An artificial intelligence-based atrial fibrillation recurrence prediction method

Abstract

The application discloses an atrial fibrillation recurrence prediction method based on artificial intelligence and belongs to the medical information technology field, and specifically comprises the following steps: inputting discrete clinical event records of a target patient after ablation, and inversely deriving a continuous internal state evolution path from the event-driven hidden state deduction model; dividing the path into multiple recovery stages and generating a feature vector; taking the last stage as a query object, searching for similar historical stages in a pre-constructed group recovery process graph, and extracting a complete stage chain until a clear outcome; and mapping back to a physiological state space, forming multiple candidate future evolution chains starting from the current state of the patient through coordinate translation; finally, combining the historical event mode of the patient, calculating the likelihood score of each candidate chain, and outputting a personalized recurrence risk prediction path set after sorting and screening. The application provides a new way for dynamic prediction of atrial fibrillation recurrence risk, which has individual adaptability and time sequence explanation.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to an artificial intelligence-based method for predicting atrial fibrillation recurrence. Background Technology

[0002] Accurate prediction of recurrence risk after catheter ablation for atrial fibrillation (AF) is one of the key challenges in clinical cardiac rhythm management. Postoperative recurrence not only affects patients' quality of life but also means a reinvestment of medical resources. Therefore, developing effective prediction methods is of great clinical significance for achieving individualized postoperative management and optimizing follow-up strategies. Currently, research and practice in this field mainly revolve around the continuous monitoring and assessment of patients' physiological status after the procedure.

[0003] In existing technologies, prediction methods based on continuous physiological signals (such as long-term electrocardiogram monitoring and wearable device data) are the mainstream approach. These methods assess recurrence risk by capturing dynamic changes in electrophysiological indicators such as heart rate variability and atrial premature beat load. In addition, some studies attempt to combine patients' clinical baseline characteristics (such as age, medical history, and left atrial size) with early postoperative cardiac rhythm monitoring results to construct statistical or machine learning models for risk stratification. Another approach focuses on using cardiac imaging data (such as the degree of fibrosis shown by delayed enhancement on cardiac magnetic resonance imaging) to assess atrial matrix stability and use it as a static basis for prediction. These technological approaches all provide information for recurrence risk assessment at specific levels.

[0004] However, existing technological solutions suffer from structural limitations at the methodological level, restricting their predictive efficacy and clinical applicability. First, the technical approach relying on high-density continuous physiological signal monitoring is limited by data acquisition barriers, making large-scale, long-term implementation impractical. Furthermore, the captured electrophysiological indicators are disconnected from key dimensions such as patient subjective feelings and clinical intervention decisions. Second, for naturally occurring discrete clinical event sequences in routine medical procedures, existing predictive models typically simplify them into static categorical variables or perform simple statistical aggregations (such as counting or one-hot encoding). Such processing completely strips away the temporal correlation and sequential logic between events, failing to characterize the dynamic process of state evolution. More importantly, existing methods lack the ability to reverse-engineer continuous intrinsic physiological state evolution from sparse, asynchronous discrete events, and also fail to place an individual's current recovery fragment within the network of population evolutionary paths for similarity retrieval and deduction. This results in the predictive logic of existing models often remaining at the level of superficial feature correlations, failing to reveal the uniqueness of individual recovery trajectories or provide temporally interpretable risk evolution paths. Therefore, they exhibit significant bottlenecks in the accuracy of personalized predictions and their value in supporting clinical decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based method for predicting atrial fibrillation recurrence, addressing the following technical problems:

[0006] Existing methods for predicting atrial fibrillation recurrence primarily rely on high-cost continuous monitoring data, which limits their feasibility. They only perform static statistics on routine discrete clinical events, losing crucial temporal correlation information. Furthermore, they cannot reverse-engineer the continuous evolutionary trajectory of the patient's intrinsic physiological state from discrete events, nor can they place the individual recovery stage within a population evolutionary network for similarity extrapolation. This results in predictions lacking individualized temporal interpretability and limited value in supporting clinical decision-making.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] An artificial intelligence-based method for predicting atrial fibrillation recurrence includes the following steps:

[0009] S1. Input the discrete clinical event sequence recorded after the target patient's self-ablation, including symptom records, medication change records, and heart rhythm conclusion records. Each record is accompanied by an event timestamp and text description.

[0010] S2. Input the discrete clinical event sequence into the event-driven hidden state deduction model. The model outputs an intrinsic physiological state time sequence path that evolves from the initial postoperative state to the current state in the preset physiological state space.

[0011] S3. Based on the distribution density and transfer direction of state points in the intrinsic physiological state time-series path, identify multiple recovery stages with different evolution patterns, and generate a corresponding stage feature vector for each recovery stage.

[0012] S4. Using the last recovery stage as the query stage, retrieve all historical recovery stages whose feature vector similarity exceeds a set threshold from the pre-constructed population recovery process map.

[0013] S5. For each historical recovery stage, traverse backwards in chronological order until a node with atrial fibrillation recurrence marker or long-term stability marker is reached to obtain a complete chain of subsequent stages. Map each chain of subsequent stages back to the physiological state space to form multiple candidate future evolution chains starting from the current state.

[0014] S6. Based on the event type distribution and time interval characteristics of the discrete clinical event sequence, calculate the likelihood score of each candidate future evolution chain, sort and filter all candidate future evolution chains according to the score, and output a personalized relapse risk prediction path set.

[0015] As a further aspect of the present invention: in S2, the operation process of the event-driven hidden state deduction model is as follows:

[0016] The physiological state space is defined as a multidimensional continuous vector space, where each dimension corresponds to a potential physiological factor. A preset initial state vector is assigned to the initial moment after ablation. Each record in the discrete clinical event sequence is iteratively processed according to the timestamp order. When processing a single record, the predicted probability distribution of various clinical events is calculated based on the currently estimated state vector. The difference between the predicted probability distribution and the actual event type is calculated.

[0017] The text description of the current event record is encoded to obtain the event intensity vector. Based on the difference and the event intensity vector, the estimated values ​​of the state vector at the current time and the previous time are adjusted in reverse. After processing all event records, the intrinsic physiological state time sequence path is output.

[0018] As a further aspect of the present invention: the process of adjusting the state vector estimate in reverse based on the difference and event intensity vector is as follows:

[0019] Construct a forward state transition network that takes a state vector as input and outputs the transition probability of the state vector. Construct a backward inference network that takes clinical event observation data as input and outputs the inference probability of the state vector.

[0020] The current state vector before adjustment is input into the forward state transition network to obtain the prior state estimation probability distribution. The actual type and event intensity vector of the current event record are input into the backward inference network to obtain the posterior state estimation probability distribution. The KL divergence between the prior state estimation probability distribution and the posterior state estimation probability distribution is calculated. The gradient of the parameters of the forward state transition network and the backward inference network is calculated based on the KL divergence. The parameters of the forward state transition network and the backward inference network are updated along the gradient direction.

