Condition flow matching and three-flow joint attention continuous arterial blood pressure waveform noninvasive reconstruction method, electronic device and computer readable storage medium
Patent Information
- Application Number
- CN202611252585.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-03
AI Technical Summary
Existing technologies make it difficult to achieve continuous monitoring of arterial blood pressure waveforms in a non-invasive manner. Furthermore, discriminative methods tend to smooth key waveform morphological features, while generative methods have high computational overhead and are coupled with errors in shape estimation and pressure estimation.
A conditional flow matching and three-flow joint attention method is adopted. The arterial blood pressure waveform is generated by using the electrocardiogram signal and photoplethysmography pulse wave signal as conditions. The arterial blood pressure waveform is generated by using velocity field to drive noise. The three-flow joint attention Transformer performs cross-modal information processing to avoid information leakage and error coupling.
It enables efficient generation of arterial blood pressure waveforms under non-invasive conditions, reduces computational overhead, eliminates the coupling of shape estimation and pressure estimation errors, and meets clinical-grade accuracy requirements.
Smart Images
Figure CN123023852A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer science and biomedical engineering, specifically to a non-invasive method for reconstructing continuous arterial blood pressure waveforms using conditional flow matching and three-flow joint attention, an electronic device, and a computer-readable storage medium. Background Technology
[0002] Blood pressure is a major modifiable risk factor for cardiovascular disease. In critical care settings, blood pressure can change significantly within seconds, necessitating continuous monitoring. The gold standard for obtaining continuous arterial blood pressure waveforms is intra-arterial catheterization, where arterial pressure waveforms are recorded frame-by-frame using a pressure sensor. However, this method is invasive and primarily limited to intensive care settings. Prolonged arterial catheterization carries clinical risks including infection, thrombosis, distal limb ischemia, and bleeding, and these risks increase with prolonged catheter placement. Traditional non-invasive cuff blood pressure measurement obtains discrete measurements of systolic and diastolic blood pressure through periodic inflation and deflation, making it difficult to continuously record blood pressure changes over one or more cardiac cycles. It also fails to fully capture waveform morphological features such as the systolic ascending limb, systolic peak, diastolic descending limb, and dicrotic notch.
[0003] In deep learning-based non-invasive arterial blood pressure waveform reconstruction methods, discriminative modeling methods fit the waveform using pointwise regression, easily smoothing key morphological features such as systolic peaks and dicrotic notches. Some unified discriminative frameworks employ a two-stage strategy: first reconstructing a normalized waveform, then rescaling it using separately predicted systolic and diastolic blood pressure. This results in shape estimation errors being coupled with pressure estimation errors. Generative modeling methods often rely on multi-step random sampling processes, involving numerous inference steps and high computational costs. Evaluations tend to focus on blood pressure parameter errors, lacking a comprehensive assessment that simultaneously considers waveform morphology and blood pressure parameters. Summary of the Invention
[0004] To address the shortcomings of existing technologies, the present invention aims to provide a non-invasive reconstruction method, electronic device, and computer-readable storage medium for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow joint attention, thereby solving the problems of discriminative methods easily smoothing key waveform morphological features, two-stage strategies causing coupling of shape and pressure errors, and generative methods having high overhead due to multi-step random sampling inference.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow joint attention, comprising: Receive synchronously acquired and time-aligned electrocardiogram signals and / or photoplethysmography (PPG) signals, and receive arterial blood pressure signals; Using a conditional flow matching framework, with electrocardiogram signals and / or photoplethysmography pulse wave signals as conditions, arterial blood pressure waveforms are generated from Gaussian noise through a velocity field. The velocity field is carried by a three-stream joint attention Transformer, which includes a target stream corresponding to the arterial blood pressure signal, a conditional stream corresponding to the electrocardiogram signal and / or photoplethysmography pulse wave signal, and a missing stream of learnable token occupants. Additive masks in task routing prevent conditional flow attention to noisy target flows at the structural level; The target flow patch is generated from the noise using the trained velocity field. The patch is then stitched together along the time dimension to obtain the arterial blood pressure waveform.
