Self-adaptive parameter regulation and control system for ECMO pulsating perfusion system

Abstract

The invention relates to the technical field of intensive care medicine adjuvant therapy, in particular to a self-adaptive parameter regulation and control system for an ECMO pulsating perfusion system. The system comprises a multi-modal signal reconstruction module, a physiological state analysis module, a heart and brain signal interaction module, a pulsation feature extraction module, a physiological demand analysis module, a collaborative perception feature extraction module, a comprehensive decision confidence analysis module, a dynamic regulation and control baseline generation module and a regulation and control instruction generation module. By means of multi-dimensional feature collaborative perception and accurate decision making, parameter adjustment deviation is reduced, heart and brain perfusion safety is guaranteed, autonomic nerve balance is maintained, clinical complication risks are reduced, a scientific, efficient and safe intelligent regulation and control scheme is provided for ECMO pulsating perfusion, and the treatment effect of critical patients is promoted.

Description

Technical Field

[0001] This invention relates to the field of critical care medicine auxiliary treatment technology, specifically to an adaptive parameter control system for ECMO pulsatile perfusion systems. Background Technology

[0002] ECMO (Extracorporeal Membrane Oxygenation), as a core technology for life support in critically ill patients, is widely used in the treatment of severe respiratory failure, heart failure, and other critical illnesses. The precise control of its pulsatile perfusion parameters (including blood flow, pulsation frequency, pulsation amplitude, centrifugal pump speed, and oxygen concentration) directly determines the perfusion effect on vital organs such as the heart and brain, thus profoundly affecting the patient's treatment prognosis, complication rate, and long-term survival rate. Currently, in clinical practice, the control of ECMO pulsatile perfusion parameters still relies primarily on the clinical experience and judgment of medical staff, lacking systematic and scientific quantitative decision-making support. This inevitably leads to prominent problems such as strong subjectivity, lag in control, and poor individual adaptability. Differences in the clinical experience and judgment standards of different medical staff can easily lead to parameter adjustment deviations. Furthermore, medical staff need to manually monitor multimodal physiological signals and assess the patient's condition before adjusting parameters, making it difficult to keep up with the dynamic changes in the patient's physiological state, resulting in significant control lag.

[0003] The core flaw of existing control methods is their inability to accurately capture multi-dimensional physiological dynamic changes in patients' cardiovascular and cerebrovascular states, autonomic nervous system balance, and other aspects. This lack of synergistic analysis of multi-source information leads to parameter adjustment biases: for example, excessively low pulsation amplitude can cause insufficient cerebral perfusion and brain tissue hypoxia and necrosis, while excessively high pulsation amplitude may increase myocardial oxygen consumption and induce complications such as arrhythmias. These problems significantly increase the treatment risks for critically ill patients and can even endanger their lives. Simultaneously, traditional control models lack systematic multi-source feature fusion mechanisms and scientific decision-making algorithms. They cannot effectively combine patients' real-time physiological states, individual baseline characteristics (age, weight, underlying diseases, etc.), and historical control experience, making it difficult to achieve real-time dynamic optimization of perfusion parameters. Furthermore, they cannot adapt to the individual physiological differences of different patients—for example, there are significant differences in vascular elasticity and cerebral perfusion regulation ability between elderly and younger patients, and the physiological needs of heart failure patients and respiratory failure patients are drastically different. Traditional, uniform control models cannot meet the individualized treatment needs of various patients, severely restricting further improvements in ECMO treatment efficacy and increasing the workload and decision-making pressure on medical staff. Therefore, the field of critical care medicine urgently needs an intelligent ECMO pulsatile perfusion method that can achieve multi-dimensional collaborative perception, precise decision-making, and real-time control, effectively addressing many shortcomings of existing technologies and providing safer and more efficient life support for critically ill patients. Summary of the Invention

[0004] To address the shortcomings of existing methods and the needs of practical applications, and in order to solve the aforementioned problems, this invention provides an adaptive parameter control system for ECMO pulsation perfusion systems, comprising: The system comprises the following modules: a multimodal signal reconstruction module for cleaning and reconstructing multimodal signals and extracting physiological temporal features; a physiological state analysis module for extracting the static state and dynamic trends of the cardiovascular system from the physiological temporal features, obtaining a cardiovascular system state feature set; a cardiovascular-brain signal interaction module for generating time-varying autonomic balance features based on the information transmission strength between cardiovascular and brain signals; a pulsation feature extraction module for obtaining pulsation features using physical information neural networks and filter neural network accumulation algorithms; and a physiological demand analysis module for identifying the causal relationship between cardiovascular and cerebrovascular states and perfusion parameters, quantifying the influence of different parameters on individual states, and generating individual... The system comprises the following modules: an integrated physiological demand vector; a collaborative perception feature extraction module, which integrates cardiovascular and cerebrovascular state feature sets, autonomic nervous balance features, and an individualized physiological demand vector to obtain collaborative perception features; a comprehensive decision confidence analysis module, which fuses autonomic nervous balance features and pulsation features through DS evidence theory to obtain comprehensive decision confidence; a dynamic regulation baseline generation module, which generates a dynamic regulation baseline based on historical regulation experience data through multi-agent deep reinforcement learning and knowledge distillation; and a regulation instruction generation module, which generates the final pulsation parameter instruction by fusing collaborative perception features, comprehensive decision confidence, and dynamic regulation baseline through a graph neural network.

[0005] Optionally, the cleaning and reconstruction of the multimodal signal to extract physiological temporal features includes the following steps: Acquire multimodal signals, including arterial blood pressure, heart rate, cerebral oxygen saturation, blood flow, and respiratory rate of ECMO patients; denoise the multimodal signals using an adaptive wavelet threshold, reconstruct the denoised physiological signals using a filter neural network; and extract physiological temporal feature matrices using the physiological signals.

[0006] Optionally, the step of extracting the static state and dynamic trend of the cardiovascular system from physiological time-series features to obtain a cardiovascular system state feature set includes the following steps: A temporal convolutional network is used to extract local state features of the cardiovascular and cerebrovascular systems from physiological temporal features. A bidirectional gated recurrent unit network is introduced, and based on the local state features, a bidirectional temporal feature vector is generated through positive and negative synergistic effects. The local state features and the bidirectional temporal feature vector are integrated to obtain a set of cardiovascular and cerebrovascular state features.

[0007] Optionally, generating time-varying autonomic balance characteristics based on the information transmission intensity between heart and brain signals includes the following steps: Based on information theory, the intensity of bidirectional information transmission between the heart and brain is quantified to obtain a dynamic temporal curve of heart-brain information transmission; a heart-brain coupling network is constructed, and the dynamic curve of heart-brain coupling intensity is analyzed through the heart-brain coupling network; combining the dynamic temporal curve and the dynamic curve of heart-brain coupling intensity, the autonomic nervous balance characteristics that change dynamically over time are generated.

[0008] Optionally, obtaining pulsation features using a physical information neural network and a filter neural network accumulation algorithm includes the following steps: A physical information neural network is constructed, and the single-cycle pulsation features are obtained by combining the physical information neural network with the arterial blood pressure wave. A filter neural network accumulation algorithm is introduced to capture the temporal accumulation effect of the single-cycle pulsation features. The pulsation feature set is obtained by combining the single-cycle pulsation features and the accumulated pulsation features.

[0009] Optionally, the step of identifying the causal relationship between cardiovascular and cerebrovascular status and perfusion parameters, quantifying the influence of different parameters on individual status, and generating an individualized physiological demand vector includes the following steps: A feature-perfusion parameter association dataset is constructed using a cardiovascular and cerebrovascular state feature set. A causal discovery algorithm is used to identify the direct causal relationship between perfusion parameters and state. A parameter sensitivity network is constructed with the direct causal relationship between perfusion parameters and state as edges to quantify the sensitivity coefficient of perfusion parameters to cardiovascular and cerebrovascular state. An individualized physiological demand vector is generated by combining the sensitivity coefficient and individual baseline features.

[0010] Optionally, the process of integrating the cardiovascular and cerebrovascular state feature set, autonomic nervous balance features, and individualized physiological demand vectors to obtain collaborative perception features includes the following steps: A dynamic routing mechanism is constructed; based on the dynamic routing mechanism, the cardiovascular and cerebrovascular state feature set, autonomic nervous balance features, and individualized physiological demand vectors are integrated to obtain collaborative perception features.

[0011] Optionally, the step of fusing autonomic balance features and pulsation features through DS evidence theory to obtain a comprehensive decision confidence level includes the following steps: An evidence body is constructed using autonomic balance and pulsation characteristics; based on rule- and data-driven BPA allocation principles, a comprehensive decision confidence level is obtained using the evidence body.

[0012] Optionally, the process of generating a dynamic regulatory baseline based on historical regulatory experience data through multi-agent deep reinforcement learning and knowledge distillation includes the following steps: Multiple collaborative agents are constructed, and a state space, action space, and reward function are defined. An offline MARL model is generated by training with a historical regulatory experience dataset. The decision knowledge of the offline MARL model is transferred to a lightweight online student model through knowledge distillation. Based on the knowledge-distilled lightweight online student model and combined with the patient's real-time physiological state, a dynamic regulatory baseline is generated.

