Atrial fibrillation burden prediction method and system based on intermittent ppg signals

Abstract

This invention provides a method and system for predicting atrial fibrillation (AF) load based on intermittent PPG signals, belonging to the field of data processing technology. The method includes: training a first neural network using a first PPG signal with known AF results to obtain an AF probability prediction model; using the AF probability prediction model to predict the AF probability of a second PPG signal within a target time period and mapping it to an AF risk value; determining the AF probability at any given time based on the risk values ​​of two adjacent second PPG signals, integrating to obtain the total AF time, and thus obtaining the AF load; training a second neural network using historical AF loads from multiple time periods to obtain an AF load prediction model, used to predict the AF load for multiple future time periods. This invention achieves quantitative assessment and future trend prediction of AF load from discrete PPG signals to continuous AF load.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method and system for predicting atrial fibrillation load based on intermittent PPG signals. Background Technology

[0002] Atrial fibrillation (AF) is one of the most common arrhythmias in clinical practice, and its prevalence increases significantly with age. AF burden, defined as the proportion of AF episodes to total duration within a specific monitoring period, is a key quantitative indicator for assessing disease severity, guiding treatment decisions, and predicting stroke risk.

[0003] Currently, clinical assessment of atrial fibrillation burden is based on continuous electrocardiogram (ECG) monitoring in hospitals (such as 24-hour or longer Holter monitoring). This method provides high-precision, continuous ECG signals, which has irreplaceable reference value for in-hospital diagnosis of atrial fibrillation. However, its application to long-term, universally accessible health monitoring in home, community, and other out-of-hospital settings faces some limitations: the equipment and monitoring costs are high, the comfort and convenience of long-term wear are insufficient, and the routine monitoring duration (usually several days) is insufficient to fully capture the dynamic evolution of paroxysmal atrial fibrillation over weeks or months. Therefore, there is still a gap in long-term, continuous, low-burden atrial fibrillation burden monitoring methods outside of hospitals.

[0004] The widespread adoption of photoplethysmography (PPG) technology has provided effective technical support for bridging the gap in disease management between hospitals and homes. PPG indirectly senses the blood flow volume pulse wave in subcutaneous capillaries by detecting changes in the intensity of reflected or transmitted light emitted from the skin surface and received by photoelectric sensors. By analyzing the rhythmic irregularities of the PPG pulse wave sequence, the occurrence of atrial fibrillation can be inferred. Existing PPG-based atrial fibrillation detection technologies primarily rely on smart wearable devices (such as smartwatches) for intermittent measurements (e.g., automatic sampling every 30 minutes) and algorithms for real-time identification of atrial fibrillation events from single PPG signals. These algorithms are based on several technical principles: analysis of heart rate variability (HRV) time-domain / frequency-domain features, mining of pulse wave waveform morphology features, and end-to-end classification using deep learning neural networks. However, existing solutions focusing on interpreting atrial fibrillation events from single PPG signals have inherent limitations and fail to fully realize their application potential in long-term management. Specifically, the limitations are as follows: (1) Functional limitations: The output is only discrete binary labels, which cannot provide continuously quantified atrial fibrillation load values; (2) Data modeling limitations: Each intermittent measurement is treated as an isolated event, which fails to model the continuity of atrial fibrillation status in the time dimension, resulting in the inability to accurately reconstruct continuous time load by directly aggregating discrete detection results; (3) Prediction limitations: It only provides a static description of the events that have occurred and lacks the ability to make forward-looking predictions of future load evolution trends using historical time-series data. Therefore, the existing detection technology based on intermittent PPG has significant gaps in realizing the complete advanced chain from "discrete event detection" to "continuous load quantification" and then to "dynamic trend prediction". In summary, the current technical path based on intermittent PPG detection is difficult to meet the advanced requirements for modeling and predicting the dynamic evolution process of atrial fibrillation. Summary of the Invention

[0005] This invention provides a method and system for predicting atrial fibrillation load based on intermittent PPG signals. It can continuously predict atrial fibrillation load in the time dimension using multiple intermittent PPG signals and can predict future atrial fibrillation load using historical atrial fibrillation load over multiple time periods.

[0006] In a first aspect, the present invention provides a method for predicting atrial fibrillation burden based on intermittent PPG signals, comprising:

[0007] The first PPG signal of multiple known atrial fibrillation results is used as the training set to train the first neural network, and the trained first neural network is used as the atrial fibrillation probability prediction model. Acquire all valid second PPG signals within the target time period; The probability of atrial fibrillation corresponding to each second PPG signal is predicted using an atrial fibrillation probability prediction model. Map the probability of atrial fibrillation corresponding to each second PPG signal to the risk value of atrial fibrillation corresponding to each second PPG signal; The probability of atrial fibrillation at any time within the measurement period of two adjacent second PPG signals is determined based on the atrial fibrillation risk value corresponding to two adjacent second PPG signals. Integrate the atrial fibrillation probability at all times within the target time period to obtain the total atrial fibrillation time within the target time period; The ratio of the total atrial fibrillation time within the target time period to the duration of the target time period is taken as the atrial fibrillation load within the target time period. The historical atrial fibrillation load from multiple time periods is used as a training set to train the second neural network, which is then used as an atrial fibrillation load prediction model. Atrial fibrillation load prediction model is used to predict atrial fibrillation load in multiple future time periods.

