk probability value to obtain the final effective risk probability value.

[0109] It should be noted that the direction of deviation is synthesized using the direction aggregation rule of the difference direction within the drift evidence to avoid the deviation direction being dragged by the abnormal value of a single field. The mapping of the deviation magnitude level uses the grading rule of the difference magnitude within the drift evidence to support the grading calibration magnitude retrieval of the individualized baseline calibration factor value. The determination of deviation persistence uses the continuous occurrence of drift evidence in adjacent time windows to support the trigger control of the rolling update of the individualized baseline calibration factor.

[0110] S6: Bind the final effective risk probability value with the corresponding time window identifier, generate a patient status record, and write it back to the health event sequence carrier for closed-loop optimization;

[0111] S6.1: Read the left boundary time marker and right boundary time marker of the time window corresponding to the final effective risk probability value from the time window sample package, and generate the time window identifier using a fixed splicing rule;

[0112] Furthermore, the time window sample package is read and the left and right boundary timestamps of the time window corresponding to the final effective risk probability value are located. The splicing rule is fixed and the splicing is performed in the order of the left boundary timestamp first and the right boundary timestamp last. A separator field is written between the left and right boundary timestamps of the time window. The splicing result is written into the time window identifier to obtain the time window identifier.

[0113] S6.2: Write the unified time stamp, source identifier, and record type identifier into the time window identifier and the final effective risk probability value to generate a patient status record;

[0114] Furthermore, a patient status record is created and written into the time window identifier field and the final effective risk probability value field. The unified time stamp field is written into the right boundary time stamp of the time window boundary, the source identifier field is written into the risk prediction model based on the time interval decay gating cycle, and the record type identifier is written into the patient status record to obtain the patient status record.

[0115] S6.3: Patient status records are appended to the health event sequence carrier according to the unified time stamp field to form a closed loop;

[0116] Furthermore, the patient status record is read and the append position is located in the health event sequence carrier according to the unified timestamp field. The append position satisfies the ascending order arrangement rule of the unified timestamp field. The patient status record is written into the health event sequence carrier to complete the append write, and the written-back health event sequence carrier is obtained.

[0117] It should be noted that the time window identifier transforms the time markers of the left and right boundaries of the time window into searchable identifiers through a fixed splicing rule, enabling the patient status record to establish a correspondence with the time window sample package in a unified time marker dimension. After the patient status record is appended to the health event sequence carrier, the health event sequence carrier can be used for the establishment of subsequent event triggering conditions and the extraction of individualized baseline calibration factor values, thereby forming a closed-loop optimization.

[0118] To facilitate consistency verification of the risk probability output after write-back, the original risk probability value, the calibrated risk probability value, and the final effective risk probability value are recorded synchronously in the patient status record in ascending order of the time markers at the right boundary of the time window. A rollback flag is also recorded synchronously when the validity verification fails and a rollback is triggered. Figure 4 The above data is presented as a trajectory over a continuous time window, with the upper image showing the global trend and the lower image showing local differences within a magnified window.

[0119] This embodiment also provides an AI-based dynamic prediction system for cerebrovascular disease risk, comprising: a time-stamping module for connecting multi-source medical and health data of a target individual to the same acquisition channel and writing a unified time stamp, sorting the multi-source medical and health data according to the time stamp, and writing it into a health event sequence carrier; a triggering and labeling module for establishing event triggering conditions based on the health event sequence carrier, and writing newly arriving multi-source medical and health data into the labeled event stream after triggering and labeling; and a time window sampling module for using the corresponding time point as the time window boundary when the event triggering conditions are met, based on the labeled event stream, and dynamically adjusting the time window span according to irregular sampling density. The system generates a dynamic set of time windows, which is mapped to a time window sample package. A risk drift module inputs the time window sample package into a risk prediction model based on time interval decay gating loops to obtain the original risk probability value and simultaneously acquires drift evidence deviating from the individual's past state. A calibration and verification module performs online recalibration on the original risk probability value based on the drift evidence, updates the individualized baseline calibration factor to form a calibrated risk probability value, and performs validity verification to obtain the final effective risk probability value. A write-back closed-loop module binds the final effective risk probability value to the corresponding time window identifier, generates a patient status record, and writes it back to the health event sequence carrier for closed-loop optimization.