[0006] As a further improvement of the present invention, the three streams are divided into patches according to the same non-overlapping block width and share the same N-token time grid, so that tokens with the same time index correspond to the same time interval in the three modes; cross-stream time phase alignment is achieved by sharing rotation position encoding.
[0007] As a further improvement of the present invention, the conditional flow matching framework employs a modified flow, with a linear probability path and its constant conditional velocity, to regress the parameterized velocity field with the flow matching target.
[0008] As a further improvement of the present invention, in the conditional flow matching framework, noise is added only to the target mode, while the conditional mode is always input with a clean signal.
[0009] As a further improvement of the present invention, the target stream uses a target-specific encoder to embed a noisy patch, the clean conditional stream uses a per-modality conditional encoder to embed a clean patch, and the missing stream uses a learned per-stream token to broadcast to N positions. The three-stream joint attention Transformer is composed of stacked joint attention blocks and single-stream blocks.
[0010] As a further improvement of the present invention, the arterial blood pressure waveform is cropped to [20, 250] mmHg and z-score normalized, and the electrocardiogram signal and photoplethysmography pulse wave signal are re-centered; The arterial blood pressure waveform is obtained through inverse normalization during the inference phase.
[0011] As a further improvement of the present invention, in the conditional flow matching framework, logit-normal sampling is used for the time step, so that the supervision signal is concentrated in the middle section of the path; The output head of the velocity field is initialized with zero.
[0012] As a further improvement of the present invention, during the training process, a generation direction is independently extracted from the task distribution for each sample, so that a single model can cover multiple cross-modal generation directions; The velocity field is pre-trained end-to-end on the source domain mixed dataset. After pre-training, it is fine-tuned at the subject level on the target dataset, and the optimal model parameters are selected on the validation set to obtain the offline model.
[0013] An electronic device includes a processor and a memory, the memory storing a computer program, the processor executing the computer program to implement the steps of the non-invasive reconstruction method for continuous arterial blood pressure waveforms with conditional flow matching and three-flow joint attention as described above.
[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the non-invasive reconstruction method for continuous arterial blood pressure waveforms using conditional flow matching and three-flow joint attention as described above.
[0015] The beneficial effects of the present invention are as follows: the conditional flow matching framework uses electrocardiogram signals and / or photoplethysmography pulse wave signals as conditions, generates arterial blood pressure waveforms by driving noise through velocity field, completes inference with fewer sampling steps, and reduces computational overhead; The three-stream joint attention Transformer prevents conditional stream attention to the noisy target stream at the structural level through task routing additive masking, thus eliminating the interference of target information leakage on waveform reconstruction. The velocity field starts from the noise and is driven by the velocity field to directly generate the arterial blood pressure waveform. It does not rely on systolic and diastolic pressure for rescaling, thus avoiding the error coupling between shape estimation and pressure estimation. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the overall process of the present invention.
[0017] Figure 2 This is the conditional flow matching and task routing graph of the present invention.
[0018] Figure 3 This is a schematic diagram of the three-stream joint attention Transformer structure of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific implementation steps.
[0020] Step 1: Data Preprocessing Assume that ECG, PPG, and ABP signals have been synchronously acquired at the same sampling rate and time-aligned, with the duration of a single sample denoted as L. This invention directly receives these cleaned and aligned signals as input.
[0021] ABP was cropped to a physiologically reasonable range [20, 250] mmHg to remove acquisition artifacts, and then z-score normalization was performed using a fixed dataset-level constant.
[0022] Where μ and σ are the mean and standard deviation estimated at once on the training set. This fixed z-score preserves absolute magnitude information while making the numerical range suitable for neural network training, and the mmHg level can be accurately recovered during the inference stage through inverse normalization. ECG and PPG have been normalized to the [0,1] interval at the upstream dataset level. This invention uses a fixed constant to re-center them to [-1 / 2, 1 / 2] to match the input numerical distribution of the Transformer.