[0008] Optionally, the step of using the first PPG signals of multiple known atrial fibrillation results as a training set to train the first neural network, and then using the trained first neural network as an atrial fibrillation probability prediction model, includes: The first neural network is used to map the first PPG signal to the first atrial fibrillation probability; The total loss of the first neural network is calculated based on the first PPG signal of each known atrial fibrillation result and the first atrial fibrillation probability corresponding to each first PPG signal. The first neural network is trained according to the loss function until it converges, and the trained first neural network is obtained as a model for predicting the probability of atrial fibrillation.

[0009] Optionally, the step of calculating the total loss of the first neural network based on the first PPG signal for each known atrial fibrillation result and the first atrial fibrillation probability corresponding to each first PPG signal includes: Calculate the total loss of the first neural network using the following formula: ; Among them, L det α represents the total loss of the first neural network; N is the total number of first PPG signals with known atrial fibrillation results; α represents the weighting coefficient of the first PPG signal corresponding to the occurrence of atrial fibrillation. The probability of atrial fibrillation corresponding to the nth first PPG signal; γ represents the modulation factor, which is a preset constant; y n The result of atrial fibrillation corresponding to the nth first PPG signal; y n =1 indicates atrial fibrillation has occurred; y n =0 indicates that atrial fibrillation did not occur.

[0010] Optionally, mapping the atrial fibrillation probability corresponding to each second PPG signal to the atrial fibrillation risk value corresponding to each second PPG signal includes: If the probability of atrial fibrillation corresponding to the second PPG signal is greater than the first probability threshold, the risk value of atrial fibrillation corresponding to the second PPG signal is set to 1. If the probability of atrial fibrillation corresponding to the second PPG signal is less than the second probability threshold, the risk value of atrial fibrillation corresponding to the second PPG signal is determined to be 0. If the probability of atrial fibrillation corresponding to the second PPG signal is less than or equal to the first probability threshold and greater than or equal to the second probability threshold, the probability value of atrial fibrillation corresponding to the second PPG signal shall be used as the risk value of atrial fibrillation corresponding to the second PPG signal.

[0011] Optionally, determining the probability of atrial fibrillation at any time within the measurement period of two adjacent second PPG signals based on the atrial fibrillation risk values ​​corresponding to two adjacent second PPG signals includes: The probability of atrial fibrillation at any time within the measurement period of two adjacent second PPG signals is calculated using the following formula: ; in, R represents the probability of atrial fibrillation at time t; a The risk value of atrial fibrillation corresponding to the earlier measured second PPG signal among two adjacent second PPG signals; w a For R a The weights of R; b The risk value of atrial fibrillation corresponding to the later-measured second PPG signal among two adjacent second PPG signals; w b For R b The weights are: e is the natural constant; λ is a predefined positive constant.

[0012] Optionally, training the second neural network using historical atrial fibrillation load from multiple time periods as a training set, and then using the trained second neural network as an atrial fibrillation load prediction model, includes: The historical atrial fibrillation load from multiple time periods is input into the encoder to encode the historical atrial fibrillation load within each time period into a hidden state; The attention mechanism is used in conjunction with the hidden state of the encoder at each time interval to determine the attention weight of the decoder during decoding at each time interval; The hidden states of the encoder at each time interval are weighted and summed using the attention weights of the decoder during decoding at each time interval to obtain the context vector of the decoder at each time interval. The predicted atrial fibrillation load for the current time period is determined based on the decoder's predicted atrial fibrillation load for the previous time period and the context vector for the current time period. The second neural network is trained by minimizing the mean square error between the predicted and actual values ​​of atrial fibrillation load in each time period until the difference between the predicted and actual values ​​of atrial fibrillation load in each time period is within a preset range. This trained second neural network is then used as an atrial fibrillation load prediction model.

[0013] Secondly, the present invention provides an atrial fibrillation load prediction system based on intermittent PPG signals, comprising: The first training module is used to train the first neural network with the first PPG signals of multiple known atrial fibrillation results as the training set, and to use the trained first neural network as an atrial fibrillation probability prediction model. The acquisition module is used to acquire all valid second PPG signals within the target time period. The first prediction module is used to predict the probability of atrial fibrillation corresponding to each second PPG signal using an atrial fibrillation probability prediction model. The mapping module is used to map the atrial fibrillation probability corresponding to each second PPG signal to the atrial fibrillation risk value corresponding to each second PPG signal. The first determining module is used to determine the probability of atrial fibrillation at any time within the measurement time period of two adjacent second PPG signals based on the atrial fibrillation risk value corresponding to two adjacent second PPG signals. The integration module is used to integrate the atrial fibrillation probability at all times within the target time period to obtain the total atrial fibrillation time within the target time period. The second determining module is used to take the ratio of the total atrial fibrillation time within the target time period to the duration of the target time period as the atrial fibrillation load within the target time period. The second training module is used to train the second neural network with historical atrial fibrillation load from multiple time periods as the training set, so that the trained second neural network can be used as an atrial fibrillation load prediction model. The second prediction module is used to predict the atrial fibrillation load for multiple future time periods using an atrial fibrillation load prediction model.

[0014] Thirdly, the present invention provides a computer device including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the atrial fibrillation load prediction method based on intermittent PPG signals described in the first aspect.

[0015] Fourthly, the present invention provides a computer-readable storage medium for storing a computer program; when executed by a processor, the computer program implements the steps of the atrial fibrillation load prediction method based on intermittent PPG signals described in the first aspect.