[0013] Optionally, the step of generating the final pulsation parameter command by fusing collaborative sensing features, comprehensive decision confidence, and dynamic control baseline through a graph neural network includes the following steps: By combining collaborative perception features, comprehensive decision confidence, and dynamic control baseline, a dynamic graph structure network is constructed. Based on the global feature vector generated by the dynamic graph structure network, preliminary parameter instructions are generated through fully connected layer inference. The preliminary parameter instructions are corrected according to ECMO clinical safety constraints to generate final pulsation parameter instructions.

[0014] This invention, guided by the clinical needs of ECMO pulsatile perfusion, first cleanses and reconstructs multimodal physiological signals, extracting three core features: cardiovascular and cerebrovascular status, autonomic nervous system balance, and pulsatile characteristics. It then identifies the direct correlation between parameters and status through causal discovery, and fuses multi-source features to generate collaboratively perceived features. Using DS evidence theory, it quantifies decision confidence, and constructs a lightweight model through MARL training and knowledge distillation to generate a dynamic control baseline. Finally, it leverages GNN to achieve correlation reasoning between multiple features and parameters, generating individualized parameter instructions constrained by clinical safety. This effectively overcomes the limitations of ECMO control, such as strong subjectivity and significant lag, achieving multi-dimensional feature collaborative perception and precise decision-making, balancing real-time performance and individualization, reducing parameter adjustment bias, ensuring cardiovascular and cerebrovascular perfusion safety, maintaining autonomic nervous system balance, and reducing the risk of clinical complications. It provides a scientific, efficient, and safe intelligent control solution for ECMO pulsatile perfusion, helping to improve the treatment outcomes of critically ill patients. Attached Figure Description

[0015] Figure 1 This is a framework diagram of an adaptive parameter control system for an ECMO pulsatile perfusion system provided in an embodiment of the present invention. Detailed Implementation

[0016] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0017] Throughout this specification, references to an embodiment, example, or illustration mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, phrases appearing in various places throughout the specification, such as "in one embodiment," "in an embodiment," "an example," or "an illustration," do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please see Figure 1 To address the above problems, this invention provides an adaptive parameter control system for ECMO pulsatile perfusion systems, such as... Figure 1 As shown, it includes: It includes a multimodal signal reconstruction module 1, a physiological state analysis module 2, a heart-brain signal interaction module 3, a pulsation feature extraction module 4, a physiological demand analysis module 5, a collaborative perception feature extraction module 6, a comprehensive decision confidence analysis module 7, a dynamic regulation baseline generation module 8, and a regulation command generation module 9.

[0019] The multimodal signal reconstruction module 1 is used to clean and reconstruct the multimodal signal and extract physiological temporal features, including the following steps: S11. Acquire multimodal signals, including arterial blood pressure, heart rate, cerebral oxygen saturation, blood flow, and respiratory rate of the ECMO patient.

[0020] Raw time-series signals of arterial blood pressure (ABP), heart rate (HR), cerebral oxygen saturation (rSO2), blood flow (FBF), and respiratory rate (RR) were collected from ECMO patients. All signals came from the sensors built into the ECMO device and the clinical monitor to ensure the synchronicity and accuracy of signal acquisition.

[0021] To eliminate feature extraction bias caused by differences in sampling frequencies of different devices, all original signals were standardized to a sampling frequency of 250Hz to ensure consistency of timing features.

[0022] Based on this, considering the differences in physiological baseline among different patients, the average signal value during the first 5 minutes of stable operation of ECMO was used as the benchmark to eliminate zero-point drift of the equipment and individual baseline deviation. Then, a combination of linear interpolation and neighborhood mean correction was used to supplement the small number of missing values ​​caused by brief equipment failures and patient movements during signal acquisition.

[0023] S12. The multimodal signal is denoised using an adaptive wavelet threshold, and the denoised physiological signal is reconstructed using a filter neural network.

[0024] To address the characteristics of ECMO signals (such as the systolic / diastolic characteristic frequencies of the ABP waveform, the periodic fluctuations of HR, and the slow variation trend of rSO2), the db6 wavelet was selected as the basis function. This wavelet basis possesses excellent time-domain localization characteristics and frequency-domain resolution, accurately adapting to the non-stationarity of ECMO physiological signals and effectively distinguishing the frequency differences between physiological signals and noise, thus avoiding the distortion of physiological characteristics caused by using a general wavelet basis. Furthermore, based on the periodic characteristics of different signals, different wavelet decomposition levels were set (6 levels for ABP signals, 5 levels for HR signals, and 4 levels for rSO2 signals) to ensure that the coefficients of each level accurately correspond to the signal components in different frequency bands after decomposition.

[0025] Furthermore, using 100 sampling points as a local window, the signal-to-noise ratio (SNR) within each window is calculated. Then, based on the threshold corresponding to the window, the high-frequency coefficients after wavelet decomposition are subjected to threshold shrinkage. During the shrinkage process, a soft threshold shrinkage method is adopted to avoid signal abrupt changes caused by hard threshold shrinkage, while preserving high-frequency components related to physiological activities (such as minute fluctuations in the pulse wave, transient variations in heart rate, and rapid fluctuations in rSO2). This precisely eliminates high-frequency noise generated by the ECMO device (such as 10-50Hz noise from the centrifugal pump), electromagnetic interference from the power line (50Hz power frequency interference), and high-frequency noise caused by patient movement, achieving a balance between noise removal and preservation of physiological characteristics.

[0026] After threshold shrinkage, inverse wavelet transform is performed on the wavelet coefficients (high-frequency coefficients + low-frequency coefficients) of each layer to reconstruct the decomposed signal components into a complete time-series signal, i.e., the preliminary denoised signal. Simultaneously, the preliminary denoised signal is initially verified by calculating the signal-to-noise ratio (SNR) improvement before and after denoising. If the SNR improvement is lower than a preset threshold (e.g., less than 3dB), the process is repeated by adjusting the wavelet decomposition layer and the local window size, and recalculating the threshold and shrinking process until the denoising effect requirements are met. This provides a high-quality preliminary denoised signal foundation for subsequent filter neural network reconstruction.

[0027] After adaptive wavelet threshold denoising, although most noise has been removed, two problems may still exist: First, during wavelet denoising, some weak physiological features with frequencies close to the noise (such as slow fluctuations in rSO2 and low-frequency components of HRV) may be mistakenly removed; second, the frequency domain characteristics of the initially denoised signal still have slight distortions and cannot fully match the frequency domain distribution of clean physiological signals in ECMO clinical scenarios. Therefore, a lightweight one-dimensional filter neural network is constructed to form a synergistic mechanism of time-domain denoising and frequency-domain reconstruction with wavelet denoising, achieving further optimization and repair of the signal. The specific operation is as follows: A lightweight one-dimensional filter neural network was constructed. The network adopts a simplified structure of input layer-convolutional layer-pooling layer-fully connected layer-output layer. The input is the preliminary denoised signal obtained by wavelet denoising. The convolutional layer uses multiple one-dimensional convolutional kernels of different sizes. The kernel size is strictly adapted to the periodic characteristics of various physiological signals of ECMO. For example, the cardiac cycle of ABP signal is usually 0.8-1.2s, corresponding to 200-300 sampling points (250Hz×0.8s=200, 250Hz×1.2s=300). Therefore, three kernel sizes of 200, 250, and 300 are set to capture the ABP waveform characteristics of different cycles. The HR signal cycle is 0.5-2s, corresponding to a convolutional kernel size of 125-500. The rSO2 signal fluctuates slowly (cycle >10s), corresponding to a convolutional kernel size of 2500 or more. The multi-size convolutional kernel achieves accurate adaptation to different signals. Meanwhile, the network introduces a batch normalization layer to accelerate training convergence, avoid gradient vanishing, and ensure that the network can stably learn the frequency domain characteristics of ECMO signals.

[0028] The network training objective and reconstruction process: The core training objective of the network is to learn the frequency domain characteristics of clean signals in ECMO clinical scenarios. These clean signals are derived from a standard signal library used in ECMO extracorporeal circulation experiments and signals from patients with stable conditions and no significant noise interference in clinical settings (confirmed by clinicians). During training, the network parameters are optimized through backpropagation using initially denoised signals as input and clean signals as labels. This allows the network to accurately identify and retain the characteristic frequency bands of various physiological signals—for example, the characteristic frequency band of HR signals is 0.5-2.5Hz (corresponding to a heart rate of 30-150 beats / minute, covering the heart rate range of ECMO patients), the characteristic frequency band of rSO2 signals is 0.01-0.1Hz (corresponding to slow fluctuations in cerebral oxygen saturation, reflecting dynamic changes in cerebral perfusion), and the characteristic frequency band of ABP signals is 0.5-5Hz (corresponding to cardiac cycle and pulse wave conduction characteristics).

