CN-121709264-B - Atrial fibrillation recurrence prediction method based on artificial intelligence

CN121709264BCN 121709264 BCN121709264 BCN 121709264BCN-121709264-B

Abstract

The invention discloses an artificial intelligence-based atrial fibrillation recurrence prediction method, which belongs to the technical field of medical information, and specifically comprises the steps of inputting a discrete clinical event record of a target patient after ablation operation, and inverting a continuous internal state evolution path from the discrete clinical event record through an event-driven hidden state deduction model. The path is divided into a plurality of restoration stages and feature vectors are generated. And taking the last stage as a query object, searching similar historical stages in the pre-constructed group recovery process map, and extracting a complete stage chain from the last stage to a clear ending. And mapping back to a physiological state space, and forming a plurality of candidate future evolution chains starting from the current state of the patient through coordinate translation. Finally, combining the patient history event mode, calculating likelihood scores of the candidate chains, and outputting a personalized recurrence risk prediction path set after sorting and screening. The invention provides a new dynamic prediction approach for atrial fibrillation recurrence risk with individual adaptability and time sequence interpretation.

Inventors

LIAO XUEWEN
HUANG XIAOQI
HUANG HUASHAN
LIAN LIANGHUA
LIN YAZHOU

Assignees

福州大学附属省立医院

Dates

Publication Date: 20260512
Application Date: 20260206

Claims (7)