[0016] Fifthly, the present invention provides a computer program product comprising computer-executable instructions or a computer program, wherein when the computer-executable instructions or the computer program are executed by a processor, the steps of the atrial fibrillation load prediction method based on intermittent PPG signals described in the first aspect are implemented.

[0017] This invention provides a method for predicting atrial fibrillation load based on intermittent PPG signals. By constructing a two-stage deep learning architecture of "atrial fibrillation probability prediction model - continuous load quantification - atrial fibrillation load prediction model", it realizes a complete technical chain from discrete PPG measurement to continuous atrial fibrillation load assessment, and then to future load trend prediction, effectively solving the three core problems existing in the prior art: First, this invention represents a technological leap from discrete event detection to continuous load quantification. Existing technologies only output binary labels and cannot provide continuously quantified atrial fibrillation load values. This invention maps the atrial fibrillation probability of intermittent PPG signals to risk values, and constructs a continuous-time atrial fibrillation probability function based on the risk values ​​of adjacent measurement points using an exponentially decaying weighted interpolation algorithm. The total atrial fibrillation time and load within the target time period are then calculated through integration. This technical solution transforms sparse, discrete detection point information into a smooth estimate of physiological states over a continuous time period, enabling the calculation of continuous atrial fibrillation load based on intermittent measurement data.

[0018] Second, a continuity modeling mechanism for atrial fibrillation (AF) status in the temporal dimension was established. Existing technologies treat each intermittent measurement as an isolated event, failing to model the temporal continuity of AF status. This invention introduces an exponentially decaying temporal interpolation algorithm, utilizing the AF risk values ​​and time weights of two adjacent PPG signal measurement points to determine the probability of AF at any time within the measurement interval. This fully considers the prior continuity of physiological events in time—that is, the AF status at a certain moment is more likely to be similar to its neighboring measurement results, and the similarity decays exponentially with the increase of the time interval. This modeling method reflects the physiological characteristics of AF attacks better than simple linear interpolation, significantly improving the accuracy and reliability of continuous load estimation.

[0019] Third, it enables proactive prediction of future atrial fibrillation load trends. Existing technologies only provide static descriptions of past events and lack predictive capabilities. This invention, based on continuous historical atrial fibrillation load data, further constructs a second neural network (an encoder-decoder architecture combined with an attention mechanism) to learn long-term dependencies and periodic patterns in historical atrial fibrillation load sequences, thereby predicting atrial fibrillation load over multiple future time periods. This provides clinicians with early warning information on disease progression, helping to develop intervention strategies in advance and shifting the treatment model from "passive monitoring" to "active prediction."

[0020] Fourth, an end-to-end automated atrial fibrillation load assessment and prediction system has been established. This invention achieves fully automated processing from raw PPG signals to final load prediction through the cascading of two neural networks, eliminating the need for manual feature engineering, lowering the technical implementation threshold, and improving the system's deployability and practicality. It is particularly suitable for long-term outpatient monitoring scenarios such as smart wearable devices. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a method for predicting atrial fibrillation load based on intermittent PPG signals, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the first neural network provided in an embodiment of the present invention; Figure 3 A schematic diagram of the intermittent PPG atrial fibrillation burden assessment process provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the structure of the second neural network provided in an embodiment of the present invention; Figure 5 A schematic diagram of the continuous load prediction results of atrial fibrillation based on intermittent PPG signals provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a system for predicting atrial fibrillation load based on intermittent PPG signals, provided as an embodiment of the present invention. Detailed Implementation

[0023] 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.

[0024] Example 1 like Figure 1 As shown, this embodiment provides a method for predicting atrial fibrillation burden based on intermittent PPG signals, including: Step 101: Use the first PPG signals of multiple known atrial fibrillation results as a training set to train the first neural network, and use the trained first neural network as an atrial fibrillation probability prediction model.

[0025] To capture the complex morphological and rhythmic features related to atrial fibrillation in the first PPG signal, this embodiment constructs an end-to-end deep learning model, namely the first neural network, to replace traditional analysis methods that rely on manual feature engineering. The deep learning model can automatically learn discriminative time-domain and frequency-domain representations from the original signal, making it particularly suitable for physiological signals like PPG, which exhibit strong individual variability and noise sensitivity. The first neural network employs a hybrid architecture of a one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (BiLSTM), as shown in the structure below. Figure 2 As shown.

[0026] For example, this step uses a first neural network to map the first PPG signal to a first atrial fibrillation probability: .

[0027] Among them, P AF (t') represents the probability of the first atrial fibrillation corresponding to the first PPG signal x(t') of the t'th known atrial fibrillation result; σ(·) is the Sigmoid activation function; W o The output layer weights are represented by BiLSTM(·), which represents a bidirectional LSTM layer responsible for modeling long-term temporal dependencies of signals and identifying rhythmic irregularities; CNN(·;θ) represents the output layer weights. c ) indicates that the parameter is θ c The 1D-CNN stack is responsible for extracting local waveform features; b o This is the first bias term.

[0028] In training an atrial fibrillation probability prediction model based on the first PPG signal, there are a large number of "easily classified samples" (such as regular sinus rhythm or typical atrial fibrillation), while a small number of "difficult-to-classify samples" (such as segments with premature atrial contractions, motion noise, or poor signal quality) are crucial to the model's generalization ability. Traditional weighted cross-entropy loss only statically adjusts weights according to class frequency, failing to differentiate sample difficulty. This results in the training process being dominated by a large number of easily classified samples, making it difficult for the model to focus on overcoming key challenges. This embodiment introduces a dynamic difficulty modulation factor γ into the loss function. Its mechanism is to reduce the loss contribution of (easily classified) samples correctly classified by the model with high confidence, while maintaining attention to low-confidence (difficult-to-classify) samples.