[0029] The network further removes residual weak noise (such as harmonic components of power frequency interference) from the initially denoised signal through frequency domain filtering. Simultaneously, it reconstructs weak physiological features mistakenly removed by wavelet denoising in the time domain, such as restoring slow fluctuations in rSO2 and low-frequency components of HRV, ensuring that the reconstructed signal is free of noise interference and fully preserves all physiological features related to ECMO perfusion. After training, the effectiveness of the reconstructed signal is verified by calculating the similarity between the reconstructed signal and the clean signal. A similarity ≥95% is considered satisfactory; otherwise, the convolutional kernel size and network parameters are adjusted, and retraining is performed to obtain high-quality, highly complete physiological signals.

[0030] S13. Extract the physiological time-series feature matrix using the physiological signals.

[0031] Based on the physiological signals reconstructed by the filter neural network, a systematic temporal feature extraction is performed to ensure that the extracted features can comprehensively and accurately reflect the physiological state of ECMO patients, providing core input for subsequent steps such as cardiovascular and cerebrovascular status analysis and autonomic nervous system balance assessment.

[0032] The extraction process follows the principles of signal type classification and feature standardization, and the specific operations are as follows: First, the reconstructed multimodal signals were split according to signal type, and features were extracted from ABP, HR, rSO2, FBF, and RR signals respectively. For ABP signals, the core features related to pulsation and perfusion were extracted, including systolic blood pressure (peak value of each cardiac cycle), diastolic blood pressure (trough value of each cardiac cycle), pulse pressure difference (difference between systolic and diastolic blood pressure), pulse wave rise slope (reflecting myocardial contractility and perfusion pressure), and diastolic fall rate (reflecting vascular resistance). For HR signals, instantaneous heart rate value, heart rate variability (including time-domain indices SDNN and RMSSD, reflecting autonomic nerve activity), and heart rate fluctuation amplitude were extracted. For rSO2 signals, the mean, slope (reflecting the trend of brain oxygenation changes), and coefficient of variation (reflecting brain perfusion stability) were extracted every 5 seconds. For FBF and RR signals, basic temporal features such as instantaneous value, mean, and variability were extracted.

[0033] After extraction, all features are standardized (using Z-score standardization to eliminate dimensional differences between different features). All features are arranged by time step to form a standardized physiological time-series feature matrix. The matrix dimension is time step × number of features, where the time step is divided into 2s and the number of features covers all extracted physiological features, ensuring that the matrix can comprehensively and orderly present the dynamic changes of the patient's physiological state.

[0034] The physiological state analysis module 2 is used to extract the static state and dynamic trend of the cardiovascular system from physiological time-series features to obtain a cardiovascular system state feature set, including the following steps: S21. Use temporal convolutional networks to extract local state features of the cardiovascular and cerebrovascular systems from physiological temporal features.

[0035] During ECMO perfusion, changes in the cardiovascular and cerebrovascular status exhibit both local transient features (such as morphological abnormalities in a single ABP waveform) and long-term temporal dependencies (such as insufficient cerebral perfusion caused by prolonged hypotension). This invention constructs a multi-layer TCN network and utilizes the properties of dilated convolution to simultaneously achieve accurate extraction of both local features and long-term temporal dependencies. The specific operation is as follows: First, a multi-layer TCN network (temporal convolutional network) was constructed with 4 layers. Each layer adopted a dilated convolutional structure, and the dilation coefficient was adapted and optimized according to the ECMO perfusion cycle, set to 1, 2, 4, and 8 respectively, corresponding to the temporal dependence range of 10s, 20s, 40s, and 80s. This range not only covers the routine adjustment cycle of ECMO parameters (5-10s) but also the typical duration of significant changes in cardiovascular and cerebrovascular status (such as insufficient cerebral perfusion, which often occurs when hypotension lasts for more than 30s), and can comprehensively capture the temporal correlations in different duration ranges.

[0036] The convolutional layers employ one-dimensional convolutional kernels with a kernel size of 3 to adapt to the temporal continuity of physiological signals. They focus on capturing local features, such as the systolic peak, diastolic trough, and dicrotic wave morphology of a single ABP waveform, as well as local temporal features like instantaneous HR fluctuations and short-term changes in rSO2. Through the hole effect of dilated convolution, the receptive field is expanded without increasing network parameters or computational load, capturing long-range temporal dependencies. Examples include the correlation between a sustained increase in HR and a sustained decrease in rSO2 within 5 minutes, and the long-term correlation between ABP fluctuation amplitude and FBF changes, achieving simultaneous extraction of local and long-range dependent features. Simultaneously, residual connections are introduced into the network to avoid the gradient vanishing problem during deep network training, ensuring the network can stably learn the temporal characteristics of ECMO cardiovascular signals.

[0037] Furthermore, the TCN network outputs local cardiovascular and cerebrovascular state features. This feature set focuses on instantaneous and short-term state descriptions, covering two core aspects: First, local morphological abnormalities of physiological signals, such as abnormal current ABP waveform (excessively high / low systolic peak, disappearance of dicrotic waves), instantaneous HR fluctuations exceeding the normal range (>150 beats / min or <60 beats / min), and a sudden drop in rSO2 (a decrease of more than 5% within 5 seconds); Second, local temporal correlation features, such as increased HR variability (indicating abnormal autonomic nerve activity), narrowed ABP pulse pressure (indicating abnormal vascular resistance), and synchronous decrease in FBF and ABP (indicating insufficient perfusion pressure). These local state features can accurately reflect the current instantaneous state of the cardiovascular and cerebrovascular systems.

[0038] S22. Introduce a bidirectional gated recurrent unit network, and generate a bidirectional temporal feature vector based on the local state features through positive and negative synergistic effects.

[0039] The local state features extracted by TCN can only reflect instantaneous and short-term states and dependencies, and cannot capture the influence of past states on the current state or the prediction of future trends by the current state. However, the regulation of cardiovascular and cerebrovascular states in ECMO requires combining historical trends and future predictions. Therefore, a bidirectional gated recurrent unit network (BiGRU network) is introduced. Through the synergistic effect of forward and backward GRUs, bidirectional temporal correlations are modeled. The specific operation is as follows: The local state features output by TCN are used as the input to the BiGRU network. BiGRU consists of two independent GRU units: forward GRU and backward GRU. The two units operate in parallel and output together.

[0040] The forward GRU processes the sample sequence in chronological order (past → present), focusing on learning the influence trend of historical states on the current state. For example, whether the continuous decrease of rSO2 in the first 30 seconds is accompanied by the continuous decrease of ABP, and whether the increase of HR in the first 10 seconds is related to FBF adjustment. By capturing historical time series correlations, the causes of the current cardiovascular and cerebrovascular state can be determined (such as whether insufficient cerebral perfusion is caused by previous hypotension).

[0041] The inverse GRU processes the sample sequence in reverse chronological order (current → future), focusing on learning how the current state predicts future trends. For example, if the current HR continues to rise and rSO2 begins to decline, does this indicate that severe cerebral insufficiency will occur later? If the current ABP pulse pressure is continuously narrowing, does this indicate that vascular resistance will further increase? By predicting future trends, it provides early warning for subsequent parameter regulation.

[0042] The outputs of the two GRU units are spliced ​​and fused to form a feature vector containing bidirectional temporal correlations, which retains both historical correlation information and future prediction information, comprehensively depicting the dynamic changes in cardiovascular and cerebrovascular status.

[0043] In this embodiment, to enhance the specificity and effectiveness of the features, an attention mechanism is introduced to weight and optimize the bidirectional temporal features processed by BiGRU, highlighting key time-step features related to ECMO perfusion regulation. Based on actual ECMO clinical operations, attention weight allocation rules are set: within 10 seconds of adjusting ECMO flow rate, pulse rate, and other parameters, a significant response in cardiovascular and cerebrovascular status will occur; therefore, the feature weight for this time period is set to 0.8-0.9, higher than other time periods (0.1-0.7). Simultaneously, time steps with abnormal cardiovascular and cerebrovascular status (such as rSO2 < 50%, ABP < 60 mmHg) are also assigned higher weights to ensure that these key status features are given priority. Through the weighting effect of the attention mechanism, the most valuable temporal features for judging cardiovascular and cerebrovascular status are selected, while irrelevant or secondary features are eliminated, improving the accuracy of subsequent feature fusion and decision-making.

[0044] S23. Integrate the local state features and the bidirectional temporal feature vector to obtain the cardiovascular and cerebrovascular state feature set.

[0045] First, the local state features of TCN are concatenated with the bidirectional temporal correlation features of BiGRU to form a fusion feature matrix, which covers multi-dimensional information such as the instantaneous state, local correlation, historical trends and future predictions of cardiovascular and cerebrovascular diseases.

[0046] Subsequently, a fully connected layer is introduced for dimensionality reduction. The fully connected layer has two hidden layers. The ReLU activation function is used to remove redundant features and compress the feature dimension, while retaining the core feature information, thus avoiding feature distortion during the dimensionality reduction process.