[0021] As a further aspect of the present invention: in step S3, the process of identifying the recovery phase and generating a phase feature vector is as follows:

[0022] In the physiological state space, a fixed-length sliding window is used to traverse the intrinsic physiological state time sequence path. The mean vector of all state points within each sliding window is calculated as the window center point. The variance of the state points within each sliding window in each dimension is calculated. When the Euclidean distance between the center points of two consecutive sliding windows exceeds the stage transition distance threshold, the stage boundary point is recorded at the beginning position of the window. When the variance change rate of two consecutive sliding windows exceeds the variance stability threshold, the stage boundary point is recorded at the beginning position of the window.

[0023] The temporal path is divided into multiple continuous segments based on all stage boundary points. Each segment is a recovery stage. For each recovery stage, its stage feature vector is composed of the mean and standard deviation of the state points in each dimension within the stage, the duration of the stage, the difference vector between the start and end state vectors of the stage, and the sum of the one-hot encoded vectors of the clinical event types within the stage.

[0024] As a further aspect of the present invention: in step S4, the process of retrieving historical recovery stages from the population recovery process map is as follows:

[0025] The nodes in the group recovery process graph store the stage feature vectors of the historical recovery stages. Directed edges connect adjacent stage nodes of the same historical patient, and the edge weight is the similarity of the stage feature vectors of the two nodes. The cosine similarity between the stage feature vector of the query stage and the stage feature vectors of each node in the graph is calculated. Nodes with a cosine similarity exceeding the retrieval threshold are selected to form the preliminary retrieval results.

[0026] For each node in the preliminary search results, the matching degree between the features of the previous recovery stage on the original path of the query stage and the features of the previous node in the graph is calculated. The cosine similarity and the matching degree are fused to perform a weighted score, and the set of historical recovery stages is selected based on the score.

[0027] As a further aspect of the present invention: in step S5, the process of obtaining a complete chain of subsequent stages is as follows:

[0028] In the population recovery process graph, the node corresponding to the retrieved historical recovery stage is located as the starting node. Starting from the starting node, the graph is traversed backward along the directed edges that start from the starting node and point to nodes later in the time sequence. When visiting the first subsequent node, the attributes of the starting node are checked to see if they contain an atrial fibrillation recurrence marker or a long-term stable marker. If not, the current subsequent node is recorded as the current chain node and the next subsequent node is visited along the directed edges. The node attribute checking and recording operation is repeated until a terminal node containing an atrial fibrillation recurrence marker or a long-term stable marker is visited. The starting node, all recorded current chain nodes, and the terminal node are connected in the order of visit time to form a complete subsequent stage chain.

[0029] As a further aspect of the present invention: in step S5, the process of forming multiple candidate future evolution chains starting from the current state is as follows:

[0030] For each node in the subsequent stage chain, the mean vector component of the state point and the duration component of the stage are extracted from its stored stage feature vector. In the physiological state space, a trajectory segment representing the continuous evolution of the state is generated with the mean vector of the state point as the geometric center and the time interval corresponding to the duration of the stage as the domain. According to the time order of the nodes in the chain, adjacent trajectory segments are connected at the time endpoints. The trajectory shape at the connection point is spatially interpolated and smoothed according to the weight value of the directed edge connecting the corresponding nodes in the graph, forming a complete and continuous initial trajectory.

[0031] Calculate the vector difference between the coordinates of the starting point of the initial trajectory in the physiological state space and the coordinates of the target patient's current state. Subtract the vector difference from the coordinates of all points on the initial trajectory to complete the translation and alignment of the trajectory space position. Mark the translated and aligned trajectory as a candidate future evolution chain starting from the current state.

[0032] As a further aspect of the present invention: in step S6, the process of calculating the likelihood score for each candidate future evolutionary chain is as follows:

[0033] Based on the state evolution sequence of the candidate future evolution chain, a predicted clinical event type sequence is mapped, and the event type distribution, average time interval between adjacent events, and frequency of specific event combinations in the sequence are statistically analyzed. The historical event type distribution, historical average time interval, and historical combination frequency corresponding to the discrete clinical event sequence of the target patient are obtained. The first degree of agreement between the prediction and the historical event type distribution is calculated, the reciprocal of the difference between the prediction and the historical average time interval is calculated as the second degree of agreement, and the average of the reciprocals of the differences between the prediction and the historical frequency of each specific event combination is calculated as the third degree of agreement.

[0034] The trajectory curvature value and total path length value of the evolutionary chain in the physiological state space are calculated. The first, second, and third conformity values, the reciprocal of the curvature value, and the reciprocal of the total length value are multiplied by preset weight coefficients and then summed to obtain the likelihood score.

[0035] As a further aspect of the present invention: In step S6, the process of sorting, filtering, and outputting the path set based on the scores is as follows:

[0036] Set a minimum score threshold, compare the likelihood score of each candidate future evolutionary chain with the threshold, and remove all evolutionary chains with scores below the threshold. Check the endpoint attribute labels of the remaining evolutionary chains, and assign evolutionary chains labeled as atrial fibrillation recurrence to the first subset and those labeled as long-term stable to the second subset. Sort the evolutionary chains in the first and second subsets in descending order of their likelihood scores. Select the top M evolutionary chains from the first subset and the top K evolutionary chains from the second subset, where M and K are preset positive integers. Merge all selected evolutionary chains into the final output personalized recurrence risk prediction path set.

[0037] The beneficial effects of this invention are:

[0038] This invention constructs an event-driven latent state inference model, enabling the reverse reconstruction of continuous intrinsic physiological state time-series paths from sparse, asynchronous, and discrete clinical event records after patient surgery. This transforms easily obtainable symptoms, medications, and examination conclusions from routine follow-ups into quantifiable and analyzable state evolution trajectories. Furthermore, by identifying key recovery stages within these trajectories and performing structural similarity retrieval and matching within a pre-constructed population recovery process atlas, the current recovery segment of an individual can be precisely located within a massive historical patient evolution network. Building upon this, by extracting historical subsequent stage chains similar to the current stage and mapping them back to a unified physiological state space to form multiple candidate future evolution chains, a multi-dimensional and visualized inference of the individual's possible recovery paths is achieved. Finally, by integrating the statistical characteristics of the patient's historical event sequence with the geometric properties of the candidate evolution chains for likelihood calculation and screening, a highly personalized set of recurrence risk prediction paths is output.

[0039] This invention fundamentally overcomes the dependence of existing technologies on expensive continuous monitoring equipment and makes full use of the temporal logic and state correlations that are often overlooked in routine medical data, providing a new approach for dynamic prediction of atrial fibrillation recurrence risk that is feasible, individually adaptable, and temporally interpretable. Attached Figure Description

[0040] The invention will now be further described with reference to the accompanying drawings.

[0041] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Please see Figure 1 As shown, this invention is an artificial intelligence-based method for predicting atrial fibrillation recurrence, comprising the following steps:

[0044] S1. Input the discrete clinical event sequence recorded after the target patient's self-ablation, including symptom records, medication change records, and heart rhythm conclusion records. Each record is accompanied by an event timestamp and text description.