[0029] The total loss of the first neural network is calculated based on the first PPG signal for each known atrial fibrillation outcome and the first atrial fibrillation probability corresponding to each first PPG signal. For example, the total loss of the first neural network is calculated according to the following formula: .

[0030] Among them, L detα represents the total loss of the first neural network; N represents the total number of first PPG signals with known atrial fibrillation results (including atrial fibrillation events and non-atrial fibrillation events); α represents the weighting coefficient of the first PPG signal corresponding to atrial fibrillation; α is used to balance the imbalance between the number of positive samples (first PPG signals labeled as atrial fibrillation events) and negative samples (first PPG signals labeled as non-atrial fibrillation events), and is set according to the ratio of positive samples to negative samples in the training set; The probability of atrial fibrillation corresponding to the nth first PPG signal; γ represents the modulation factor, which is a preset constant; y n The result of atrial fibrillation corresponding to the nth first PPG signal; y n =1 indicates atrial fibrillation has occurred; y n =0 indicates that atrial fibrillation did not occur.

[0031] By introducing a modulation factor γ into the loss function, the loss contribution of easily distinguishable samples is dynamically reduced, while attention is maintained on difficult samples (such as segments with atrial premature beats, motion noise, or poor signal quality). This solves the problems of class imbalance and sample difficulty differences, and significantly improves the model's generalization ability and recognition accuracy for complex boundary situations.

[0032] The first neural network is trained according to the loss function until it converges, and the trained first neural network is obtained as a model for predicting the probability of atrial fibrillation.

[0033] By employing an end-to-end deep learning architecture (a hybrid of 1D-CNN and BiLSTM), discriminative time-domain and frequency-domain representations are automatically learned from the original PPG signal, replacing traditional manual feature engineering. This effectively addresses the issues of large individual differences and noise sensitivity in PPG signals, improving the accuracy and robustness of atrial fibrillation probability prediction.

[0034] Step 102: Obtain all valid second PPG signals within the target time period.

[0035] Step 103: Use the atrial fibrillation probability prediction model to predict the atrial fibrillation probability corresponding to each second PPG signal.

[0036] In steps 102-103, during the evaluation period T total Within a 24-hour period, assume there are M valid measurements of the (second PPG signal), with each measurement lasting for T seconds. measure The corresponding timestamp is The atrial fibrillation probability prediction model outputs the atrial fibrillation probability as follows: .

[0037] Step 104: Map the atrial fibrillation probability corresponding to each second PPG signal to the atrial fibrillation risk value corresponding to each second PPG signal.

[0038] If a fixed threshold (such as 0.5) is used directly for binarization, the model confidence information contained in the probability value will be lost, resulting in coarse subsequent estimates. For example, probabilities of 0.51 and 0.95 are equivalent after binarization, but their reliability is significantly different.

[0039] Therefore, this embodiment introduces a high-confidence positive threshold. (First probability threshold, such as 0.8) and low confidence negative threshold (The second probability threshold, such as 0.2), is used to soften the probability through a transition rule. Mapped to a risk value R that better reflects the degree of certainty m : .

[0040] Among them, R m The risk value of atrial fibrillation corresponding to the second PPG signal measured in the m-th measurement; The probability value of atrial fibrillation event is the output of the atrial fibrillation probability prediction model for the second PPG signal measured m times; I(·) is an indicator function, which takes a value of 1 when the condition in parentheses is true, and 0 otherwise, i.e.: If the probability of atrial fibrillation corresponding to the second PPG signal is greater than the first probability threshold, the risk value of atrial fibrillation corresponding to the second PPG signal is set to 1.

[0041] If the probability of atrial fibrillation corresponding to the second PPG signal is less than the second probability threshold, the risk value of atrial fibrillation corresponding to the second PPG signal is determined to be 0.

[0042] If the probability of atrial fibrillation corresponding to the second PPG signal is less than or equal to the first probability threshold and greater than or equal to the second probability threshold, the probability value of atrial fibrillation corresponding to the second PPG signal shall be used as the risk value of atrial fibrillation corresponding to the second PPG signal.

[0043] Step 105: Determine the probability of atrial fibrillation at any time within the measurement period of the two adjacent second PPG signals based on the atrial fibrillation risk values ​​corresponding to the two adjacent second PPG signals.

[0044] Step 105 aims to address the challenge of estimating continuous-time atrial fibrillation load from discrete data. To achieve this goal, this embodiment proposes a time-series interpolation algorithm based on exponentially decaying weighted data, from a discrete risk value sequence {R}. m ,t m Construct a system covering the entire evaluation period T total The probability function of continuous atrial fibrillation The basic idea is that the atrial fibrillation state at a certain moment is more likely to be similar to the measurement results in its vicinity, and the similarity decreases exponentially with the increase of time interval. This is more reflective of the spatiotemporal correlation of physiological events than simple linear interpolation.

[0045] Specifically, for any time point t within the evaluation period, find the timestamps t of the two most recent valid measurements before and after it. a and t b (satisfying t) a ≤t≤t b The corresponding risk values ​​are R. a and R b The probability estimate of atrial fibrillation at time t is the weighted average of the two values: .