[0047] Finally, combining ECMO clinical knowledge, the reduced-dimensional features were labeled and quantified to generate a cardiovascular and cerebrovascular status feature set containing three major categories of core information: first, cardiac function status (such as estimated myocardial oxygen consumption, estimated cardiac output, and myocardial contractility evaluation indicators); second, cerebrovascular status (such as estimated cerebral perfusion pressure, estimated cerebral blood flow velocity, and cerebral oxygen supply and demand balance evaluation indicators); and third, perfusion risk level (combined with cardiovascular and cerebrovascular status indicators, classified into three risk levels of low, medium, and high according to clinical standards, such as rSO2 < 50% and cerebral perfusion pressure < 50 mmHg as high risk).

[0048] The heart-brain signal interaction module 3 is used to generate autonomic nervous balance characteristics that change dynamically over time based on the information transmission intensity between heart and brain signals, including the following steps: S31. Based on information theory, quantify the intensity of bidirectional information transmission between the heart and brain to obtain the dynamic time-series curve of information transmission between the heart and brain.

[0049] From physiological temporal characteristics, core signals closely related to autonomic nervous system regulation are extracted and categorized according to the cardiac-brain side, providing targeted input for subsequent information transmission quantification and coupling strength analysis. The specific operation is as follows: The core signals on the cardiac side focus on the autonomic nervous system's regulation and feedback of cardiac function. HR variability (HRV) and systolic blood pressure variability (SBPV) are selected as core indicators. HRV directly reflects the autonomic nervous system's ability to regulate heart rate. Its low-frequency component (LF) corresponds to sympathetic nerve activity, and its high-frequency component (HF) corresponds to parasympathetic nerve activity, making it a classic indicator for assessing autonomic nervous system balance. SBPV reflects the autonomic nervous system's regulation of vasomotor function and is directly related to the sympathetic nervous system's vasomotor effect, thus supplementing the shortcomings of HRV in vascular regulation.

[0050] The core signals from the cerebral side focus on the feedback of heart-brain interaction under the regulation of the autonomic nervous system. rSO2 variability and the low-frequency / high-frequency ratio of EEG are selected as core indicators. rSO2 variability reflects the dynamic changes in cerebral perfusion, and abnormal autonomic nervous system balance directly affects the stability of cerebral blood flow perfusion, leading to abnormal rSO2 variability. The low-frequency / high-frequency ratio of EEG (usually referring to the power ratio of the 0.5-4Hz low-frequency band to the 8-13Hz high-frequency band) directly reflects the neural activity state of the cerebral cortex and is highly correlated with the excitability of the autonomic nervous system, indirectly demonstrating the regulatory role of the autonomic nervous system on brain function. After segmentation, the two types of signals undergo standardized preprocessing to ensure uniform signal dimensions.

[0051] Furthermore, by employing the transfer entropy method from information theory, the intensity of bidirectional information transmission between the heart and brain, and between the brain and heart, is quantified to accurately capture the dynamic regulatory relationship between the heart and brain.

[0052] Transfer entropy can effectively quantify the unidirectional information transmission between two time-series signals. For the unique clinical scenario of ECMO patients—rapid fluctuations in patient condition, short physiological signal durations (mostly short-sequence data from real-time monitoring), and significant signal interference from ECMO equipment noise (such as signal interference caused by centrifugal pump operation and pipeline vibration)—the traditional transfer entropy calculation method is optimized for this scenario: First, the time window is adjusted, shortening the common 60s calculation window to 20s to adapt to the rapid changes in the physiological state of ECMO patients, ensuring timely capture of dynamic changes in cardiac-brain information transmission and avoiding information lag caused by excessively long windows. Second, a partial transfer entropy method is introduced, using noise signals generated by ECMO equipment operation (such as equipment interference components in blood flow FBF and sensor electromagnetic interference signals) as conditional variables to eliminate their interference in quantifying cardiac-brain signal information transmission, ensuring that the transfer entropy calculation results truly reflect the physiological interaction between the heart and brain, rather than spurious correlations caused by equipment interference.

[0053] Then, the transfer entropy value is calculated for each 5-second time window to obtain a dynamic time-series curve of heart-brain information transmission. Combined with the standard ECMO parameter adjustment cycle (5-10 seconds), a 5-second calculation step size for transfer entropy is set, with a 20-second calculation window. A sliding window approach is used to continuously calculate the transfer entropy values ​​between the heart and brain, and between the brain and heart, within each time window. The heart-brain transfer entropy value reflects the intensity of the heart's signal regulation of the brain signal, while the brain-heart transfer entropy value reflects the intensity of the brain's feedback to the heart signal. The transfer entropy values ​​of all time windows are arranged chronologically to form a dynamic time-series curve of heart-brain information transmission. The fluctuation trend of the curve directly reflects the dynamic changes in the bidirectional interaction between the heart and brain—for example, when the sympathetic nervous system is dominant, the heart's regulation of the brain is enhanced, and the heart-brain transfer entropy value increases significantly; when the parasympathetic nervous system is dominant, the brain's feedback to the heart is enhanced, and the brain-heart transfer entropy value increases significantly.

[0054] S32. Construct a heart-brain coupling network and analyze the dynamic curve of heart-brain coupling strength through the heart-brain coupling network.

[0055] Information transmission between the heart and brain involves not only unidirectional regulation and feedback, but also complex coupling relationships among multiple indicators. The dynamic changes in this coupling strength reflect the overall state of autonomic nervous system balance. Based on network physiology, this study constructs a complex network to quantify the coupling patterns and overall correlation strength of various indicators within the system. Combining characteristic indicators of core heart-brain signals, a heart-brain coupling network is constructed, and its global coupling strength is analyzed. The specific operation is as follows: A cardio-brain coupling network was constructed, with nodes defined as key characteristic indicators of cardio-brain signals. Specifically, these included: the high-frequency component (HF, reflecting parasympathetic activity) and low-frequency component (LF, reflecting sympathetic activity) of HRV on the cardiac side, and the mean of SBPV (reflecting the overall state of vasoconstriction); and the slope of rSO2 on the cerebral side (reflecting the trend of cerebral oxygenation changes, related to cerebral blood perfusion), the coefficient of variation of rSO2 (reflecting cerebral perfusion stability), and the low-frequency / high-frequency ratio of EEG (reflecting the state of cerebral neural activity). Edge weights were defined as the phase synchronization between two nodes (indicators), calculated using Phase-Locking Value (PLV). PLV accurately quantifies the degree of phase synchronization between two time-series signals, ranging from 0 to 1. A PLV value closer to 1 indicates stronger phase synchronization and higher coupling between the two indicators; a PLV value closer to 0 indicates lower coupling. The PLV values ​​between all nodes are calculated and used as the weights of the network edges to complete the construction of the heart-brain coupling network. This network can intuitively present the coupling relationship between various indicators of the heart and brain, as well as the correlation between this relationship and the activity of the autonomic nervous system.

[0056] Furthermore, the global coupling strength of the network (such as the average clustering coefficient and feature path length) is calculated to reflect the overall coupling state of the heart-brain system. Global coupling strength is a core indicator for measuring the overall interconnectedness of the heart-brain coupled network, directly reflecting the heart-brain synergy under autonomic nervous system equilibrium. The average clustering coefficient and feature path length are selected as the core evaluation indicators. The average clustering coefficient reflects the degree of clustering of nodes in the network, that is, the tightness of coupling between each node and its neighboring nodes. The higher the average clustering coefficient, the tighter the local coupling between various heart-brain indicators, and the stronger the heart-brain synergy. The feature path length reflects the average shortest path between any two nodes in the network. The shorter the feature path length, the higher the information transmission efficiency between various heart-brain indicators, and the stronger the overall coupling.

[0057] In the context of ECMO clinical scenarios, a normal range for coupling strength was defined (based on network parameter statistics from clinically stable patients). If the calculated average clustering coefficient is lower than the normal range and the feature path length is higher than the normal range, it indicates a decrease in heart-brain coupling strength and potential abnormalities in autonomic nervous system balance (such as excessive sympathetic nerve excitation leading to a decline in heart-brain coordination). Simultaneously, the global coupling strength of the heart-brain coupling network was dynamically calculated for each time window with a 5-second time step, resulting in a dynamic curve of heart-brain coupling strength, comprehensively characterizing the dynamic features of heart-brain interaction.

[0058] S33. Combining the dynamic time-series curve and the dynamic curve of heart-brain coupling intensity, generate autonomic nervous balance characteristics that change dynamically over time.

[0059] First, based on the directional strength of transfer entropy, the activity of the sympathetic and parasympathetic nervous systems was quantified: when the sympathetic nervous system was dominant, the heart's regulatory effect on the brain was enhanced, the transfer entropy value from heart to brain (TEc2b) was significantly higher than that from brain to heart (TEb2c), and the low-frequency component (LF) of HRV increased while the high-frequency component (HF) decreased; when the parasympathetic nervous system was dominant, the brain's feedback effect on the heart was enhanced, TEb2c was significantly higher than TEc2b, and HF increased while LF decreased.

[0060] Secondly, the activity quantification results are corrected by combining the dynamic changes in the heart-brain coupling strength: when the coupling strength is high, the heart-brain coordination ability is strong and the autonomic nervous regulation is more stable, and the direction strength of the transfer entropy reflects the activity more accurately; when the coupling strength is low, the heart-brain coordination ability is weak and there are abnormalities in the autonomic nervous regulation, and the weight of the direction strength of the transfer entropy needs to be reduced.