1. An artificial intelligence-based atrial fibrillation recurrence prediction method is characterized by comprising the following steps: S1, inputting a discrete clinical event sequence recorded after self-ablation of a target patient, wherein the discrete clinical event sequence comprises symptom records, medication change records and heart rhythm conclusion records, and each record is attached with an event time stamp and text description; s2, inputting the discrete clinical event sequence into an event-driven hidden state deduction model, and outputting an internal physiological state time sequence path which is evolved from an initial state to a current state after operation in a preset physiological state space by the model; s3, identifying a plurality of recovery stages with different evolution modes according to the distribution density and the transfer direction of the state points in the internal physiological state time sequence path, and generating a corresponding stage feature vector for each recovery stage; s4, taking the last recovery stage as a query stage, and searching out historical recovery stages with feature vector similarity exceeding a set threshold value in all stages in a pre-constructed group recovery process map; S5, traversing each history recovery stage backwards along a time sequence until reaching a node with a atrial fibrillation recurrence mark or a long-term stable mark, obtaining a complete subsequent stage chain, and mapping each subsequent stage chain back to the physiological state space to form a plurality of candidate future evolution chains starting from the current state; s6, calculating likelihood scores of each candidate future evolution chain according to event type distribution and time interval characteristics of the discrete clinical event sequence, and sequencing and screening all candidate future evolution chains according to the scores to output a personalized recurrence risk prediction path set; in the step S2, the operation process of the event-driven hidden state deduction model is as follows: Defining the physiological state space as a multidimensional continuous vector space, wherein each dimension corresponds to a potential physiological factor, distributing a preset initial state vector for the initial moment after ablation operation, iteratively processing each record in the discrete clinical event sequence according to the time stamp sequence, calculating the prediction probability distribution of various clinical events based on the current estimated state vector when processing a single record, and calculating the difference between the prediction probability distribution and the actual event type; the text description of the current event record is encoded to obtain an event intensity vector, the state vector estimated values at the current moment and the previous moment are reversely adjusted according to the difference and the event intensity vector, and the internal physiological state time sequence path is output after the processing of all the event records is completed; in the step S3, the process of identifying the recovery stage and generating the stage feature vector is as follows: in the physiological state space, traversing the internal physiological state time sequence path by adopting a sliding window with fixed length, calculating the mean vector of all state points in each sliding window as a window center point, calculating the variance of the state points in each sliding window in each dimension, and recording a stage boundary point at a window starting position when the Euclidean distance between the center points of two continuous sliding windows exceeds a stage transition distance threshold value; The time sequence path is divided into a plurality of continuous paragraphs according to boundary points of all phases, each paragraph is a recovery phase, and for each recovery phase, a phase characteristic vector is formed by splicing a mean value and a standard deviation of a state point in the phase in each dimension, a duration of the phase, a difference vector of a start state vector and a finish state vector of the phase and a single-heat coding summation vector of a clinical event type in the phase.
2. The method of claim 1, wherein the step of adjusting the state vector estimate in the opposite direction to the event intensity vector based on the difference is: constructing a forward state transition network, wherein the network takes a state vector as an input and outputs transition probability of the state vector, and constructing a reverse inference network, wherein the network takes clinical event observation data as an input and outputs inference probability of the state vector; Inputting the current state vector before adjustment into the forward state transition network to obtain prior state estimation probability distribution, inputting the actual type of the current event record and the event intensity vector into the reverse inference network to obtain posterior state estimation probability distribution, calculating KL divergence between the prior state estimation probability distribution and the posterior state estimation probability distribution, calculating gradients of parameters of the forward state transition network and the reverse inference network according to the KL divergence, and updating the parameters of the forward state transition network and the reverse inference network along the gradient direction.
3. The method for predicting atrial fibrillation recurrence based on artificial intelligence as claimed in claim 1, wherein in S4, the process of searching the historical recovery phase in the population recovery process map is as follows: The nodes in the group recovery process map store phase feature vectors of a history recovery phase, directed edges are connected with adjacent phase nodes of the same history patient, the edge weights are the similarity of the two node phase feature vectors, the cosine similarity of the phase feature vectors of the query phase and the phase feature vectors of all nodes in the map is calculated, and the nodes with the cosine similarity exceeding a retrieval threshold are selected to form a preliminary retrieval result; And for each node in the preliminary search result, calculating the matching degree between the characteristic of the previous recovery stage of the query stage on the original path and the characteristic of the previous node of the node in the map, fusing the cosine similarity and the matching degree to carry out weighted scoring, and screening a historical recovery stage set according to the scoring.
4. The method for predicting atrial fibrillation recurrence based on artificial intelligence as claimed in claim 1, wherein the step of obtaining a complete chain of subsequent phases in S5 is as follows: And locating the node corresponding to the retrieved history recovery stage in the group recovery process map as an initial node, starting from the initial node, traversing backwards along the directed edge which starts from the initial node and points to the node with the time sequence behind, checking whether the attribute of the initial node contains a atrial fibrillation mark or a long-term stability mark when accessing the first subsequent node, if not, recording the current subsequent node as a current chain node and continuing to access the next subsequent node along the directed edge, repeatedly executing the node attribute checking and recording operation until the access to a termination node with the attribute containing the atrial fibrillation mark or the long-term stability mark, and connecting the initial node, all recorded current chain nodes and the termination node according to the access time sequence to form a complete subsequent stage chain.
5. The method for predicting atrial fibrillation recurrence based on artificial intelligence as defined in claim 4, wherein in S5, the process of forming a plurality of candidate future evolution chains from the current state is as follows: Extracting a state point mean value vector component and a stage duration component of a stored stage feature vector from each node in the subsequent stage chain, generating a track subsection of continuous evolution of a characterization state by taking the state point mean value vector as a geometric center and a time interval corresponding to the stage duration as a definition domain in the physiological state space, butting adjacent track subsections at time endpoints according to the time sequence of the nodes in the chain, and performing spatial interpolation smoothing processing on track morphology at the butting position according to directed edge weight values of connecting corresponding nodes in the map to form a complete and continuous initial track; calculating vector difference between the coordinates of the initial track starting point in a physiological state space and the current state coordinates of the target patient, subtracting the vector difference from all the point coordinates on the initial track to finish translation alignment of the track space position, and marking the track after translation alignment as a candidate future evolution chain starting from the current state.
6. The artificial intelligence-based atrial fibrillation recurrence prediction method as claimed in claim 1, wherein in S6, the process of calculating the likelihood score of each candidate future evolution chain is as follows: The method comprises the steps of mapping a predicted clinical event type sequence according to a state evolution sequence of a candidate future evolution chain, counting event type distribution, adjacent event average time intervals and specific event combination occurrence frequency of the sequence, acquiring historical event type distribution, historical average time intervals and historical combination frequency corresponding to a target patient discrete clinical event sequence, calculating a first coincidence degree value between the predicted event type distribution and the historical event type distribution, calculating the reciprocal of a difference value between the predicted event type distribution and the historical average time interval as a second coincidence degree value, and calculating the average value of reciprocal of difference values between each specific event combination prediction and the historical frequency as a third coincidence degree value; And calculating a track curvature value and a path total length value of the evolution chain in a physiological state space, multiplying the first, second and third coincidence degree values, the reciprocal of the curvature value and the reciprocal of the total length value by a preset weight coefficient respectively, and then summing to obtain a likelihood score.
7. The method for predicting atrial fibrillation recurrence based on artificial intelligence as defined in claim 6, wherein in S6, the process of sorting and screening and outputting the path set according to the scores is as follows: Setting a lowest score threshold, comparing the likelihood score of each candidate future evolution chain with the threshold, removing all evolution chains with scores lower than the threshold, checking the end attribute marks of the rest evolution chains, classifying the evolution chains marked as atrial fibrillation recurrence into a first subset, classifying the evolution chains marked as long-term stable into a second subset, respectively arranging the evolution chains in the first subset and the second subset in descending order according to the likelihood scores, selecting the first M evolution chains from the first subset after arrangement, selecting the first K evolution chains from the second subset after arrangement, wherein M and K are all preset positive integers, and combining all the selected evolution chains into a final output personalized recurrence risk prediction path set.