[0046] in, R represents the probability of atrial fibrillation at time t; a The risk value of atrial fibrillation corresponding to the earlier measured second PPG signal among two adjacent second PPG signals; w a For R a The weights of R; b The risk value of atrial fibrillation corresponding to the later-measured second PPG signal among two adjacent second PPG signals; w b For R b The weight of ; e is the natural constant; λ represents the attenuation coefficient, which is a preset positive constant used to control the attenuation rate of the influence of the measurement point as the time interval increases.

[0047] Step 106: Integrate the atrial fibrillation probability at all times within the target time period to obtain the total atrial fibrillation time within the target time period.

[0048] After obtaining the probability of continuous atrial fibrillation, the total atrial fibrillation time T within the assessment period is estimated by integrating this function over the assessment period (or summing the densely sampled points in the discrete implementation). AF : .

[0049] Here, dt represents the integral infinitesimal element, which represents the infinite subdivision of time.

[0050] Step 107: The ratio of the total atrial fibrillation time within the target time period to the duration of the target time period is taken as the atrial fibrillation load within the target time period.

[0051] Atrial fibrillation burden B is obtained by calculating the proportion of atrial fibrillation time to the total assessment time and expressed as a percentage: .

[0052] Through mathematical modeling, discrete and sparse detection point information is transformed into a smooth estimate of physiological state over a continuous period of time, thereby achieving the core objective of calculating continuous atrial fibrillation load based on intermittent measurement data.

[0053] In steps 102-107, the output of the atrial fibrillation probability prediction model is used as input, and the output is the atrial fibrillation load value, aiming to overcome the inherent technical limitations of intermittent PPG measurement. Although intermittent measurement has the advantages of convenience and low load, its data is discretely distributed in time, making it impossible to directly obtain the duration of atrial fibrillation episodes under continuous monitoring, thus making it difficult to calculate the atrial fibrillation load using traditional methods. Addressing the core issue of "how to reliably estimate the overall atrial fibrillation load over a continuous time period from discrete, discontinuous atrial fibrillation event identification results," this embodiment proposes an innovative solution: The core of this embodiment is to construct a mapping model from discrete probability sequences to continuous load estimation. This model takes the intermittent atrial fibrillation probability sequence output by the atrial fibrillation probability prediction model as input, and through two key steps—probability softening and risk value sequence construction, and time series interpolation and continuous integral estimation—it ultimately outputs a quantified atrial fibrillation load value, thus providing a continuous and reliable time-series data foundation for subsequent trend prediction. This process systematically addresses the shortcomings of intermittent data in terms of continuity, information utilization, and quantitative assessment. The overall process is as follows: Figure 3 As shown.

[0054] Step 108: Use historical atrial fibrillation load from multiple time periods as a training set to train the second neural network, and use the trained second neural network as an atrial fibrillation load prediction model.

[0055] The historical atrial fibrillation load sequence, composed of historical loads across multiple time periods, reveals complex trends and periodic dependencies. To effectively capture these long-term and periodic patterns, this embodiment employs an end-to-end deep learning architecture capable of learning and fusing these intrinsic patterns. Traditional recurrent neural networks may suffer from information decay when processing long sequences. Therefore, this embodiment constructs an encoder-decoder architecture and introduces an attention mechanism. The goal of this model (the second neural network) is to learn a mapping function f: ,in, H is the prediction step size (e.g., H=7 days), and the structure of the second neural network is as follows: Figure 4 As shown.

[0056] The encoder is implemented using a Long Short-Term Memory (LSTM) network to encode the input historical sequence into a series of hidden states, which condense the sequence information up to each time point. In this embodiment, historical atrial fibrillation load from multiple time periods is input into the encoder to encode the historical atrial fibrillation load within each time period into a hidden state: .

[0057] Among them, h t''This represents the hidden state vector of the encoder at the t'' time interval; LSTM enc (·) represents the long short-term memory network unit of the encoder; B t'' h represents the historical atrial fibrillation load value for the t''th time period; t'' Let represent the hidden state vector of the encoder in the t''-1 time period.

[0058] When generating a future prediction at each step, the decoder does not rely equally on all historical hidden states. The role of the attention mechanism is to dynamically calculate which parts of the historical sequence the decoder should "pay attention to" in the current step. This embodiment uses the attention mechanism in conjunction with the encoder's hidden states at each time interval to determine the attention weights of the decoder during decoding at each time interval: .

[0059] in, Let be the attention weight for the t''-th hidden state of the encoder at the j-th decoding step; exp(·) denotes the natural exponential function; score(·) is a relevance scoring function used to calculate the encoder hidden state h. t'' With the previous hidden state s of the decoder j-1 The degree of correlation between them; s j-1 This represents the hidden state of the decoder at step j-1; D represents the total number of time steps in the historical atrial fibrillation load sequence.

[0060] By using the attention weights of the decoder during each time interval, the hidden states of the encoder at each time interval are weighted and summed to obtain the context vector of the decoder at each time interval: .

[0061] Among them, c j This is the context vector of the decoder at step j (the j-th time interval).

[0062] The predicted atrial fibrillation load for the current time period is determined based on the decoder's predicted atrial fibrillation load from the previous time period and the context vector for the current time period: .