[0061] Based on this, the Autonomic Nervous Balance Index (ANI) is defined as: ANI = (sympathetic nerve activity - parasympathetic nerve activity) / (sympathetic nerve activity + parasympathetic nerve activity), where sympathetic nerve activity = α × TEc2b + β × LF, parasympathetic nerve activity = α × TEb2c + β × HF, α is the weight of the transfer entropy direction intensity (α = 0.6 when the coupling strength is high, α = 0.4 when the coupling strength is low), and β is the weight of the HRV component (β = 0.5, complementary to α), ensuring that ANI can comprehensively reflect the influence of the information transfer direction, HRV characteristics, and coupling strength. The value range of ANI is -1 to 1. When ANI > 0, the sympathetic nerve activity is higher than the parasympathetic nerve activity; when ANI < 0, the parasympathetic nerve activity is higher than the sympathetic nerve activity; when ANI = 0, the two are in a balanced state.

[0062] Furthermore, generate the ANI time-series characteristics that change over time (i.e., autonomic nervous balance characteristics), and label the sympathetic-dominant / parasympathetic-dominant / balanced state at each time point: taking 5 s as the time step, combining the transfer entropy values, HRV components, and coupling strength within each time window, calculate the corresponding ANI values, and arrange all ANI values in chronological order to form the ANI time-series characteristics, that is, the autonomic nervous balance characteristics.

[0063] Meanwhile, in combination with the ECMO clinical diagnostic criteria, set the classification threshold of ANI: when ANI ≥ 0.2, it is determined as the sympathetic-dominant state, indicating that the patient may have excessive sympathetic nerve excitement (such as compensatory excitement caused by stress response and perfusion insufficiency); when ANI ≤ -0.2, it is determined as the parasympathetic-dominant state, indicating that the patient may have excessive parasympathetic nerve inhibition or abnormal function; when -0.2 < ANI < 0.2, it is determined as the balanced state, indicating that the autonomic nerve regulation is within the normal range. Label the ANI value at each time point, and finally form the autonomic nervous balance characteristics including ANI time-series numerical values + state labels, which can dynamically and accurately reflect the regulation state of the autonomic nerves of ECMO patients.

[0064] The pulsation feature extraction module 4 is used to obtain pulsation features by using the physics-informed neural network and the filter neural network cumulative algorithm, including the following steps: S41. Construct a physics-informed neural network, and combine the physics-informed neural network and the arterial blood pressure wave to obtain single-cycle pulsation features.

[0065] First, incorporate the core equation of ECMO hemodynamics as a loss function term of the PINN (physics-informed neural network) to achieve the dual optimization of data fitting + physical law constraint, avoiding unreasonable results caused by pure data-driven.

[0066] The core equations incorporated include: first, Poiseuille's law, which describes the core relationship between blood flow, vascular resistance, and pressure difference, and is the fundamental physical law of ECMO pulsatile perfusion; second, Bernoulli's equation, which supplements the description of the conversion relationship between blood flow velocity and pressure, adapting to the dynamic characteristics of blood flow velocity changing over time in ECMO pulsatile perfusion. These two equations are transformed into constraint terms in the loss function, combined with the mean squared error (MSE) loss, to construct the total loss function of PINN (total loss = data fitting loss + λ × physical constraint loss, where λ is the constraint weight, set to 0.3 after clinical data validation to ensure a balance between physical laws and data fitting), forcing the network to output pulsatile parameters that conform to hemodynamic laws.

[0067] Furthermore, the input features include: the rising slope of the ABP waveform (reflecting the rate of increase in myocardial contractility and perfusion pressure), the amplitude of the systolic peak (reflecting the maximum perfusion pressure during systole), the end-diastolic pressure (reflecting the baseline perfusion pressure during diastole), the duration of systole, the duration of diastole, and the amplitude and location of the dicrotic wave (reflecting aortic valve function and vascular elasticity). All input features are Z-score standardized to eliminate dimensional differences and ensure the stability of network training.

[0068] The specific characteristics of a single-cycle pulsation include pulsational flow (blood flow in each cardiac cycle), pulsational pressure gradient (the rate of change of the pressure difference between systole and diastole), estimated vascular resistance (derived based on Poiseuille's law), and blood flow acceleration (reflecting the impact intensity of the pulsation). These output parameters directly reflect the physical effects of pulsational perfusion.

[0069] By combining ECMO extracorporeal circulation experimental data with real-world clinical data, a high-quality training dataset was constructed to optimize the physical constraint weights of PINN, ensuring the model's adaptability to ECMO clinical scenarios. The extracorporeal circulation experimental data, sourced from an ECMO simulation platform, encompasses standard ABP waveforms and corresponding hemodynamic parameters (measured using specialized instruments and used as labeled data) at different pulse frequencies (60-120 beats / minute) and pulse amplitudes (10-30 mmHg), used to calibrate the parameters of the physical constraint equations. The real-world clinical data, derived from ECMO patient monitoring records, covers patients with varying conditions (severe heart failure, respiratory failure) and different baseline characteristics (age, weight, underlying diseases), used to optimize the weighting ratio between data fitting loss and physical constraint loss.

[0070] During training, the gradient descent method is used to iteratively optimize the network parameters. Every 100 iterations, the deviation between the model output and the true label is verified. If the output parameters violate physical laws (such as the negative correlation between pulsation flow and pressure gradient), the weight of the physical constraint loss is increased until the model output conforms to the laws of hemodynamics and the data fitting error is less than 5%. The PINN training is then completed and the model parameters are saved for real-time estimation of single-cycle pulsation features.

[0071] S42. Introduce a filter neural network accumulation algorithm to capture the temporal accumulation effect of the single-cycle pulsation characteristics.

[0072] The single-cycle pulsation features output by the PINN network can only reflect the pulsation state of a single cardiac cycle. However, the effectiveness of ECMO pulsation perfusion depends on the cumulative effect of multiple consecutive cycles (e.g., 10 consecutive cycles of low-amplitude pulsations, even with normal single-cycle parameters, may lead to insufficient cerebral perfusion). Therefore, a filter neural network accumulation algorithm is introduced to capture the temporal cumulative effect of pulsation features and avoid the limitations of single-cycle features. The specific operation is as follows: A one-dimensional filter neural network was constructed, employing a simplified structure of input layer-convolutional layer-batch normalization layer-cumulative pooling layer-output layer. The input consisted of single-cycle pulsation features (including pulsational flow, pressure gradient, vascular resistance, etc.) output by the PINN, arranged chronologically to form a temporal feature sequence. The convolutional layer used multiple one-dimensional convolutional kernels, with kernel sizes set according to the ECMO pulsation cycle range. The standard clinical pulsation frequency for ECMO is 60-120 beats / minute, corresponding to a single cycle duration of 0.5-1 second (125-250 sampling points at a 250Hz sampling frequency). Therefore, three kernel sizes (125, 175, and 250) were used to capture feature patterns at different pulsation frequencies. The stride of the convolutional kernel was set to 1 to ensure no feature from any cycle was missed. Simultaneously, a batch normalization layer was introduced to accelerate network training convergence, avoid gradient vanishing, and ensure the network could stably capture the temporal changes in pulsation features.

[0073] Then, a cumulative pooling layer is used to accumulate and calculate the pulsation features of 10 consecutive cardiac cycles, extracting temporal cumulative features to comprehensively reflect the long-term effects of pulsation perfusion. The window size of the cumulative pooling layer is set to 10 (corresponding to 10 consecutive cardiac cycles, with a duration of approximately 5-10 seconds, adapted to the ECMO parameter adjustment cycle), and the sliding step size is set to 1, so that the cumulative features are updated once for each new cardiac cycle, ensuring the real-time performance of the cumulative features.

[0074] The specific cumulative calculation includes three core dimensions: first, amplitude accumulation, which calculates the average pulsation amplitude and average pulsation flow over 10 consecutive cycles, reflecting the overall intensity of the pulsation; second, stability accumulation, which calculates the coefficient of variation of pulsation frequency and the coefficient of variation of pulsation amplitude over 10 consecutive cycles, reflecting the stability of the pulsation (the smaller the coefficient of variation, the more stable the pulsation and the more balanced the perfusion effect); and third, energy accumulation, which calculates the cumulative pulsation energy over 10 consecutive cycles (based on the product of pulsation amplitude and frequency), reflecting the total energy input of pulsation perfusion. Insufficient energy may lead to insufficient tissue perfusion, while excessive energy may increase the burden on the heart. All cumulative characteristics are standardized to maintain the same dimensions as the single-cycle characteristics.

[0075] S43. Combining the single-cycle pulsation features and cumulative pulsation features, obtain the pulsation feature set.