[0023] Since all the above normalization constants are fixed constants at the dataset level and do not depend on any sample-level statistics, even if the target ABP appears at the input, its magnitude information cannot be leaked through the normalization channel.
[0024] Step 2: Model Design Step 2.1: Three-stream partitioning and shared time grid representation Suppose each signal of length L is divided into There are three non-overlapping patches, each with a width of P. The token matrix corresponding to the three streams is denoted as:
[0025] Where s=0, 1, and 2 correspond to ABP, ECG, and PPG, respectively. All three streams share the same N-token time grid, meaning the nth token corresponds to the same time interval in ABP, ECG, and PPG. This design is a prerequisite for subsequent cross-flow time alignment modeling, enabling the present invention to directly model physiological cues such as pulse propagation delay from ECG to PPG under shared rotational position encoding.
[0026] Step 2.2: Conditional Flow Matching and Task Modeling This invention models each ECG / PPG / ABP conversion as a single condition generation task:
[0027] Where m represents the target mode (to be generated), and C represents the set of observed clean conditional modes. To avoid directions with no practical deployment value, such as "using ABP to infer ECG / PPG," this invention stipulates that ABP only appears as the target mode, not as a conditional mode; ECG and PPG can appear as either targets or conditions. This invention trains the following five generation directions, uniformly covered by the same parameter set: ECG+PPG→ABP (flagship direction), ECG→ABP, PPG→ABP, ECG→PPG, and PPG→ECG.
[0028] set up A clean patch sequence for the target modality. For Gaussian noise, this invention employs a modified linear probability path for the flow:
[0029] Its conditional velocity is constant Parameterized velocity field Training by matching targets with streams:
[0030] in It carries clean condition mode and task routing information. Time step t uses logit-normal sampling, i.e. , This concentrates the supervision signal in the middle section of the path, thereby improving training efficiency and generation stability. The target is simulation-independent, and parameter updates can be completed without expanding any ODE during the training phase.
[0031] In terms of task sampling, this invention employs independent sampling per sample: for each sample in a batch, a generation direction is independently extracted from the task distribution π. A preferred value for the task distribution π is: 0.20 for the ECG+PPG→ABP direction, 0.30 for each unimodal →ABP direction, and 0.10 for each ECG↔PPG bridging direction. This value ensures that the flagship direction is adequately trained while covering the other directions; those skilled in the art can adjust this distribution according to actual deployment needs. Independent sampling per sample provides task gradient diversity within a batch, avoiding gradient homogenization caused by batch sampling.
[0032] Step 2.3: Three-stream Joint Attention Transformer velocity field Implemented by a joint attention-diffusion Transformer. Each stream is first mapped to a D-dimensional token sequence via a role-associated patch encoder:
[0033] The target stream uses a target-specific encoder to embed a noisy patch sequence. The clean conditional stream uses a per-modality conditional encoder to embed its clean patch sequence; the missing stream uses a learned per-stream token to broadcast to N locations. A preferred value is D=512.
[0034] The network is stacked with M=12 transformer blocks, of which the first J=8 are joint attention blocks and the last 4 are single-stream blocks. Global conditional vector:
[0035] Where TE represents the sinusoidal time step embedding. This is the mean of all clean conditional stream embeddings. This global conditional vector modulates each transformer block using Adaptive Layer Normalization (AdaLN).
[0036] In the joint attention block, the 3N tokens from the three streams are first subjected to per-stream Q, K, V projections, and then self-attention computation is performed under shared Rotated Position Encoding (RoPE). Since the three streams share the same N-token time grid, and tokens with the same time index share the same phase, the network can directly model cross-stream time alignment cues such as the pulse propagation phase delay from ECG to PPG. This invention introduces an additive task routing mask into the self-attention mechanism. }:
[0037] The mask imposes structural constraints on attention based on the role assignment of task τ, with the following rules: (i) the target flow query can pay attention to its own flow and all clean conditional flows; (ii) the clean conditional flow query can only pay attention to clean conditional flows and cannot pay attention to the noisy target flow; (iii) the missing flow query can only pay attention to itself, and its output is discarded in subsequent calculations; (iv) diagonal elements are always open to ensure that each query row has at least one payable position and avoid softmax row degradation.