[0063] in, W represents the predicted atrial fibrillation load for the (D+j)th time period. p The weight matrix of the output layer; LSTM dec (·) denotes the LSTM unit of the decoder, whose input is the concatenation of the previous time step's prediction and the current context vector. And the hidden state s of the decoder at the previous moment. j-1 b p This is the second bias term.

[0064] The second neural network is trained by minimizing the mean square error between the predicted and actual values ​​of atrial fibrillation load in each time period until the difference between the predicted and actual values ​​of atrial fibrillation load in each time period is within a preset range. This trained second neural network is then used as an atrial fibrillation load prediction model.

[0065] For example, the mean square error loss L of the second neural network pred for: .

[0066] Among them, B D+j This represents the actual atrial fibrillation load during the (D+j)th time period.

[0067] By introducing an encoder-decoder architecture with an attention mechanism, the decoder can dynamically focus on key time periods in historical sequences when predicting future load, effectively capturing long-term dependencies and periodic patterns. This overcomes the information decay problem of traditional RNNs and significantly improves the accuracy of multi-step atrial fibrillation load prediction.

[0068] Step 109: Use the atrial fibrillation load prediction model to predict the atrial fibrillation load for multiple future time periods.

[0069] After the atrial fibrillation load prediction model predicts atrial fibrillation load for multiple future time periods, it is necessary to extract easily understandable trend conclusions. Therefore, this embodiment calculates the relative rate of change r of the predicted average load compared to the historical average load: .

[0070] Where mean(·) is the mean function; Indicates a future atrial fibrillation load prediction sequence; B hist This represents a historical atrial fibrillation burden prediction sequence.

[0071] Based on r, future trends can be categorized as follows: Downward trend: r ≤ -5%; Stable trend: -5% < r < 5%; Upward trend: r ≥ 5%.

[0072] like Figure 5 The diagram shown illustrates the continuous load estimation process for atrial fibrillation based on intermittent PPG measurements, demonstrating the complete calculation flow from discrete detection results to continuous load estimation during the quantitative assessment of atrial fibrillation load.

[0073] (1) Input data (discrete risk sequence) Figure 5 The red dots represent the atrial fibrillation risk values ​​(R values) at a series of discrete time points obtained through the front-end PPG atrial fibrillation prediction model within a 24-hour assessment period. m Each risk value is derived from the original atrial fibrillation detection probability. Based on the probability softening rule provided in this embodiment (using high / low confidence thresholds), it is mapped to have both classification deterministic and uncertain information.

[0074] (2) Core processing (construction of continuous probability function) Figure 5 The smooth blue curve represents the continuous-time atrial fibrillation probability function constructed using the "exponential decay weighted interpolation algorithm" proposed in this embodiment. This algorithm is based on the prior knowledge of the temporal continuity of physiological events. It controls the attenuation rate of the influence of neighboring measurement points through an attenuation coefficient λ (λ=0.005 in this embodiment), thereby reducing the sparse, discrete risk points {R}. m ,t m This process transforms data into a smoothly varying continuous probability curve covering the entire monitoring period. This systematically solves the core technical challenge of estimating continuous states from intermittent measurement data.

[0075] (3) Output results (load quantization) The area of ​​the blue-filled region enclosed by the curve and the horizontal axis (time axis) represents the estimated cumulative atrial fibrillation episode time. The ratio of this area to the total monitoring time (rectangular area) is the final quantified atrial fibrillation burden B. Figure 5 The calculated example is 42.5%. The load value B objectively and continuously reflects the severity of atrial fibrillation during the monitoring period as a percentage.

[0076] (4) Demonstration of technical effects Figure 5 This clearly demonstrates that this embodiment, through innovative algorithm mapping, achieves a crucial leap from discrete, sparse "point-like" detection results to continuous, integrable "linear" load quantization. This provides a reliable data foundation for subsequent dynamic trend prediction based on continuous time-series data.

[0077] In summary, this embodiment provides a method for predicting atrial fibrillation load based on intermittent PPG signals. It maps the atrial fibrillation probability of intermittent PPG signals to risk values, and constructs a continuous-time atrial fibrillation probability function based on the risk values ​​of adjacent measurement points using an exponentially decaying weighted interpolation algorithm. The total atrial fibrillation time and load within the target time period are then calculated through integration. This technical solution transforms sparse, discrete detection point information into a smooth estimate of physiological states over a continuous time period, enabling the calculation of continuous atrial fibrillation load based on intermittent measurement data.

[0078] By introducing an exponentially decaying temporal interpolation algorithm, the probability of atrial fibrillation at any time within the measurement interval is determined using the atrial fibrillation risk values ​​and their time weights at two adjacent PPG signal measurement points. This fully considers the prior continuity of physiological events in time—that is, the atrial fibrillation state at a certain moment is more likely to be similar to its neighboring measurement results, and the similarity decays exponentially with the increase of the time interval. This modeling method reflects the physiological characteristics of atrial fibrillation attacks better than simple linear interpolation, significantly improving the accuracy and reliability of continuous load estimation.

[0079] Based on continuous historical atrial fibrillation load data, a second neural network (encoder-decoder architecture combined with an attention mechanism) is further constructed to learn the long-term dependencies and periodic patterns in the historical atrial fibrillation load sequence, thereby enabling the prediction of atrial fibrillation load in multiple future time periods. This provides clinicians with early warning information on disease progression, helping to formulate intervention strategies in advance and realizing a shift from a "passive monitoring" to a "proactive prediction" approach to diagnosis and treatment.