[0076] Single-cycle pulsation features reflect the instantaneous pulsation state, while cumulative pulsation features reflect the long-term pulsation effect. The single-cycle pulsation features (pulsation frequency, amplitude, rise time, pulsation flow, pressure gradient, and vascular resistance) are dimensionally aligned with the time-series cumulative features output by the filter neural network (average pulsation amplitude, coefficient of variation of pulsation frequency, cumulative pulsation energy, and average pulsation flow) to ensure consistent time steps for all features (divided into 5-second intervals, matching the time steps of the previously discussed autonomic nervous system balance features and cardiovascular / cerebrovascular state features). Subsequently, feature concatenation is used for fusion, and features are discarded. Redundant features (such as the non-redundant coefficient of variation of single-cycle pulsation frequency and the cumulative features, which are all retained; the complementary single-cycle pulsation amplitude and the average pulsation amplitude, which are also retained) are all retained. Finally, the fused features are standardized to generate a pulsation feature set containing core indicators such as pulsation frequency, pulsation amplitude, rise time, pulsation flow, pressure gradient, vascular resistance, average pulsation amplitude, coefficient of variation of pulsation frequency, and cumulative pulsation energy. This feature set covers both the physical parameters of instantaneous pulsation and the cumulative effect of long-term pulsation, comprehensively reflecting the effectiveness and stability of ECMO pulsation perfusion.

[0077] The physiological needs analysis module 5 is used to identify the causal relationship between cardiovascular and cerebrovascular status and perfusion parameters, quantify the degree of influence of different parameters on individual status, and generate an individualized physiological needs vector, including the following steps: S51. Construct a dataset that associates features with perfusion parameters using a cardiovascular and cerebrovascular state feature set, and identify the direct causal relationship between perfusion parameters and state through a causal discovery algorithm.

[0078] First, a high-quality state-parameter association dataset is constructed, deeply correlating the cardiovascular and cerebrovascular state feature set with clinical ECMO perfusion parameters. This ensures that the dataset can accurately reflect the dynamic correspondence between perfusion parameters and the patient's cardiovascular and cerebrovascular state, providing reliable data support for subsequent causal relationship identification and sensitivity analysis. The specific steps are as follows: The core data sources are integrated into two sets: first, a cardiovascular and cerebrovascular status feature set (including three major categories of quantitative features and labels: cardiac function status, cerebrovascular status, and perfusion risk level); and second, core clinical ECMO perfusion parameters, covering key adjustable parameters such as blood flow (4-8 L / min in the conventional range), centrifugal pump speed (1500-3000 r / min), pulse rate (60-120 beats / min), pulse amplitude (10-30 mmHg), and oxygen concentration (50%-100%). At the same time, patient baseline information is supplemented (age, weight, height, underlying diseases such as heart failure / respiratory failure type, reason for ECMO initiation, past medical history, etc.). This baseline information is the core basis for subsequent individualized weight adjustments.

[0079] Secondly, the standard for sample construction is standardized. Each sample strictly includes four parts: patient baseline information, current cardiovascular and cerebrovascular status characteristics, current ECMO perfusion parameter values, and status feedback results. The status feedback results are the changes in cardiovascular and cerebrovascular status within 5-10 seconds after parameter maintenance or adjustment (e.g., after the pulsation amplitude is increased by 5 mmHg, the cerebral perfusion pressure increases by 8 mmHg, the blood flow is maintained at 5 L / min, and the myocardial oxygen consumption is maintained within the normal range), to ensure the temporal correlation and clinical reference value of the samples.

[0080] Finally, the dataset was optimized by collecting clinical data from ECMO patients (covering different disease severity, age groups, and underlying disease types). Samples were collected from each patient at 5-second time steps, resulting in a sufficiently large associated dataset. Missing values ​​in the dataset were supplemented using K-nearest neighbor interpolation (which is superior to linear interpolation and better reflects the non-linear characteristics of physiological data). Abnormal samples (such as abnormal state feedback after parameter adjustment or invalid data caused by equipment failure) were labeled and removed. Z-score standardization was used to unify the dimensions of all features and parameters. The dataset was divided into training set (70%), validation set (15%), and test set (15%) for subsequent training and validation of causal discovery algorithms and parameter sensitivity networks.

[0081] There are numerous correlations between ECMO perfusion parameters and cardiovascular status, but not all of these correlations can guide clinical regulation (e.g., spurious correlations may lead to erroneous parameter adjustments). Therefore, this invention employs a causal discovery algorithm to accurately identify the true causal relationship between parameter adjustment and status change, eliminating spurious correlations and providing a reliable causal basis for subsequent sensitivity analysis and demand inference. The specific operation is as follows: To optimize the dynamic characteristics of ECMO time-series data, a time delay term is introduced to adapt to the actual clinical patterns of ECMO parameter adjustments: After ECMO parameter adjustments, the cardiovascular and cerebrovascular status does not change immediately, but rather there is a certain time delay (e.g., cerebral perfusion pressure will only show significant changes 3-5 seconds after blood flow adjustment, and heart rate will only tend to stabilize 5-10 seconds after pulsation frequency adjustment). Therefore, a time delay gradient of 0-10 seconds is introduced. By traversing different delay durations, the time causal chain of parameter adjustment → status change after a delay of t seconds is identified to ensure the authenticity of the causal relationship and its clinical suitability.

[0082] Specifically, by progressively eliminating conditionally independent variable pairs, a causal graph is constructed, where nodes represent perfusion parameters and cardiovascular / cerebrovascular status indicators, edges represent causal relationships, and the direction of the edges is parameter → state, with corresponding time delays labeled (e.g., increased pulsation frequency → increased cerebral perfusion pressure after 3 seconds). For example, the algorithm can clearly identify that increased pulsation frequency → increased cerebral perfusion pressure is a true causal relationship (increased pulsation frequency can increase stroke volume, thereby increasing cerebral perfusion pressure), while pulsation frequency and increased rSO2 may only be correlated (both may be affected by blood flow).

[0083] At the same time, temporal constraints are introduced to ensure that the causal relationship conforms to the time logic of parameter adjustment first and state change later, and avoids reverse causality (such as cerebral perfusion pressure increase → parameter adjustment, which is actually the clinician adjusting the parameters according to the change of cerebral perfusion pressure, which is a passive feedback, rather than the active regulation causality required by the algorithm).

[0084] Furthermore, combining ECMO clinical knowledge and algorithm verification, we distinguish between direct causality, indirect association, and spurious association to ensure the validity of causal relationships. The core characteristic of spurious associations is that the two factors have no direct physiological mechanism connection, but rather indirectly affect each other through a third-party variable. For example, equipment temperature → heart rate change may seem correlated, but in reality, when the equipment temperature rises, the centrifuge pump speed may fluctuate slightly, leading to changes in blood flow, which ultimately affects heart rate. Equipment temperature and heart rate have no direct physiological regulatory relationship, thus constituting a spurious association that must be eliminated. Similarly, respiratory rate → changes in cerebral oxygen saturation are both affected by cerebral perfusion pressure, making it an indirect association that also needs to be eliminated. Only causal relationships with direct physiological mechanisms are retained (such as blood flow → cerebral perfusion pressure, pulsation amplitude → myocardial oxygen consumption). After elimination, the remaining causal relationships undergo clinical verification and are reviewed and confirmed by ECMO clinicians to ensure that each causal relationship conforms to clinical physiological mechanisms. This ultimately forms a set of direct causal relationships between perfusion parameters and states, providing the core edge structure for the construction of the parameter sensitivity network.

[0085] S52. Construct a parameter sensitivity network using the direct causal relationship between the perfusion parameters and the state as edges, and quantify the sensitivity coefficient of the perfusion parameters to the cardiovascular and cerebrovascular state.

[0086] Causality analysis can only clarify whether perfusion parameters affect the patient's condition. However, the core requirement of individualized ECMO control is to quantify the degree of influence of parameters on the individual's condition. Therefore, a parameter sensitivity network is constructed. By quantifying the sensitivity coefficient of parameters to cardiovascular and cerebrovascular conditions and dynamically adjusting the weights based on the patient's individual baseline characteristics, individualized adaptation of sensitivity analysis can be achieved. The specific operation is as follows: Using the direct causal relationship between perfusion parameters and states as edges, a parameter sensitivity network is constructed, with nodes divided into two categories: one category is ECMO perfusion parameter nodes, including adjustable core parameters such as blood flow, centrifuge pump speed, pulsation frequency, pulsation amplitude, and oxygen concentration; the other category is cardiovascular and cerebrovascular state indicator nodes, including state indicators such as cerebral perfusion pressure, myocardial oxygen consumption, cardiac output, rSO2, and HRV.

[0087] The weight of an edge is defined as the sensitivity coefficient of a parameter to a state, which is calculated by partial derivatives. Specifically, the calculation method is as follows: with other parameters fixed, calculate the change in the corresponding cardiovascular and cerebrovascular state indicators when a certain parameter changes by 1% (e.g., if the pulsation amplitude changes by 1% and the cerebral perfusion pressure changes by 0.8%, then the sensitivity coefficient of pulsation amplitude → cerebral perfusion pressure is 0.8). The larger the absolute value of the sensitivity coefficient, the more significant the influence of the parameter on the corresponding state indicator.