[0120] In summary, this invention achieves stable and dynamic prediction of cerebrovascular disease risk by: constructing a dynamic time window driven by event triggering, which adaptively adjusts the time window span according to irregular sampling density and forms a time window sample package; outputting the original risk probability using a time interval decay-gated cyclic model; and driving online recalibration and validity verification by drift evidence, combined with a write-back closed loop.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based method for dynamic prediction of cerebrovascular disease risk, characterized in that: include, Multi-source medical and health data of the target individual are connected to the same acquisition channel and written with a unified time stamp. The multi-source medical and health data are sorted according to the time stamp and written into a health event sequence carrier. Based on the health event sequence carrier, event triggering conditions are established, and newly arriving multi-source medical and health data are marked with triggers and then written into the marked event stream; Based on the labeled event stream, when the event triggering conditions are met, the corresponding time point is used as the boundary of the time window, and the time window span is dynamically adjusted according to the irregular sampling density to generate a dynamic time window set, which is mapped to the time window sample bag; Input the time window sample package into the risk prediction model based on the time interval decay gating cycle to obtain the original risk probability value, and simultaneously obtain drift evidence that deviates from the individual's past state; Based on drift evidence, the original risk probability value is recalibrated online, and the individualized baseline calibration factor is updated to form a calibrated risk probability value. At the same time, an effectiveness verification is performed to obtain the final effective risk probability value. The final effective risk probability value is bound to the corresponding time window identifier to generate a patient status record, which is then written back to the health event sequence carrier for closed-loop optimization.

2. The AI-based dynamic prediction method for cerebrovascular disease risk as described in claim 1, characterized in that: The specific steps for writing the unified timestamp are as follows. The multi-source medical and health data of the target individual are written one by one into the receiving buffer of the same acquisition channel, deduplication verification is performed, and multi-source medical and health data that has passed the deduplication verification is generated. Read the record time field of the multi-source medical and health data that has passed the deduplication verification, convert it into millisecond-level timestamps and write it into a unified timestamp field to generate multi-source medical and health data with unified timestamps.

3. The AI-based dynamic prediction method for cerebrovascular disease risk as described in claim 2, characterized in that: The specific steps for writing the health event sequence carrier are as follows: Sort the multi-source medical and health data with the same time stamp in ascending order according to the time stamp, and generate sorted multi-source medical and health data; The sorted multi-source medical and health data are appended to the health event sequence carrier one by one according to the sorting order.

4. The AI-based dynamic prediction method for cerebrovascular disease risk as described in claim 3, characterized in that: The specific steps for triggering and labeling newly arriving multi-source medical and health data and then writing it into the labeled event stream are as follows. Extract the source identifier sequence and the unified time-stamp sequence of multi-source medical and health data from the health event sequence carrier according to the unified time stamp, and generate the arrival interval sequence by grouping according to the source identifier; The arrival interval sequence is sorted in ascending order to obtain a sorted arrival interval sequence, and event triggering conditions are established by combining the boundary interval threshold. Receive newly arrived multi-source medical and health data, read the source identifier field and the unified timestamp field, combine them with the event triggering conditions to perform trigger annotation, and write the annotated event stream.

5. The AI-based dynamic prediction method for cerebrovascular disease risk as described in claim 4, characterized in that: The specific steps for using the corresponding time point as the boundary of the time window when the event triggering condition is met are as follows: The trigger annotations are written into the multi-source medical and health data that has not yet been triggered, according to the labeled event flow, and recorded as the right boundary record of the time window; The unified time marker recorded at the right boundary of the time window is used as the right boundary time marker of the time window, and a backtracking sampling method based on the source identifier is performed in the labeled event stream to form the time window boundary.