[0080] By cascading two neural networks, the entire process from raw PPG signals to final load prediction is automated, eliminating the need for manual feature engineering, lowering the technical implementation threshold, and improving the system's deployability and practicality. It is particularly suitable for long-term off-site monitoring scenarios such as smart wearable devices.

[0081] Example 2 Based on the same inventive concept as Embodiment 1, this embodiment provides an atrial fibrillation load prediction system based on intermittent PPG signals. Since the principle of solving the problem in this system is similar to that of the atrial fibrillation load prediction method based on intermittent PPG signals described in Embodiment 1, the implementation of this system can refer to the implementation of the atrial fibrillation load prediction method based on intermittent PPG signals.

[0082] like Figure 6 As shown, this embodiment provides an atrial fibrillation load prediction system based on intermittent PPG signals, including: The first training module 10 is used to train the first neural network using the first PPG signals of multiple known atrial fibrillation results as a training set, so as to use the trained first neural network as an atrial fibrillation probability prediction model.

[0083] Acquisition module 20 is used to acquire all valid second PPG signals within the target time period.

[0084] The first prediction module 30 is used to predict the probability of atrial fibrillation corresponding to each second PPG signal using an atrial fibrillation probability prediction model.

[0085] The mapping module 40 is used to map the atrial fibrillation probability corresponding to each second PPG signal to the atrial fibrillation risk value corresponding to each second PPG signal.

[0086] The first determining module 50 is used to determine the probability of atrial fibrillation at any time within the measurement time period of two adjacent second PPG signals based on the atrial fibrillation risk values ​​corresponding to the two adjacent second PPG signals.

[0087] The integration module 60 is used to integrate the atrial fibrillation probability at all times within the target time period to obtain the total atrial fibrillation time within the target time period.

[0088] The second determining module 70 is used to take the ratio of the total atrial fibrillation time within the target time period to the duration of the target time period as the atrial fibrillation load within the target time period.

[0089] The second training module 80 is used to train the second neural network with historical atrial fibrillation load from multiple time periods as the training set, so that the trained second neural network can be used as an atrial fibrillation load prediction model.

[0090] The second prediction module 90 is used to predict the atrial fibrillation load for multiple future time periods using an atrial fibrillation load prediction model.

[0091] For more detailed information on the working process of each of the above modules, please refer to the relevant content disclosed in Example 1, which will not be repeated here.

[0092] Example 3 This embodiment provides a computer device, including a processor and a memory; wherein, when the processor executes the computer program stored in the memory, it implements the steps of the atrial fibrillation load prediction method based on intermittent PPG signals described in Embodiment 1.

[0093] For a more detailed explanation of the above method, please refer to the relevant content disclosed in Example 1, which will not be repeated here.

[0094] Example 4 This embodiment provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the atrial fibrillation load prediction method based on intermittent PPG signals described in Embodiment 1.

[0095] For a more detailed explanation of the above method, please refer to the relevant content disclosed in Example 1, which will not be repeated here.

[0096] Example 5 This embodiment provides a computer program product, including computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, they implement the steps of the atrial fibrillation load prediction method based on intermittent PPG signals described in Embodiment 1.

[0097] For a more detailed explanation of the above method, please refer to the relevant content disclosed in Example 1, which will not be repeated here.

[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems, devices, storage media, and computer program products disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0099] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0100] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0101] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0102] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0103] The present invention has been described in detail above with reference to specific embodiments and exemplary examples; however, these descriptions should not be construed as limiting the present invention. Those skilled in the art will understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and embodiments of the present invention without departing from the spirit and scope of the invention, and all such modifications and improvements fall within the scope of the present invention. The scope of protection of the present invention is defined by the appended claims.

Claims

1. A method for predicting atrial fibrillation load based on intermittent PPG signals, characterized in that, include: The first PPG signal of multiple known atrial fibrillation results is used as the training set to train the first neural network, and the trained first neural network is used as the atrial fibrillation probability prediction model. Acquire all valid second PPG signals within the target time period; The probability of atrial fibrillation corresponding to each second PPG signal is predicted using an atrial fibrillation probability prediction model. Map the probability of atrial fibrillation corresponding to each second PPG signal to the risk value of atrial fibrillation corresponding to each second PPG signal; The probability of atrial fibrillation at any time within the measurement period of two adjacent second PPG signals is determined based on the atrial fibrillation risk value corresponding to two adjacent second PPG signals. Integrate the atrial fibrillation probability at all times within the target time period to obtain the total atrial fibrillation time within the target time period; The ratio of the total atrial fibrillation time within the target time period to the duration of the target time period is taken as the atrial fibrillation load within the target time period. The historical atrial fibrillation load from multiple time periods is used as a training set to train the second neural network, which is then used as an atrial fibrillation load prediction model. Atrial fibrillation load prediction model is used to predict atrial fibrillation load in multiple future time periods.

2. The atrial fibrillation burden prediction method according to claim 1, characterized in that, The step of using the first PPG signal of multiple known atrial fibrillation results as a training set to train the first neural network, and then using the trained first neural network as an atrial fibrillation probability prediction model, includes: The first neural network is used to map the first PPG signal to the first atrial fibrillation probability; The total loss of the first neural network is calculated based on the first PPG signal of each known atrial fibrillation result and the first atrial fibrillation probability corresponding to each first PPG signal. The first neural network is trained according to the loss function until it converges, and the trained first neural network is obtained as a model for predicting the probability of atrial fibrillation.