[0088] After the network is built, the network parameters are optimized using training set data to ensure that the calculated sensitivity coefficient is consistent with actual clinical feedback (e.g., in clinical practice, the effect of pulsation amplitude on cerebral perfusion pressure is greater than that of oxygen concentration, so the sensitivity coefficient of pulsation amplitude → cerebral perfusion pressure in the network should be higher than that of oxygen concentration → cerebral perfusion pressure).

[0089] Furthermore, network weights are dynamically adjusted based on the baseline characteristics of individual patients to achieve personalized adaptation. The core adjustment is based on the patient's baseline information, including key factors such as age, weight, underlying diseases, and ECMO duration, and specific weight adjustment rules are established: For example, elderly patients (≥65 years old) have decreased vascular elasticity and weakened cerebral perfusion regulation ability, and are significantly more sensitive to blood pressure changes than younger patients. Therefore, the weight of pulsation amplitude → cerebral perfusion pressure and blood flow → cerebral perfusion pressure is increased by 20%-30%; heavier patients (≥80kg) require higher blood flow to maintain normal perfusion, so the weight of blood flow → cardiac output is increased by 15%-25%; patients with congenital heart disease have weaker myocardial function and lower tolerance to pulsation amplitude, so the weight of pulsation amplitude → myocardial oxygen consumption is increased by 30%-40% to avoid excessive pulsation amplitude leading to a surge in myocardial oxygen consumption; for patients with ECMO duration exceeding 72 hours, vascular resistance may increase, so the weight of rotation speed → blood flow is increased by 10%-20%. The weight adjustment uses an adaptive algorithm that automatically matches the adjustment coefficients based on the patient's baseline information. At the same time, it combines real-time patient status feedback (such as whether the status change after adjustment meets expectations) to dynamically fine-tune the weights, ensuring that the sensitive network can accurately adapt to the physiological characteristics of individual patients.

[0090] S53. Combine the aforementioned sensitivity coefficient and individual baseline characteristics to generate an individualized physiological demand vector.

[0091] Based on the optimized parameter sensitivity network and combined with the patient's current cardiovascular and cerebrovascular status, target values ​​for perfusion parameters that conform to individual physiological characteristics are inferred and standardized into an individualized physiological demand vector. The specific operation is as follows: Based on parameter-sensitive networks, combined with the current cardiovascular and cerebrovascular status of patients, the criteria for determining the optimal cardiovascular and cerebrovascular status are set, and then the corresponding perfusion parameter target values ​​are inferred in reverse.

[0092] The determination of optimal cardiovascular and cerebrovascular status strictly follows the ECMO clinical guidelines and is dynamically adjusted in combination with the baseline characteristics of individual patients. For example, the optimal cerebral perfusion pressure for ordinary adult patients is 60-80 mmHg, rSO2 ≥ 60%, and myocardial oxygen consumption is maintained within the normal range (25-35 ml / min·100g), while the optimal cerebral perfusion pressure for elderly patients can be lowered to 55-75 mmHg, and the myocardial oxygen consumption for heart failure patients can be lowered to 20-30 ml / min·100g.

[0093] During the reasoning process, the optimal cardiovascular and cerebrovascular state is the target. The optimal values ​​for each perfusion parameter are calculated using a sensitivity network. For example, if the current cerebral perfusion pressure is 50 mmHg (below the optimal range), the sensitivity network shows a sensitivity coefficient of 0.8 for pulsation amplitude versus cerebral perfusion pressure and 0.6 for blood flow versus cerebral perfusion pressure. Therefore, pulsation amplitude is prioritized for adjustment. Based on the current pulsation amplitude (15 mmHg), it is deduced that the pulsation amplitude needs to be adjusted to 20 mmHg, and the blood flow maintained at 5 L / min. Simultaneously, other state indicators such as myocardial oxygen consumption and rSO2 are ensured to be within their optimal ranges. This ultimately yields individualized target values ​​for perfusion parameters (e.g., blood flow 5 L / min, pulsation frequency 80 beats / min, pulsation amplitude 20 mmHg, rotational speed 2000 r / min, oxygen concentration 80%). The reasoning for the target values ​​must consider the synergistic optimization of multiple state indicators to avoid a single indicator being optimal while other indicators are abnormal (e.g., when increasing blood flow to improve cerebral perfusion pressure, myocardial oxygen consumption must not exceed the normal range).

[0094] The perfusion parameter target values ​​obtained through inference are standardized to generate a uniformly formatted individualized physiological demand vector. The dimension of the vector is defined as parameter type × target value, where parameter type covers all core adjustable parameters of ECMO (blood flow, rotational speed, pulse rate, pulse amplitude, oxygen concentration), and each parameter type corresponds to a standardized target value. For example, the standardized individualized physiological demand vector can be represented as [blood flow (standardized value 0.3), rotational speed (standardized value 0.2), pulse rate (standardized value 0.4), pulse amplitude (standardized value 0.5), oxygen concentration (standardized value 0.6)], where each value in the vector corresponds to the standardized result of the target value of that parameter, preserving the relative magnitude of the target value while achieving dimensional uniformity with other features.

[0095] The collaborative perception feature extraction module 6 is used to integrate the cardiovascular and cerebrovascular state feature set, autonomic nervous balance features, and individualized physiological demand vector to obtain collaborative perception features, including the following steps: S61. Construct a dynamic routing mechanism.

[0096] First, feature dimensions are aligned. Using the ECMO parameter adjustment cycle (5s) as a unified time step, the time dimensions of the three types of features are synchronized: the cardiovascular and cerebrovascular status feature set extracts core quantitative indicators (such as the estimated value of cerebral perfusion pressure, myocardial oxygen consumption, and perfusion risk level every 5s) in a 5s time step; the autonomic nervous balance feature extracts the ANI index and time series values ​​in a 5s time step; and the individualized physiological demand vector updates the parameter target values ​​in a 5s time step. This ensures that the time dimensions of the three types of features are completely matched and avoids fusion deviations caused by time misalignment.

[0097] Secondly, standardization and unification are performed, followed by feature splicing and integration. The three types of standardized features are spliced ​​in the order of cardiovascular and cerebrovascular status features → autonomic nervous balance features → individualized physiological demand vectors. Redundant features are removed (such as cerebral perfusion pressure in cardiovascular and cerebrovascular status and the target value of cerebral perfusion pressure in individualized demand, which have no redundancy and are both retained). Finally, a fused input feature matrix is ​​formed. The matrix dimension is time step × total number of features, where the total number of features covers all the core indicators of the three types of features. This ensures that the matrix can comprehensively and accurately present the patient's status, neuromodulation needs, and individualized perfusion needs, laying the foundation for dynamic routing mechanisms and expert sub-network reasoning.

[0098] Furthermore, a gating network is constructed (the input is the fusion feature matrix, and the output is the weights of each expert sub-network). The training objective of the gating network is to dynamically allocate higher weights to the corresponding expert sub-networks based on the patient's current state (e.g., insufficient cerebral perfusion + parasympathetic dominance with excessive myocardial oxygen consumption + sympathetic dominance), so as to ensure that the output of the expert sub-network can accurately match the patient's current core needs.

[0099] The gated network adopts a lightweight structure design, specifically an input layer, a fully connected layer, and a softmax output layer. The input layer receives the fused input feature matrix, the fully connected layer extracts feature association information through the ReLU activation function, and the softmax output layer outputs the weights of the four expert subnetworks (the sum of the weights is 1), ensuring the rationality of the weight allocation.

[0100] During training, ECMO clinical decision-making rules were incorporated as constraints. For example, when cerebral perfusion pressure was <55 mmHg and rSO2 was <55% (insufficient cerebral perfusion), the weight of the cerebrovascular regulation expert subnetwork was ≥0.6; when ANI was >0.5 (sympathetic dominance) and myocardial oxygen consumption was >35 ml / min·100g (excessive myocardial oxygen consumption), the weight of the cardiovascular regulation expert subnetwork was ≥0.5; and when the deviation between the target value and the current value of pulsation amplitude in the individualized demand vector was >5 mmHg, the weight of the individualized demand matching expert subnetwork was ≥0.4. These constraints ensured that the weight allocation of the gating network conformed to the clinical decision-making logic, rather than a blind allocation driven by pure data.

[0101] Meanwhile, the cross-entropy loss function is used to optimize the parameters of the gating network. The training data comes from the constructed state-parameter association dataset to ensure that the gating network can accurately learn the correspondence between patient state and expert weights. After training, the gating network can output the weights of each expert sub-network within 50ms based on the patient's real-time fusion feature matrix, meeting the real-time control requirements.

[0102] Furthermore, based on the core regulatory requirements of ECMO pulsatile perfusion, four types of expert subnetworks were customized. Each type of subnetwork focuses on feature reasoning in a specific domain, achieving precise adaptation to multi-dimensional needs and avoiding the problem of insufficient adaptability of general expert subnetworks. The specific definitions are as follows: The first is the cardiovascular regulation expert sub-network, which focuses on the adaptation reasoning of cardiac function status. The inputs are cardiac function status features (myocardial oxygen consumption, cardiac output, myocardial contractility, etc.) in the fusion features and target values ​​of cardiovascular-related parameters in the individualized needs. The outputs are cardiovascular perception features such as perfusion frequency adapted to heart rate and pulsation amplitude adapted to myocardial oxygen consumption. The main focus is to solve the core problem of how to adjust parameters to optimize cardiac function.