[0038] Because the aforementioned mask structurally cuts off the access path of the conditional flow query to the noisy target, and the clean signal of the target mode is never used as a conditional input to the network, target information cannot flow back to its own reconstruction process via the conditional path. In other words, this invention eliminates target leakage at the structural level through masking and role assignment, rather than relying on soft loss terms for penalty.
[0039] After all joint attention blocks, the network collects N tokens from the target stream, refines them further using the remaining single-stream blocks, and then maps each token back to the target stream through the target-specific output header. After splicing, a prediction of the velocity field of the complete target waveform is obtained. The output head and all AdaLN modulations are initialized to zero, making the network approximately a unit flow in the early stages of training. The parameters are approximately 0, thus smoothly deviating from the identity mapping, which is beneficial for optimization stability. All encoders, missing stream tokens, and output heads participate in the computation in each forward propagation (through gating rather than skipping), ensuring that no parameters are unused in data-parallel training.
[0040] Step 3: Model Training and Fine-tuning The BPFlow model was pre-trained end-to-end on a large-scale mixed dataset in the source domain. The optimizer used was AdamW, with a preferred hyperparameter value of β=(0.9, 0.95) and weight decay disabled. The learning rate employed a strategy of maintaining a constant value after linear warm-up and decaying after the validation set loss reached the plateau. The gradient norm pruning threshold was preferably 1.0. An exponential moving average (EMA) was used for the weights, with a preferred decay coefficient of 0.9999, and all evaluations used EMA weights. The above hyperparameters are preferred values for this embodiment; those skilled in the art can adjust them according to the dataset size, memory conditions, and computational budget.
[0041] When memory and computing power allow, a larger global batch size is preferred during the pre-training phase, while a smaller batch size is preferred during the calibration and fine-tuning phase, both using a linear warm-up of approximately 1000 steps. On datasets with limited computing power, the learning rate can be scaled by the square root of the batch size, and the number of warm-up steps can be shortened accordingly.
[0042] Training is performed for a maximum of 1000 epochs, with early stopping executed on the mean absolute error (MAE) of the ABP waveform on the validation set. A fixed random seed is used to ensure reproducible results. During training, a generation direction is independently drawn from the task distribution π for each sample, ensuring task diversity in gradients within the same batch and facilitating simultaneous convergence of the model in five directions.
[0043] After pre-training in the source domain, calibration and fine-tuning are performed on the target dataset using either subject-level or segment-level partitioning. On datasets with available subject identifiers, an 8:1:1 subject-level partition is used, allowing the calibration and test sets to share subjects but not signal segments. On datasets without subject identifiers, segment-level partitioning is used. During the calibration phase, some parameters in the joint attention block and single-stream block can be selectively frozen or updated with a smaller learning rate to mitigate overfitting on small-scale target data. Finally, the optimal model parameters are selected on the validation set to obtain a deployable offline model.
[0044] Step 4: Model Inference The offline model obtained in step 3 is used to perform inference on the test segment of the target dataset. (From Gaussian noise) Starting from this point, perform a T-step integration of the velocity field using the following explicit Euler format:
[0045] One preferred value for T is 16. The obtained patch is concatenated along the time dimension, and the target mode is subjected to the reverse of step 1 (i.e., multiplied by the dataset level σ and added to the dataset level µ) to directly recover the ABP waveform at the millimeter-mercury level and output it.
[0046] A beneficial characteristic of this few-step deterministic sampling is its high insensitivity to the number of sampling steps T: in experiments with T=1, 2, 4, 8, and 16, the MAE difference of the reconstructed waveforms did not exceed 0.05 mmHg, and all met BHS Grade A and satisfied the AAMI numerical error limit. This means that in clinical deployment, a very small T can be selected based on the computational budget to achieve a single forward inference, or a larger T can be selected to further smooth the generated results.