[0045] S2. Input the discrete clinical event sequence into the event-driven hidden state deduction model. The model outputs an intrinsic physiological state time sequence path that evolves from the initial postoperative state to the current state in the preset physiological state space.

[0046] S3. Based on the distribution density and transfer direction of state points in the intrinsic physiological state time-series path, identify multiple recovery stages with different evolution patterns, and generate a corresponding stage feature vector for each recovery stage.

[0047] S4. Using the last recovery stage as the query stage, retrieve all historical recovery stages whose feature vector similarity exceeds a set threshold from the pre-constructed population recovery process map.

[0048] S5. For each historical recovery stage, traverse backwards in chronological order until a node with atrial fibrillation recurrence marker or long-term stability marker is reached to obtain a complete chain of subsequent stages. Map each chain of subsequent stages back to the physiological state space to form multiple candidate future evolution chains starting from the current state.

[0049] S6. Based on the event type distribution and time interval characteristics of the discrete clinical event sequence, calculate the likelihood score of each candidate future evolution chain, sort and filter all candidate future evolution chains according to the score, and output a personalized relapse risk prediction path set.

[0050] In a preferred embodiment of the present invention, the operation process of the event-driven hidden state deduction model in step S2 is as follows:

[0051] The core task of the event-driven latent state inference model is to infer a continuous, smooth physiological state evolution path that can reasonably explain the occurrence of all events based solely on a series of discrete and asynchronous clinical events recorded by the target patient over time after ablation.

[0052] First, the carrier of the model's operation needs to be clearly defined, namely the physiological state space. This space is not directly observable, but rather an artificially constructed mathematical framework used to characterize the patient's internal physiological state. In this embodiment, it is defined as a 50-dimensional continuous real-number vector space. Each of these 50 dimensions corresponds to an abstract, potential physiological factor. These factors are internal variables of the model, and their specific physical meaning is not predetermined, but can be understood as comprehensively representing complex physiological conditions that cannot be directly measured, such as "atrial electrical stability," "autonomic nerve tension," "myocardial tissue inflammation level," and "drug serum concentration effect." The initial state of the model, i.e., the hypothetical internal state immediately after ablation, is assigned a preset initial state vector, such as a vector with all dimensions having a value of 0.1, representing a neutral, undifferentiated initial recovery starting point.

[0053] The basic unit of data processed by the model is a set of discrete clinical event records strictly ordered by timestamps. Each record contains an event type (e.g., "palpitation," "taking amiodarone," "24-hour ECG shows increased premature atrial contractions") and its textual description. The model works by sequentially "digesting" these events and continuously adjusting its estimates of the patient's intrinsic physiological state. The process is an iterative forward prediction and backward correction loop.

[0054] When processing the t-th event record, the model calculates a predicted probability distribution based on the best estimated vector (denoted as ht) of the intrinsic physiological state at time t, using a multi-layer neural network. This distribution covers all possible clinical event types (e.g., a pre-defined dictionary containing 15 common events), outputting the likelihood probability of each event occurring in state ht. For example, the model might calculate that in state ht, the probability of "reporting palpitations" is 0.05, the probability of "recording atrial tachycardia" is 0.02, and the probability of "no special event recorded" is 0.93, etc.

[0055] Next, the model compares this predicted probability distribution with the actual event type in the t-th record. The actual event is represented as a one-hot encoded vector, with 1 only at the position corresponding to the event type and 0 otherwise. The difference between the two is calculated using the cross-entropy loss function; this difference quantifies the degree of "error" in the model's prediction based on the current state estimate.

[0056] However, the descriptive text of clinical events carries important information. For example, while "mild palpitations" and "persistent, severe palpitations" both belong to the "palpitations" category, they imply different levels of severity. To address this, the model integrates a text encoder (e.g., a lightweight, fine-tuned version of a pre-trained BERT model) to process the event description text and generate a fixed-length event strength vector. This vector transforms the qualitative description in the text into a 32-dimensional real-valued vector, quantifying the semantic strength and potential clinical weight of the event. This vector will participate in subsequent state adjustments.

[0057] The calculated prediction discrepancy and event intensity vector together drive the model to make a crucial backpropagation adjustment to the state estimate. The core idea of ​​this adjustment process is that since the current state estimate ht leads to a prediction that doesn't match the actual situation, not only might ht need correction, but the previous state h{t-1} that caused the patient to evolve to ht may also be responsible. Therefore, the adjustment will simultaneously apply to the state vectors at the current time t and the previous time t-1. The magnitude and direction of the adjustment are determined by the gradient of the prediction discrepancy and the event intensity vector, implemented through the backpropagation algorithm. Specifically, the model calculates how the state vectors ht and h{t-1} should be fine-tuned to reduce the current prediction error. After adjustment, ht is updated, and this updated ht will serve as the starting point for the "current estimated state" when processing the next event record.

[0058] To systematically and optimally implement the aforementioned "correcting past states based on future observations" reverse adjustment mechanism, this implementation employs a specific neural network architecture based on variational inference. This architecture comprises two core components: a forward state transition network and a reverse inference network.

[0059] The forward state transition network is responsible for characterizing the intrinsic dynamics of physiological state evolution. It simulates how physiological states naturally evolve over time without external intervention. The network takes the adjusted previous-time state vector h{t-1} (from the previous adjustment) as input and outputs a probability distribution describing the prior estimates of all possible states starting from h{t-1} at the next time step. This enforces the smoothness and continuity of the state path, preventing unreasonable abrupt changes in states between adjacent time steps.

[0060] The backpropagation network is responsible for inferring the most likely state based on the actual observed events. It takes the event data observed at the current time, i.e., the concatenation of the event type one-hot encoding and the event intensity vector, as input and outputs a probability distribution that describes the posterior estimate of the possible states at the current time given the current observations.

[0061] The optimization goal of model training is to make the probability distributions output by the two networks as consistent as possible. A specific metric is KL divergence. KL divergence measures the information difference between the prior distribution (from the feedforward network, based on historical states) and the posterior distribution (from the backpropagation network, based on current observations). The training process uses gradient descent to minimize the sum of the KL divergences at all event time steps. This means that the parameters learned by the model will achieve a dynamic balance between: 1. the future state prior predicted from historical states, and 2. the state posterior inferred from actual observed events. Ultimately, the feedforward network learns state transition patterns that conform to physiological laws, and the backpropagation network learns precise mappings of state information decoded from events.

[0062] After model training, the actual simulation process is as follows: Starting from the initial state vector, events are read in chronological order. For each event, the feedforward network provides a prior "predicted state" based on the previous state, while the backpropagation network provides a posterior "explained state" based on the current event. Then, the model calculates a weighted fusion of these two states (the weights can be dynamically determined by KL divergence or other confidence metrics) to form the final estimate of the state vector at that moment. This estimate respects the continuity of state evolution (from the prior) and faithfully reflects the observed information of the current event (from the posterior). This process is continued event by event, and finally, the fused state vectors at all moments are connected in chronological order to obtain the "intrinsic physiological state temporal path." This path is a continuously changing trajectory in 50-dimensional space, where each fluctuation and turn is driven and explained by the specific clinical event sequence input, thus transforming the discrete event point sequence into a continuous state evolution curve.