3. The atrial fibrillation load prediction method according to claim 2, characterized in that, The step of calculating the total loss of the first neural network based on the first PPG signal for each known atrial fibrillation result and the first atrial fibrillation probability corresponding to each first PPG signal includes: Calculate the total loss of the first neural network using the following formula: ； Among them, L det α represents the total loss of the first neural network; N is the total number of first PPG signals with known atrial fibrillation results; α represents the weighting coefficient of the first PPG signal corresponding to the occurrence of atrial fibrillation. The probability of atrial fibrillation corresponding to the nth first PPG signal; γ represents the modulation factor, which is a preset constant; y n The result of atrial fibrillation corresponding to the nth first PPG signal; y n =1 indicates atrial fibrillation has occurred; y n =0 indicates that atrial fibrillation did not occur.

4. The atrial fibrillation load prediction method according to claim 1, characterized in that, The step of mapping the atrial fibrillation probability corresponding to each second PPG signal to the atrial fibrillation risk value corresponding to each second PPG signal includes: If the probability of atrial fibrillation corresponding to the second PPG signal is greater than the first probability threshold, the risk value of atrial fibrillation corresponding to the second PPG signal is set to 1. If the probability of atrial fibrillation corresponding to the second PPG signal is less than the second probability threshold, the risk value of atrial fibrillation corresponding to the second PPG signal is determined to be 0. If the probability of atrial fibrillation corresponding to the second PPG signal is less than or equal to the first probability threshold and greater than or equal to the second probability threshold, the probability value of atrial fibrillation corresponding to the second PPG signal shall be used as the risk value of atrial fibrillation corresponding to the second PPG signal.

5. The atrial fibrillation burden prediction method according to claim 1, characterized in that, The step of determining the probability of atrial fibrillation at any time within the measurement period of two adjacent second PPG signals based on the atrial fibrillation risk values ​​corresponding to two adjacent second PPG signals includes: The probability of atrial fibrillation at any time within the measurement period of two adjacent second PPG signals is calculated using the following formula: ； in, R represents the probability of atrial fibrillation at time t; a The risk value of atrial fibrillation corresponding to the earlier measured second PPG signal among two adjacent second PPG signals; w a For R a The weights of R; b The risk value of atrial fibrillation corresponding to the later-measured second PPG signal among two adjacent second PPG signals; w b For R b The weights are: e is the natural constant; λ is a predefined positive constant.

6. The atrial fibrillation burden prediction method according to claim 1, characterized in that, The step of training a second neural network using historical atrial fibrillation load from multiple time periods as a training set, and then using the trained second neural network as an atrial fibrillation load prediction model, includes: The historical atrial fibrillation load from multiple time periods is input into the encoder to encode the historical atrial fibrillation load within each time period into a hidden state; The attention mechanism is used in conjunction with the hidden state of the encoder at each time interval to determine the attention weight of the decoder during decoding at each time interval; The hidden states of the encoder at each time interval are weighted and summed using the attention weights of the decoder during decoding at each time interval to obtain the context vector of the decoder at each time interval. The predicted atrial fibrillation load for the current time period is determined based on the decoder's predicted atrial fibrillation load for the previous time period and the context vector for the current time period. The second neural network is trained by minimizing the mean square error between the predicted and actual values ​​of atrial fibrillation load in each time period until the difference between the predicted and actual values ​​of atrial fibrillation load in each time period is within a preset range. This trained second neural network is then used as an atrial fibrillation load prediction model.

7. An atrial fibrillation load prediction system based on intermittent PPG signals, characterized in that, include: The first training module is used to train the first neural network using the first PPG signals of multiple known atrial fibrillation results as a training set, and to use the trained first neural network as an atrial fibrillation probability prediction model. The acquisition module is used to acquire all valid second PPG signals within the target time period; The first prediction module is used to predict the probability of atrial fibrillation corresponding to each second PPG signal using an atrial fibrillation probability prediction model. The mapping module is used to map the atrial fibrillation probability corresponding to each second PPG signal to the atrial fibrillation risk value corresponding to each second PPG signal. The first determining module is used to determine the probability of atrial fibrillation at any time within the measurement time period of two adjacent second PPG signals based on the atrial fibrillation risk value corresponding to two adjacent second PPG signals. The integration module is used to integrate the atrial fibrillation probability at all times within the target time period to obtain the total atrial fibrillation time within the target time period. The second determining module is used to take the ratio of the total atrial fibrillation time within the target time period to the duration of the target time period as the atrial fibrillation load within the target time period. The second training module is used to train the second neural network with historical atrial fibrillation load from multiple time periods as the training set, and the trained second neural network is used as an atrial fibrillation load prediction model. The second prediction module is used to predict the atrial fibrillation load for multiple future time periods using an atrial fibrillation load prediction model.

8. A computer device, characterized in that, It includes a processor and a memory; wherein, when the processor executes the computer program stored in the memory, it implements the steps of the atrial fibrillation load prediction method based on intermittent PPG signals as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer programs; when executed by a processor, the computer programs implement the steps of the atrial fibrillation load prediction method based on intermittent PPG signals as described in any one of claims 1-6.

10. A computer program product, characterized in that, It includes computer-executable instructions or computer programs, which, when executed by a processor, implement the steps of the atrial fibrillation load prediction method based on intermittent PPG signals as described in claims 1-6.

Images