[0047] The same offline model can support five generation directions—ECG+PPG→ABP, ECG→ABP, PPG→ABP, ECG→PPG, and PPG→ECG—by switching task routes τ during the inference phase, without the need to train a separate expert model for each direction. In the flagship direction of ECG+PPG→ABP, this invention achieves waveform MAE of approximately 2.16 mmHg and 2.21 mmHg on the PulseDB and UCI datasets, respectively. The segment-level AAMI errors for SBP, DBP, and MAP all meet the requirements of |ME|≤5 mmHg and σ≤8 mmHg, respectively, and achieve BHS Grade A, which can meet the accuracy requirements of clinical-grade continuous non-invasive blood pressure monitoring.
[0048] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow joint attention, characterized in that, include: Receive synchronously acquired and time-aligned electrocardiogram signals and / or photoplethysmography (PPG) signals, and receive arterial blood pressure signals; Using a conditional flow matching framework, with electrocardiogram signals and / or photoplethysmography pulse wave signals as conditions, arterial blood pressure waveforms are generated from Gaussian noise through a velocity field. The velocity field is carried by a three-stream joint attention Transformer, which includes a target stream corresponding to the arterial blood pressure signal, a conditional stream corresponding to the electrocardiogram signal and / or photoplethysmography pulse wave signal, and a missing stream of learnable token occupants. Additive masks in task routing prevent conditional flow attention to noisy target flows at the structural level; The target flow patch is generated from the noise using the trained velocity field. The patch is then stitched together along the time dimension to obtain the arterial blood pressure waveform.
2. The non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow joint attention according to claim 1, characterized in that, The three streams are divided into patches with the same non-overlapping block width and share the same N-token time grid, so that tokens with the same time index correspond to the same time interval in the three modes; cross-stream time phase alignment is achieved by sharing rotation position encoding.
3. The non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow combined attention according to claim 1, characterized in that, The conditional flow matching framework employs a modified flow, with a linear probability path and its constant conditional velocity, to regress the parameterized velocity field against the flow matching objective.
4. The non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow joint attention according to claim 1, characterized in that, In the conditional flow matching framework, noise is added only to the target mode, while the conditional modes are always input with a clean signal.
5. The non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow combined attention according to claim 1, characterized in that, The target stream uses a target-specific encoder to embed a noisy patch, the clean conditional stream uses a per-modality conditional encoder to embed a clean patch, and the missing stream uses a learned per-stream token to broadcast to N positions. The three-stream joint attention Transformer is composed of stacked joint attention blocks and single-stream blocks.
6. The non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow combined attention according to claim 1, characterized in that, The arterial blood pressure waveform was cropped to [20, 250] mmHg and normalized by z-score. The electrocardiogram signal and photoplethysmography pulse wave signal were re-centered. The arterial blood pressure waveform is obtained through inverse normalization during the inference phase.
7. The non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow combined attention according to claim 1, characterized in that, In the conditional flow matching framework, logit-normal sampling is used at the time step to concentrate the monitoring signal in the middle of the path; The output head of the velocity field is initialized with zero.
8. The non-invasive reconstruction method for continuous arterial blood pressure waveforms based on conditional flow matching and three-flow combined attention according to claim 1, characterized in that, During training, a generation direction is independently extracted from the task distribution for each sample, enabling a single model to cover multiple cross-modal generation directions; The velocity field is pre-trained end-to-end on the source domain mixed dataset. After pre-training, it is fine-tuned at the subject level on the target dataset, and the optimal model parameters are selected on the validation set to obtain the offline model.
9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the non-invasive reconstruction method of continuous arterial blood pressure waveforms with conditional flow matching and three-flow joint attention as claimed in any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the non-invasive reconstruction method of continuous arterial blood pressure waveforms based on conditional flow matching and three-flow joint attention as claimed in any one of claims 1 to 8.