[0063] In another preferred embodiment of the present invention, the process of identifying the recovery phase and generating a phase feature vector in step S3 is as follows:

[0064] The purpose of this step is to automatically segment the continuous intrinsic physiological state timeline into multiple continuous segments with a consistent physiological pattern. Each segment is called a recovery stage, and a fixed-dimensional feature vector, i.e., a stage feature vector, is generated for each stage to comprehensively and quantitatively characterize its properties.

[0065] First, the input is the time-series path of the target patient's intrinsic physiological state obtained in step S2. This path consists of a series of state vectors sampled at fixed time intervals. Each state vector is a 50-dimensional real-valued vector representing the patient's potential physiological state at that sampling moment. For example, if the sampling interval is one day, the path is a time series of a 50-dimensional vector for each day.

[0066] To delineate meaningful stages from a continuous path, a sliding window of fixed time length is used to traverse and analyze the path. The length of the sliding window needs to be set based on prior clinical knowledge, for example, 30 days. This means that each window covers 30 consecutive state vectors, corresponding to a 30-day observation period. The sliding step size of the window is set to 7 days, allowing for a 23-day overlap between adjacent windows. This improves the robustness of stage boundary detection and avoids missing boundaries due to the randomness of window division.

[0067] For each sliding window, two core calculations are performed. The first is to calculate the window center point vector. Specifically, this is done by calculating the arithmetic mean of all state vectors contained within the window in each dimension, resulting in a 50-dimensional mean vector. This vector represents the average level of the patient's physiological state during the window period. The second is to calculate the variance of the state points within the window in each dimension, i.e., calculating the variance of all state values ​​in each dimension, resulting in a 50-dimensional variance vector. This vector reflects the volatility or stability of the patient's physiological state across various underlying physiological factors during the window period.

[0068] Next, stage boundaries are detected based on the center point vector and variance vector of two consecutive sliding windows. Two decision criteria are introduced here.

[0069] The first criterion is a stage transition criterion based on center point displacement. The Euclidean distance between the center point vectors of two consecutive windows is calculated. Euclidean distance measures the overall shift in the average state level between two time periods. A stage transition distance threshold is set; for example, this threshold can be obtained by analyzing the distribution of center point distances across all windows in historical data and taking the 85th percentile. When the calculated Euclidean distance exceeds this threshold, a significant physiological state pattern shift is considered to have occurred between the two windows, and therefore the start time point of the current window is recorded as a candidate stage boundary point.

[0070] The second criterion is a stage transition criterion based on variance stability. The variance change rate is calculated for two consecutive windows. Specifically, the absolute difference between the two variance vectors in each dimension is calculated first, then these differences are summed, and finally divided by the sum of the variance vectors in the previous window for normalization. The variance change rate quantifies the severity of state volatility changes. A variance stability threshold is set, for example, the 90th percentile of the historical variance change rate distribution. When the calculated variance change rate exceeds this threshold, it means that the instability of the state has jumped. Even if the average level change is not large, it may mark the beginning of a new stage. Therefore, the starting point of the current window is also recorded as a candidate boundary point.

[0071] After traversing the entire path, a series of candidate boundary points that may be densely distributed in time will be obtained. To obtain a reasonable stage division, boundary point merging and filtering are required. A minimum stage duration is set, for example, 14 days. If the time interval between two candidate boundary points is less than this minimum, they are merged into a single boundary point, usually taking the midpoint as the final boundary. This operation ensures that each stage has a clinically interpretable minimum length.

[0072] Based on the final determined sequence of boundary points, the original intrinsic physiological state time path is divided into multiple consecutive time segments, each of which is defined as a recovery phase.

[0073] For each identified recovery stage, a high-dimensional feature vector needs to be constructed to uniquely represent it. The stage feature vector is composed of five components, as follows.

[0074] The first part is the mean vector of all state points within this stage across all dimensions. The state vectors at all sampling times covered by this stage are calculated, and their averages are taken across all 50 dimensions, resulting in a 50-dimensional vector. This vector summarizes the baseline level of physiological state during this stage.

[0075] The second part is the standard deviation vector of all state points in each dimension within this stage. Calculating the standard deviation of the state vector in each dimension within this stage yields another 50-dimensional vector. This characterizes the fluctuation range and stability of the physiological state during this stage.

[0076] The third part is the duration of this phase, expressed as a scalar. For example, if a phase lasts for 45 days, it is recorded as the value 45. It reflects the time span of the pattern's duration.

[0077] The fourth part is the difference vector between the initial and final state vectors of this stage. Subtracting the initial state vector from the final state vector of the stage yields a 50-dimensional vector. This vector describes the overall trend and magnitude of the patient's physiological changes during this stage.

[0078] The fifth part is the aggregated vector of clinical event types occurring within this phase. First, a dictionary containing 15 common postoperative clinical events is defined, each event corresponding to a unique index. For each clinical event recorded within this phase, a 15-dimensional one-hot encoded vector is generated based on its event type, i.e., 1 at the event type index and 0 at the rest. Then, the one-hot encoded vectors of all events within this phase are summed element-wise to obtain a 15-dimensional integer vector. Each element of this vector represents the frequency of the corresponding event type occurring within this phase, recording external clinical observations accompanying the evolution of the physiological state.

[0079] Finally, the five parts mentioned above are connected in sequence to form a stage feature vector with a total dimension of 166 (50 + 50 + 1 + 50 + 15). This vector provides a comprehensive digital description of the recovery stage from five aspects: average state level, volatility, time length, trend of change, and accompanying events, providing a solid foundation for subsequent similarity retrieval.

[0080] In another preferred embodiment of the present invention, the process of retrieving historical recovery stages in the population recovery process map in step S4 is as follows:

[0081] The goal of this step is to find a series of historical recovery stages that are highly similar in feature structure for the current patient's last recovery stage, i.e., the query stage, in the pre-generated and stored population recovery process map, as a reference for subsequent inference of future paths.

[0082] First, the construction and data structure of the population recovery process graph are described. This graph is a directed weighted graph, constructed based on a large-scale dataset of historical patients. For each historical patient, steps S2 and S3 are executed to obtain a sequence of recovery stages and a 166-dimensional stage feature vector for each stage. In the graph, a node is created for each recovery stage of a historical patient. The attributes stored in the node mainly include: the stage feature vector, the patient's anonymous unique identifier, and the sequential number of that stage in the patient's own recovery process. Edges in the graph are used to represent the temporal evolution relationship between stages. For the same historical patient, a directed edge is established between every two adjacent stage nodes according to the chronological order of their recovery stages, pointing from the previous stage to the next. Each edge is accompanied by a weight value, which is defined as the cosine similarity between the stage feature vectors of the two connected nodes. The cosine similarity is obtained by calculating the cosine of the angle between two 166-dimensional vectors, ranging from -1 to +1; a larger value indicates a greater similarity in the feature patterns of the two stages. Through this construction method, the graph not only gathers the static features of all stages in the population, but also encodes the similarity relationships of dynamic transitions between stages.