Description

Atrial fibrillation recurrence prediction method based on artificial intelligence Technical Field The invention relates to the technical field of medical information, in particular to an atrial fibrillation recurrence prediction method based on artificial intelligence. Background Accurate prediction of risk of recurrence after atrial fibrillation (atrial fibrillation) catheter ablation is one of the key challenges in the field of clinical rhythm management. Postoperative recurrence not only affects the quality of life of the patient, but also means the reinjection of medical resources. Therefore, developing an effective prediction method has important clinical significance for realizing personalized postoperative management and optimizing follow-up strategies. Currently, research and practice in this field is largely focused around the continuous monitoring and assessment of the physiological state of postoperative patients. In the prior art, a prediction method based on continuous physiological signals (such as long-range electrocardiographic monitoring and wearable equipment data) is a main stream direction. Such methods assess risk of recurrence by capturing dynamic changes in electrophysiological indices such as heart rate variability, atrial premature beat burden, etc. In addition, some studies attempt to build statistical models or machine learning models for risk stratification in combination with clinical baseline characteristics of the patient (e.g., age, medical history, left atrial size) and post-operative early cardiac rhythm monitoring results. Another concept focuses on assessing atrial substrate stability using cardiac imaging data (e.g., the degree of fibrosis shown by cardiac magnetic resonance delay enhancement) and using it as a static basis for prediction. These technical paths all provide information for recurrence risk assessment at a specific level. However, the prior art solutions have structural limitations at the methodology level, limiting their predictive efficacy and clinical utility. Firstly, the technical path relying on high-density continuous physiological signal monitoring is limited by a data acquisition threshold, the feasibility of large-scale and long-period implementation is insufficient, and the captured electrophysiological index has a gap from key dimensions such as subjective feeling of a patient, clinical intervention decision and the like. Secondly, for the discrete clinical event sequence naturally generated in the conventional medical process, the existing prediction model usually simplifies the discrete clinical event sequence into static classification variables or carries out simple statistical aggregation (such as counting and single-heat coding), and the processing mode completely strips the time association and sequential logic between the events, so that the dynamic process of state evolution cannot be described. More importantly, the existing method lacks the capability of reversely constructing continuous internal physiological state evolution from sparse and asynchronous discrete events, and also fails to place the current recovery fragments of the individuals in a network of a group evolution path for similarity retrieval and deduction. This results in the prediction logic of the existing model often staying at the surface feature correlation level, which makes it difficult to reveal the uniqueness of the individual recovery trajectory, and also fails to provide a risk evolution path with time sequence interpretation, so that there is a significant bottleneck in the accuracy of personalized prediction and the clinical decision support value. Disclosure of Invention The invention aims to provide an artificial intelligence-based atrial fibrillation recurrence prediction method, which solves the following technical problems: The existing atrial fibrillation recurrence prediction method mainly depends on high-cost continuous monitoring data, has limited feasibility, only performs static statistics on conventional discrete clinical events, and loses key time sequence associated information. And the continuous evolution track of the internal physiological state of the patient cannot be reversely constructed from the discrete event, and the individual recovery stage cannot be placed in the group evolution network to carry out similarity deduction, so that the prediction result lacks individual time sequence interpretation and the clinical decision support value is limited. The aim of the invention can be achieved by the following technical scheme: an artificial intelligence-based atrial fibrillation recurrence prediction method comprises the following steps: S1, inputting a discrete clinical event sequence recorded after self-ablation of a target patient, wherein the discrete clinical event sequence comprises symptom records, medication change records and heart rhythm conclusion records, and each record is attached with an event time stamp and text descript