6. The AI-based dynamic prediction method for cerebrovascular disease risk as described in claim 5, characterized in that: The mapping is a time window sample package, and the specific steps are as follows. By selecting multi-source medical and health data from the labeled event stream according to a unified time mark, which fall within the range of the time mark from the left boundary of the time window to the right boundary of the time window, a dynamic time window set is obtained. Based on the dynamic time window set, the multi-source medical and health data corresponding to the time window boundary are grouped according to the record type, and the grouping results, along with the time markers of the left and right boundaries of the time window, are written into the time window sample package.

7. The AI-based dynamic prediction method for cerebrovascular disease risk as described in claim 6, characterized in that: The process of obtaining the original risk probability value and simultaneously acquiring drift evidence deviating from the individual's past state involves the following specific steps: Based on time window sample packages, the multi-source medical and health data in the grouping results are organized sequentially according to a unified time stamp to obtain a risk prediction sequence. A risk prediction model based on time interval decay gated recurrent units (GRU), residual convolutional neural networks, and fully connected neural networks is constructed. The historical risk prediction sequence is bound to the cerebrovascular disease outcome events that occur in close temporal order in the health event sequence carrier to form a training sample set, and then a validation sample set is obtained. The training sample set is input into the risk prediction model to perform supervised training, and the effectiveness of the training process is determined by the validation sample set to obtain the trained risk prediction model. Input the risk prediction sequence into the trained risk prediction model, perform irregular sampling time series probability inference, and output the original risk probability value; Field values ​​are extracted from the grouping results of the time window sample package and the previous state time window sample package, and item-by-item comparison processing is performed to generate drift evidence.

8. The AI-based dynamic prediction method for cerebrovascular disease risk as described in claim 7, characterized in that: The specific steps for obtaining the final effective risk probability value are as follows. The deviation direction, deviation magnitude level, and deviation persistence are extracted from drift evidence, and individualized baseline calibration factor values ​​are extracted from health event sequence vectors. The deviation direction, deviation magnitude level, deviation persistence, and individualized baseline calibration factor values ​​are bound to the online recalibration context; Based on the online recalibration context, probability mapping calibration is performed on the original risk probability value to obtain the calibrated risk probability value, and the individualized baseline calibration factor is rolled over with the deviation persistence as the update condition to obtain the updated individualized baseline calibration factor. The validity of the calibrated risk probability value is verified under three types of constraints: consistency of probability range, consistency of calibration magnitude, and continuity of time, and the final effective risk probability value is obtained.

9. The AI-based dynamic prediction method for cerebrovascular disease risk as described in claim 8, characterized in that: The specific steps for generating the patient status record are as follows. Read the left and right boundary time markers of the time window corresponding to the final effective risk probability value from the time window sample package, and generate the time window identifier using a fixed splicing rule; A patient status record is generated by writing a unified time stamp, source identifier, and record type identifier into the time window identifier and the final effective risk probability value.

10. An AI-based dynamic prediction system for cerebrovascular disease risk, based on the AI-based dynamic prediction method for cerebrovascular disease risk according to any one of claims 1 to 9, characterized in that: include, The data acquisition time stamp module is used to connect the multi-source medical and health data of the target individual to the same acquisition channel, write a unified time stamp, sort the multi-source medical and health data according to the time stamp, and write it into the health event sequence carrier. The trigger annotation module is used to establish event triggering conditions based on the health event sequence carrier, and to write the newly arrived multi-source medical and health data into the labeled event stream after triggering annotation. The time window sample module is used to generate a dynamic time window set based on the labeled event stream, using the corresponding time point as the boundary of the time window when the event triggering condition is met, and dynamically adjusting the time window span according to the irregular sampling density, which is then mapped to the time window sample package. The risk drift module is used to input the time window sample package into the risk prediction model based on the time interval decay gating cycle to obtain the original risk probability value, and simultaneously obtain drift evidence that deviates from the individual's past state. The calibration and verification module is used to perform online recalibration on the original risk probability value based on drift evidence, update the individualized baseline calibration factor to form a calibrated risk probability value, and perform validity verification to obtain the final effective risk probability value. The write-back closed-loop module is used to bind the final effective risk probability value with the corresponding time window identifier, generate patient status records, and write them back to the health event sequence carrier for closed-loop optimization.

Images