[0083] The retrieval process employs a two-stage screening strategy to balance efficiency and accuracy.

[0084] The first stage is preliminary retrieval, based on the direct similarity of feature vectors. Given the current patient's query stage Q, its stage feature vector is denoted as Vq. For each node in the atlas, the cosine similarity between the node's stored feature vector Vn and Vq is calculated. A retrieval threshold is set, for example, 0.85. All nodes whose cosine similarity calculation results exceed this threshold are selected, forming the preliminary retrieval result set. This set contains a group of historical stages that are closest to the query stage in the feature space direction.

[0085] The second stage involves refined screening, introducing contextual matching to improve retrieval accuracy. Considering the continuous evolution of physiological states, the validity of a stage partly depends on its predecessor stage. Therefore, it is necessary to evaluate the contextual consistency between the local path of each node in the preliminary retrieval results and the path of the query stage. For each node N in the preliminary retrieval result set, the following operations are performed: First, find the predecessor node P of node N in the graph, i.e., the node directly pointing to N via a directed edge. If N is the first stage of a patient, then P is empty. Simultaneously, obtain the previous recovery stage Qprev and its feature vector of the current patient's query stage Q. Next, calculate the contextual matching. If both P and Qprev exist, calculate the cosine similarity between the feature vectors of P and Qprev as the matching score. If neither P nor Qprev exists, set the matching score to a default value, such as 0.5. Then, perform a fusion score. The cosine similarity between node N and query stage Q is denoted as the direct similarity Sdir, and the calculated contextual matching score is denoted as Sctx. A weighted summation method is used to calculate the comprehensive score ScoreN for node N, with the formula ScoreN = w1 × Sdir + w2 × Sctx. The weight coefficients w1 and w2 are preset values, for example, set to 0.7 and 0.3 respectively, to assign higher weight to direct similarity. Finally, all nodes in the initial search results are sorted in descending order based on the comprehensive score, and the top K nodes with the highest rankings are selected, where K is a preset value, for example, 10, to form the final retrieved historical recovery stage set.

[0086] In another preferred embodiment of the present invention, the process of obtaining a complete chain of subsequent stages in S5 is as follows:

[0087] First, input the set of historical recovery stages retrieved in step S4. Each element in this set corresponds to a specific node in the graph, which serves as the starting node for subsequent traversals. The population recovery process graph is essentially a large knowledge network storing a vast amount of historical patient recovery stages and their transition relationships. Each node in the graph stores not only a 166-dimensional stage feature vector but also several attribute labels. Among these, two key terminal labels are the "atrial fibrillation recurrence label" and the "long-term stability label." When a historical patient's follow-up outcome is clearly determined to be atrial fibrillation recurrence, the last stage node in their recovery stage sequence associated with the recurrence event time is assigned the "atrial fibrillation recurrence label." Similarly, if a patient has no recurrence during long-term follow-up after surgery (e.g., more than 2 years), their recorded last stage node is assigned the "long-term stability label." These two labels provide a clear clinical outcome indication for each historical path.

[0088] The operation to obtain the subsequent stage chain involves performing a depth-first, time-oriented path traversal in the graph, with each starting node as the root. The specific algorithm is as follows:

[0089] 1. Location and Initialization: The graph node corresponding to a certain historical recovery stage is formally set as the "starting node" for the current traversal. An empty linked list is created to store the node sequence that makes up the chain, and the starting node is added to the linked list as the first element.

[0090] 2. Backward Traversal and Condition Check: Starting from the starting node, examine all directed edges emanating from that node. These directed edges represent the possibility of evolution from the current stage to the next. Since the graph is constructed strictly in chronological order connecting different stages of the same patient, any edge emanating from a node will point to a node that is later in time. Select the only directed edge emanating from that node (because the stage order of the same patient is fixed) and visit the "next node" it points to. When visiting this node, the core operation is to check its attribute fields. The program will read the node's attribute list and query whether it contains a Boolean flag for "atrial fibrillation recurrence" or "long-term stability".

[0091] 3. Recursive Extension and Termination: If the currently visited node does not contain any of the aforementioned termination markers, it indicates that the node is merely an intermediate recovery stage. In this case, append the node to the aforementioned linked list and use it as the new current node. Repeat step 2 to continue visiting its next node. This process repeats continuously, forming a traversal process that jumps through stages along the timeline.

[0092] 4. Path Completion: When a node is visited, if its attributes are checked and it contains either a "relapse marker" or a "long-term stable marker," the traversal immediately stops. At this point, the node with the termination marker is added to the end of the linked list as the "terminal node." Thus, the sequence of nodes stored in the linked list in the order of visit constitutes a complete evolutionary path starting from the initial node (a similar historical stage), going through several intermediate stages, and finally reaching a clear clinical outcome (relapse or stability). We define this as a "subsequent stage chain."

[0093] For example, suppose we retrieve a historical stage node A. Starting from A, we traverse through nodes B and C, finally reaching node D. Upon inspection, node D carries a "atrial fibrillation recurrence marker." Then, the node sequence A, B, C, and D constitutes a complete chain of subsequent stages, demonstrating the specific process of a historical patient similar to the current patient's final stage, who subsequently experienced the intermediate stages B and C before ultimately relapsing. For each historical stage node in the search result set, we independently perform the above traversal process once, thus obtaining multiple chains of subsequent stages equal to the number of search results. These chains serve as the initial blueprint for constructing the predicted trajectory.

[0094] In S5, the process of forming multiple candidate future evolution chains starting from the current state is as follows:

[0095] The purpose of this process is to restore and convert the "subsequent stage chain" obtained in the previous step, which is represented by a discrete stage node sequence, into a smooth trajectory in a continuous physiological state space, and then extend it from the patient's current state point through geometric transformation, ultimately forming a candidate future evolution chain that can be directly used for risk analysis.

[0096] First, the discrete stage representation of the chain needs to be transformed into a continuous trajectory. For a given chain of subsequent stages, it contains N nodes, corresponding to N historical recovery stages. In the 166-dimensional stage feature vector stored in each node, the first 50 dimensions are the mean vector of the state points of that stage, and the next dimension is the duration of that stage (in days). These two components are crucial.

[0097] 1. Trajectory Segment Generation: For the i-th node in the chain, extract its state point mean vector Mi (50-dimensional) and stage duration Ti (scalar, e.g., 28 days). In the physiological state space, the coordinates of the mean vector Mi represent the "central position" of the physiological state in this stage. We will construct a continuous trajectory segment representing the state evolution of this stage, with Mi as the geometric center. Specifically, we consider the time interval of this duration Ti as the domain, assuming that the state fluctuates slightly around the center Mi in a smooth manner within this stage. A simplified implementation model is to generate a smooth curve from the virtual starting point to the virtual ending point within this time interval, such that the state value at the midpoint of the time interval is exactly equal to Mi, and the average state value throughout the interval also approximates Mi. This can be achieved by constructing a bell-shaped curve or polynomial curve of appropriate amplitude with the midpoint of time as the center of symmetry, and superimposing it on the constant vector Mi. Ultimately, this trajectory segment is composed of a series of discrete time-state coordinate points. For example, if sampling is done at a 1-day interval, and Ti is 28 days, then 28 state points are generated. The set of these points in the state space appears as a smooth segment closely distributed around the center Mi.

[0098] 2. Segment Joining and Smoothing: Following the chain node order, the trajectory segments generated in stage i and stage i+1 are joined end-to-end on the timeline. However, direct joining may result in unnatural sharp turns or jumps in the trajectory at stage boundaries because the centers Mi and M{i+1} of the two stages may be different. To simulate a more natural transition between stages, spatial interpolation smoothing is required based on the weights of the directed edges connecting the two nodes in the graph. The edge weight (denoted as W) is the cosine similarity of the feature vectors of the two node stages, ranging from 0 to 1. A larger value indicates a more similar pattern between the two stages, and theoretically, the state transition should be smoother. Smoothing is performed within a time window near the joining point (e.g., the last 7 days of the previous segment and the first 7 days of the next segment). We use a cubic Hermitian interpolation function, adjusting the interpolation "tension" parameter based on the weight W. A higher W value allows for a larger tension parameter setting, making the interpolation curve more linear and gradual in the transition region. A lower W value allows for a greater curvature in the transition curve, enabling a faster transition from one state center to another. Through this process, two originally separate trajectory segments are merged into a single, continuous, and first-order smooth trajectory (i.e., the tangent direction changes continuously). This operation is repeated for all adjacent segments in the chain, ultimately resulting in a trajectory that represents the complete evolutionary history of the subsequent stage of the chain and extends continuously in the physiological state space; this is called the "initial trajectory."

[0099] However, the starting coordinates of this initial trajectory are the state center M1 of the first historical stage in the first subsequent stage chain (i.e., the stage similar to the final stage of our current patient). This coordinate lies within the state space of the historical patient itself. To predict the future of the current patient using this coordinate, spatial alignment is necessary.

[0100] 3. Coordinate Translation and Alignment: This is the crucial step in "grafting" the historical trajectory onto the current patient's state space. Let the starting point coordinates of the initial trajectory be Pstart (a 50-dimensional vector). Let the latest intrinsic physiological state vector obtained by the current patient through step S2 be Pcurrent (also a 50-dimensional vector), which represents the patient's physiological state coordinates "at this moment". Calculate the translation vector Voffset = Pcurrent - Pstart. This is a vector difference that indicates the direction and distance from the starting point of the historical trajectory to the current patient's state point. Then, for the coordinates of each sampling point on the initial trajectory (assuming there are L points), perform the following operation: For the coordinates Pj of the j-th point, calculate the new aligned coordinates Pj' = Pj + Voffset. This means that the entire historical trajectory has been rigidly translated, so that the starting point of the translated trajectory completely coincides with the current patient's state point Pcurrent.

[0101] 4. Generating Candidate Evolutionary Chains: After coordinate translation and alignment, the resulting new trajectory aligns its temporal starting point with the "present" and its spatial starting point with the patient's current state. This trajectory describes a possible future path starting from the current state, mimicking the morphological and evolutionary pattern of a specific subsequent historical stage chain. We formally label this processed trajectory as a "candidate future evolutionary chain." Its prediction of future states covers the total time elapsed from the current moment until its corresponding historical outcome (relapse or stabilization).

[0102] Finally, for each subsequent stage chain derived from each historical recovery stage retrieved in step S4, the entire process of trajectory segment generation, splicing smoothing, and coordinate translation alignment described above is performed independently and completely. Assuming 10 similar historical stages are retrieved, with each stage deriving one subsequent stage chain, then 10 candidate future evolution chains will be generated. This set of chains constitutes a diverse, data-driven, and visualized projection of the current patient's possible future fate, providing direct input for the final probability calculation and risk path selection. Each chain is a concrete and clear computable path in a 50-dimensional physiological state space, pointing from the known present to the unknown future.

[0103] In another preferred embodiment of the present invention, the process of calculating the likelihood score of each candidate future evolution chain in S6 is as follows:

[0104] Likelihood score is a comprehensive metric used to quantify the degree to which a candidate future evolutionary path matches the target patient's personal historical behavioral patterns, while also considering the geometric plausibility of the evolutionary path within the physiological state space. A higher score indicates that the predicted path is more likely to reflect the patient's true future evolutionary trend. Its calculation integrates information from two dimensions: event sequence consistency and trajectory geometric characteristics.

[0105] The first dimension assesses the alignment between the predicted event sequence and the patient's historical patterns. Each candidate future evolutionary chain is essentially a continuous trajectory in a 50-dimensional physiological state space. To evaluate its plausibility, this abstract state trajectory must first be "translated" back into a clinically understandable observational level, i.e., predicting the sequence of clinical events that may accompany its evolution. This is achieved through a pre-trained "state-event decoder." This decoder is a neural network model whose input is a physiological state vector, and whose output is the probability distribution of various clinical event types occurring in that state. Along the trajectory of the candidate evolutionary chain, state points are sampled at fixed time intervals (e.g., every 7 days). The state vector of each sampled point is input into the decoder, and the event type with the highest output probability is selected as the predicted event at that time point. Arranging the predicted events of all sampled points along the entire trajectory in chronological order yields a "predicted clinical event type sequence."

[0106] Next, we compare this predicted sequence with the patient's actual historical event sequence using three specific statistical features. First is the event type distribution. We statistically analyze the frequency of each of the 15 predefined event types in the predicted sequence, forming a 15-dimensional predicted distribution vector Ptype. Similarly, we statistically analyze the frequency of each type of event in the target patient's existing discrete clinical event sequences, forming a historical distribution vector Htype. We then calculate the cosine similarity between these two vectors, a value ranging from 0 to 1. A value closer to 1 indicates a greater similarity between the predicted event type composition and the patient's historical habits. This calculation result is called the first degree of consistency value F1.

[0107] Secondly, the average time interval between adjacent events is calculated. The average number of days between any two consecutive events in the predicted event sequence is denoted as Ipred. The average interval between adjacent events in the patient's historical event sequence is denoted as Ihist. The absolute value of the difference between these two average intervals, Dinterval, is equal to the absolute value of Ipred minus the absolute value of Ihist. Patients have individual preferences regarding event density; the smaller the difference in intervals, the more the predicted rhythm matches the individual pattern. The reciprocal of this difference is used to measure the degree of agreement, i.e., the second degree of agreement value F2 is equal to 1.0 divided by (Dinterval plus a very small constant, such as 0.01, to prevent division by zero errors). The larger the F2 value, the closer the predicted event rhythm is to the patient's historical rhythm.

[0108] Thirdly, the frequency of specific event combinations is considered. Based on clinical knowledge, a set of clinically relevant event combinations is defined. For example, "recording palpitation symptoms" and "adding antiarrhythmic medication within the next 7 days" are considered combination A, and "recording an increase in premature atrial contractions" and "re-examining an electrocardiogram within the next 14 days" are considered combination B. The frequency of each specific combination in the predicted event sequence is statistically analyzed to obtain a combination frequency vector Pcombo. Similarly, the frequency of these combinations in the historical sequence is statistically analyzed to obtain Hcombo. For each combination k, the absolute value Dcombok of the difference between its predicted frequency and historical frequency is calculated. Similarly, the consistency component Ck of each combination is calculated as 1.0 divided by (Dcombok plus 0.01). Finally, the arithmetic mean of Ck for all combinations is taken to obtain the third consistency value F3. F3 reflects the degree of matching between the association pattern between key clinical events in the predicted event sequence and the patient's historical pattern.

[0109] The second dimension assesses the geometric characteristics of the evolutionary chain trajectory itself. A reasonable physiological state evolution path should be relatively smooth in state space, without violent oscillations or unnecessary detours. To this end, two geometric metrics are introduced. The first is the trajectory curvature value. Curvature is a quantity in calculus describing the degree of bending of a curve. For the spatial path of the candidate evolutionary chain, which consists of discrete sampling points, an approximate method is used to calculate its average curvature. Specifically, for three consecutive sampling points on the trajectory, the curvature of the arc they form can be calculated. The curvature of all three consecutive point windows on the entire trajectory is calculated and averaged to obtain the average curvature value of the chain, Curvature. The larger the value, the sharper the trajectory turns, and the more unstable the evolution process may be. The second is the total path length value. The sum of the lengths of the line segments between all consecutive sampling points on the trajectory is calculated to obtain PathLength. This value measures the overall magnitude of state evolution. Within the same time span, an excessively long path may indicate that the state has undergone unnecessary frequent fluctuations or large-scale reversals in space.

[0110] Directly using compliance values ​​F1, F2, F3, curvature, and path length is inconvenient because they have different dimensions and meanings, and their contributions to "reasonableness" are inconsistent. F1, F2, and F3 are better the higher they are, while Curvature and PathLength are better the lower they are. Therefore, standardization and homogenization are necessary. For curvature and path length, their reciprocals are taken, making them also better the higher they are; that is, InvCurv equals 1.0 divided by Curvature, and InvLen equals 1.0 divided by PathLength. To prevent extreme values, Curvature and PathLength are smoothed before calculating the reciprocals, for example, by adding a baseline value.

[0111] Next, pre-defined weight coefficients are assigned to these five positiveized indicators to reflect their relative importance in the comprehensive evaluation. These weight coefficients are based on statistical analysis of historical data. For example, a feasible weight allocation is: event type distribution conformity weight W1 = 0.30, event interval conformity weight W2 = 0.25, event combination conformity weight W3 = 0.20, trajectory curvature reciprocal weight W4 = 0.15, and path length reciprocal weight W5 = 0.10. The sum of all weights is 1.0. Finally, the likelihood score Scorech of a candidate evolutionary chain Ch is calculated using the formula: Scorech = W1 × F1 + W2 × F2 + W3 × F3 + W4 × InvCurv + W5 × InvLen. This weighted summation yields a comprehensive score between 0 and 1, which quantitatively evaluates each candidate chain from two aspects: the conformity of event prediction to individual habits and the smoothness and rationality of the trajectory spatial morphology.

[0112] In step S6, the process of sorting, filtering, and outputting the path set based on the scores is as follows:

[0113] The first step is to set a quality threshold, i.e., a minimum score threshold. This threshold can be determined by analyzing historical validation data. For example, the scores of candidate chains generated from all historical patient data can be sorted, and the 20th percentile can be used as the threshold Tmin, which might be 0.65. The score of each candidate evolutionary chain is compared with Tmin. If a chain's score is lower than Tmin, it is considered to have too low a match with the patient's individual pattern or that the trajectory morphology is too unreasonable, and it is removed. This step filters out obviously unreliable noisy predictions.

[0114] The second step is to classify and sort the remaining evolutionary chains that have passed the threshold screening. Each chain inherits from its original historical subsequent stage chain, and therefore naturally carries the outcome label of its terminal node, namely "atrial fibrillation recurrence" or "long-term stability". Examining the terminal attributes of all remaining chains, they are clearly divided into two mutually exclusive subsets: the first subset contains all evolutionary chains with the endpoint labeled as atrial fibrillation recurrence, representing the path predicted to lead to recurrence; the second subset contains all evolutionary chains with the endpoint labeled as long-term stability, representing the path predicted to lead to stability. This classification intuitively demonstrates the probability of the two main future outcomes.

[0115] Next, within each subset, the chains are sorted in descending order based on their likelihood scores. In the first subset (relapse group), the chain with the highest score is ranked first, representing the path that best matches the patient's personal history and has the most reasonable trajectory among all predicted relapse paths, and is therefore the high-risk path that requires the most attention. The same descending order is applied within the second subset (stable group).

[0116] The third step is quota selection. To avoid information overload caused by too many output paths, it is necessary to select a number of the most representative paths from each subset. Two positive integers, M and K, are set, for example, M equals 3 and K equals 2. This means selecting the top 3 evolutionary chains in terms of score from the relapse group and the top 2 evolutionary chains in terms of score from the stable group. The values ​​of M and K can be adjusted according to different clinical emphases on risk warning or reassurance.

[0117] Finally, the selected M relapse prediction chains and K stability prediction chains are merged to form the final "personalized relapse risk prediction pathway set." This set typically contains about five pathways with high individual likelihood, clearly distinguishing between relapse and stability outcomes, and providing several representative evolutionary trajectories for each outcome. This pathway set is output in a visualized graphical and data format, providing clinicians with a data-driven, highly personalized, and rigorously rationally screened blueprint for future risk evolution, directly supporting precise clinical decisions regarding enhanced monitoring, treatment adjustments, or reassurance.

[0118] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An artificial intelligence-based method for predicting atrial fibrillation recurrence, characterized in that, Includes the following steps: S1. Input the discrete clinical event sequence recorded after the target patient's self-ablation, including symptom records, medication change records, and heart rhythm conclusion records. Each record is accompanied by an event timestamp and text description. S2. Input the discrete clinical event sequence into the event-driven hidden state deduction model. The model outputs an intrinsic physiological state time sequence path that evolves from the initial postoperative state to the current state in the preset physiological state space. S3. Based on the distribution density and transfer direction of state points in the intrinsic physiological state time-series path, identify multiple recovery stages with different evolution patterns, and generate a corresponding stage feature vector for each recovery stage. S4. Using the last recovery stage as the query stage, retrieve all historical recovery stages whose feature vector similarity exceeds a set threshold from the pre-constructed population recovery process map. S5. For each historical recovery stage, traverse backwards in chronological order until a node with atrial fibrillation recurrence marker or long-term stability marker is reached to obtain a complete chain of subsequent stages. Map each chain of subsequent stages back to the physiological state space to form multiple candidate future evolution chains starting from the current state. S6. Based on the event type distribution and time interval characteristics of the discrete clinical event sequence, calculate the likelihood score of each candidate future evolution chain, sort and filter all candidate future evolution chains according to the score, and output a personalized relapse risk prediction path set. In S2, the operation process of the event-driven hidden state deduction model is as follows: The physiological state space is defined as a multidimensional continuous vector space, where each dimension corresponds to a potential physiological factor. A preset initial state vector is assigned to the initial moment after ablation. Each record in the discrete clinical event sequence is iteratively processed according to the timestamp order. When processing a single record, the predicted probability distribution of various clinical events is calculated based on the currently estimated state vector. The difference between the predicted probability distribution and the actual event type is calculated. The text description of the current event record is encoded to obtain the event intensity vector. Based on the difference and the event intensity vector, the state vector estimates of the current time and the previous time are adjusted in reverse. After processing all event records, the intrinsic physiological state time sequence path is output. In step S3, the process of identifying the recovery phase and generating a phase feature vector is as follows: In the physiological state space, a fixed-length sliding window is used to traverse the intrinsic physiological state time sequence path. The mean vector of all state points in each sliding window is calculated as the center point of the window. The variance of the state points in each dimension in each sliding window is calculated. When the Euclidean distance between the center points of two consecutive sliding windows exceeds the stage transition distance threshold, the stage boundary point is recorded at the starting position of the window. When the rate of change of variance for two consecutive sliding windows exceeds the variance stability threshold, the stage boundary point is recorded at the beginning of the window. The temporal path is divided into multiple continuous segments based on all stage boundary points. Each segment is a recovery stage. For each recovery stage, its stage feature vector is composed of the mean and standard deviation of the state points in each dimension within the stage, the duration of the stage, the difference vector between the start and end state vectors of the stage, and the sum of the one-hot encoded vectors of the clinical event types within the stage.

2. The atrial fibrillation recurrence prediction method based on artificial intelligence according to claim 1, characterized in that, The process of adjusting the state vector estimate in reverse based on the difference and the event intensity vector is as follows: Construct a forward state transition network that takes a state vector as input and outputs the transition probability of the state vector. Construct a backward inference network that takes clinical event observation data as input and outputs the inference probability of the state vector. The current state vector before adjustment is input into the forward state transition network to obtain the prior state estimation probability distribution. The actual type and event intensity vector of the current event record are input into the backward inference network to obtain the posterior state estimation probability distribution. The KL divergence between the prior state estimation probability distribution and the posterior state estimation probability distribution is calculated. The gradient of the parameters of the forward state transition network and the backward inference network is calculated based on the KL divergence. The parameters of the forward state transition network and the backward inference network are updated along the gradient direction.

3. The atrial fibrillation recurrence prediction method based on artificial intelligence according to claim 1, characterized in that, In step S4, the process of retrieving historical recovery stages from the population recovery process map is as follows: The nodes in the group recovery process graph store the stage feature vectors of the historical recovery stages. Directed edges connect adjacent stage nodes of the same historical patient, and the edge weight is the similarity of the stage feature vectors of the two nodes. The cosine similarity between the stage feature vector of the query stage and the stage feature vectors of each node in the graph is calculated. Nodes with a cosine similarity exceeding the retrieval threshold are selected to form the preliminary retrieval results. For each node in the preliminary search results, the matching degree between the features of the previous recovery stage on the original path of the query stage and the features of the previous node in the graph is calculated. The cosine similarity and the matching degree are fused to perform a weighted score, and the set of historical recovery stages is selected based on the score.

4. The atrial fibrillation recurrence prediction method based on artificial intelligence according to claim 1, characterized in that, In S5, the process of obtaining the complete subsequent stage chain is as follows: In the population recovery process graph, the node corresponding to the retrieved historical recovery stage is located as the starting node. Starting from the starting node, the graph is traversed backward along the directed edges that start from the starting node and point to nodes later in the time sequence. When visiting the first subsequent node, the attributes of the starting node are checked to see if they contain an atrial fibrillation recurrence marker or a long-term stable marker. If not, the current subsequent node is recorded as the current chain node and the next subsequent node is visited along the directed edges. The node attribute checking and recording operation is repeated until a terminal node containing an atrial fibrillation recurrence marker or a long-term stable marker is visited. The starting node, all recorded current chain nodes, and the terminal node are connected in the order of visit time to form a complete subsequent stage chain.

5. The atrial fibrillation recurrence prediction method based on artificial intelligence according to claim 4, characterized in that, In S5, the process of forming multiple candidate future evolution chains starting from the current state is as follows: For each node in the subsequent stage chain, the mean vector component of the state point and the duration component of the stage are extracted from its stored stage feature vector. In the physiological state space, a trajectory segment representing the continuous evolution of the state is generated with the mean vector of the state point as the geometric center and the time interval corresponding to the duration of the stage as the domain. According to the time order of the nodes in the chain, adjacent trajectory segments are connected at the time endpoints. The trajectory shape at the connection point is spatially interpolated and smoothed according to the weight value of the directed edge connecting the corresponding nodes in the graph, forming a complete and continuous initial trajectory. Calculate the vector difference between the coordinates of the starting point of the initial trajectory in the physiological state space and the coordinates of the target patient's current state. Subtract the vector difference from the coordinates of all points on the initial trajectory to complete the translation and alignment of the trajectory space position. Mark the translated and aligned trajectory as a candidate future evolution chain starting from the current state.

6. The atrial fibrillation recurrence prediction method based on artificial intelligence according to claim 1, characterized in that, In step S6, the process of calculating the likelihood score for each candidate future evolutionary chain is as follows: Based on the state evolution sequence of the candidate future evolution chain, a predicted clinical event type sequence is mapped, and the event type distribution, average time interval between adjacent events, and frequency of specific event combinations in the sequence are statistically analyzed. The historical event type distribution, historical average time interval, and historical combination frequency corresponding to the discrete clinical event sequence of the target patient are obtained. The first degree of agreement between the prediction and the historical event type distribution is calculated, the reciprocal of the difference between the prediction and the historical average time interval is calculated as the second degree of agreement, and the average of the reciprocals of the differences between the prediction and the historical frequency of each specific event combination is calculated as the third degree of agreement. The trajectory curvature value and total path length value of the evolutionary chain in the physiological state space are calculated. The first, second, and third conformity values, the reciprocal of the curvature value, and the reciprocal of the total length value are multiplied by preset weight coefficients and then summed to obtain the likelihood score.

7. The atrial fibrillation recurrence prediction method based on artificial intelligence according to claim 6, characterized in that, In step S6, the process of sorting, filtering, and outputting the path set based on the scores is as follows: Set a minimum score threshold, compare the likelihood score of each candidate future evolutionary chain with the threshold, and remove all evolutionary chains with scores below the threshold. Check the endpoint attribute labels of the remaining evolutionary chains, and assign evolutionary chains labeled as atrial fibrillation recurrence to the first subset and those labeled as long-term stable to the second subset. Sort the evolutionary chains in the first and second subsets in descending order of their likelihood scores. Select the top M evolutionary chains from the first subset and the top K evolutionary chains from the second subset, where M and K are preset positive integers. Merge all selected evolutionary chains into the final output personalized recurrence risk prediction path set.

Images