[0103] The second is the cerebrovascular regulation expert subnetwork, which focuses on the adaptation reasoning of cerebrovascular status. The inputs are cerebrovascular status features (cerebral perfusion pressure, rSO2, cerebral blood flow velocity, etc.) in the fusion features and the target values ​​of cerebrovascular-related parameters in the individualized needs. The outputs are cerebrovascular domain perception features such as perfusion pressure for cerebral oxygen adaptation and pulsation frequency for cerebral perfusion stability adaptation. The main focus is to solve the core problem of how to adjust parameters to ensure cerebral perfusion safety.

[0104] Third is the autonomic nervous system regulation expert subnetwork, which focuses on the adaptation reasoning of the autonomic nervous system balance state. The input is the autonomic nervous system balance features (ANI index, sympathetic / parasympathetic state labels, etc.) in the fusion features, and the output is the perceptual features in the field of neuroregulation such as the pulsation amplitude and the perfusion frequency of the autonomic nervous system balance adaptation. It focuses on solving the core problem of how to adjust parameters to improve the autonomic nervous system balance.

[0105] Fourthly, there is the individualized needs matching expert subnetwork, which focuses on the accurate matching of individualized physiological needs. The input consists of the individualized physiological needs vector from the fused features and the current ECMO perfusion parameter values. The output includes domain-aware features such as the parameter target value matching degree and the adaptability of parameter adjustment ranges. It primarily addresses the core issue of how to adjust parameters to accurately match individual needs. Each type of expert subnetwork employs a streamlined CNN+GRU structure, balancing inference accuracy and computational efficiency. During training, parameters are optimized using clinical data from the corresponding domain to ensure that the subnetwork can accurately output the core perceptual features of that domain.

[0106] S62. Based on the dynamic routing mechanism, the cardiovascular and cerebrovascular state feature set, autonomic nervous balance features, and individualized physiological demand vector are fused to obtain collaborative perception features.

[0107] The fusion input feature matrix is ​​input into each expert sub-network, and each sub-network outputs perceptual features specific to its domain, enabling accurate reasoning across multiple domains. Specifically, the fusion input feature matrix is ​​simultaneously input into four types of expert sub-networks. Each sub-network outputs targeted perceptual features based on its own reasoning logic: the cardiovascular regulation expert sub-network, combining current myocardial oxygen consumption and cardiac output, outputs features such as the perfusion frequency appropriate to heart rate (e.g., when the heart rate is 85 beats / min, the appropriate pulsation frequency is 82-88 beats / min) and the pulsation amplitude appropriate to myocardial oxygen consumption (e.g., when myocardial oxygen consumption is 32 ml / min·100g, the appropriate pulsation amplitude is 18-22 mmHg); the cerebrovascular regulation expert sub-network, combining current cerebral perfusion pressure and rSO2, outputs features such as the perfusion pressure appropriate to cerebral oxygen (e.g., when rSO2 is 58%, the appropriate cerebral perfusion pressure is 60- The system outputs features such as 70 mmHg (corresponding to a pulsation amplitude of 20-24 mmHg) and pulsation frequency adapted to stable brain perfusion. The autonomic nervous system regulation expert subnetwork, combining the current ANI index and sympathetic / parasympathetic state, outputs features such as pulsation amplitude adapted to nervous system regulation (e.g., when the sympathetic system is dominant, the adapted pulsation amplitude is slightly higher to ensure adequate perfusion) and perfusion frequency adapted to autonomic nervous system balance. The individualized needs matching expert subnetwork, combining the current parameter values ​​with the individualized needs target values, outputs features such as the parameter target value matching degree (e.g., current pulsation amplitude 15 mmHg, target value 20 mmHg, matching degree is 0.75) and parameter adjustment range adaptability (e.g., a suggested increase of 5 mmHg, adaptability is 0.9). The outputs of each expert subnetwork are standardized to ensure consistent feature dimensions, laying the foundation for subsequent weighted fusion.

[0108] Then, the expert outputs are weighted and fused using dynamic routing weights to generate collaborative perceptual features that include multi-dimensional demand matching, state fit, and neural modulation fit. The fusion process strictly follows the principles of weight fit and clinical priority, and the specific operation is as follows: First, obtain the weights of the four types of expert subnetworks output by the gating network (e.g., cerebrovascular regulation expert weight 0.6, cardiovascular regulation expert weight 0.2, autonomic nervous regulation expert weight 0.1, and individualized demand matching expert weight 0.1).

[0109] Secondly, the perceptual features output by each expert subnetwork are weighted and calculated together with their corresponding weights. For example, the perfusion pressure feature value for brain oxygen adaptation is multiplied by 0.6, the perfusion frequency feature value for heart rate adaptation is multiplied by 0.2, and so on.

[0110] Finally, all weighted features are integrated to generate collaborative perceptual features, which contain three core dimensions: First, it assesses the multi-dimensional matching degree of needs, covering the degree of matching for cardiovascular, cerebrovascular, autonomic nervous, and individualized needs, and quantifies the fit between current parameters and various needs.

[0111] Second, state fit, which quantifies the degree of fit between the current perceptual features and the patient's cardiovascular and autonomic nervous system status, reflects the potential effect of parameter adjustment.

[0112] Thirdly, neural regulation adaptation, which quantifies the degree of adaptation between perceptual features and the autonomic nervous system balance state, to ensure that parameter adjustments can take into account the needs of neural regulation.

[0113] The dimensions of the collaborative sensing features are consistent with the fusion input feature matrix, which not only achieves complementary collaboration of multi-source features, but also preserves the core information of each domain.

[0114] The comprehensive decision confidence analysis module 7 is used to obtain comprehensive decision confidence by fusing autonomic nerve balance characteristics and pulsation characteristics through DS evidence theory, including the following steps: S71. Construct evidence through autonomic balance and pulsation characteristics.

[0115] Evidence body E1 focuses on neuroregulatory needs, with autonomic balance characteristics as its core. Specifically, it includes the time-series value of the ANI index, sympathetic / parasympathetic state labels (sympathetic dominance, parasympathetic dominance, and balance), and supplements the rate of change of ANI values ​​(such as the magnitude of ANI increase / decrease within 5 seconds). These characteristics directly reflect the patient's autonomic nervous system's regulatory needs for perfusion parameters. For example, when the sympathetic nervous system is dominant, the patient may need higher perfusion pressure to meet the body's stress requirements, while when the parasympathetic nervous system is dominant, it is necessary to avoid excessively high perfusion parameters that would increase the burden on the heart.

[0116] Evidence body E2 focuses on the physical perfusion effect, with pulsation characteristics as the core, selecting three key indicators: pulsation amplitude, pulsation frequency, and cumulative pulsation energy. It also supplements the data with the coefficient of variation of pulsation amplitude and the average pulsation flow rate. These characteristics directly reflect the physical effect of the current ECMO pulsation perfusion. For example, if the pulsation amplitude is too low or the cumulative energy is insufficient, it indicates poor perfusion effect, and it may be necessary to increase the relevant parameters. If the pulsation frequency fluctuates abnormally, it indicates insufficient perfusion stability, and the parameters need to be adjusted to maintain balanced perfusion.

[0117] The two evidence bodies provide decision-making evidence from the dimensions of neural needs and physical effects, respectively, forming a complementary relationship. At the same time, redundant features irrelevant to parametric decision-making are eliminated to ensure the relevance and effectiveness of the evidence bodies.

[0118] Furthermore, based on the core decision-making objectives of ECMO pulsation perfusion parameter regulation, an identification framework Θ={perfusion parameters need to be increased, perfusion parameters need to be decreased, and perfusion parameters remain unchanged} is set. The three propositions are independent of each other and completely exhaustive, covering all possible parameter regulation decision-making scenarios.

[0119] Among them, the perfusion parameters need to be up-regulated for the corresponding scenarios as follows: high neuroregulation demand (such as sympathetic dominance) and poor physical perfusion effect (such as insufficient pulsation amplitude, low cumulative energy), or a single evidence body indicates that up-regulation is required (such as ANI > 0.5 and normal pulsation characteristics, still need to be moderately up-regulated to match the neural demand); the perfusion parameters need to be down-regulated for the corresponding scenarios as follows: low neuroregulation demand (such as parasympathetic dominance) and excessive physical perfusion effect (such as too high pulsation amplitude, too high cumulative energy), or there is a risk of over-perfusion (such as cerebral perfusion pressure exceeding the optimal range); the perfusion parameters remain unchanged for the corresponding scenarios as follows: the neuroregulation demand matches the physical perfusion effect, and the cardiovascular and cerebrovascular status is stable (the perfusion risk level is low risk), and good perfusion effect can be maintained without adjusting the parameters. At the same time, clarify the boundary conditions of the recognition framework, avoid decision-making ambiguity, and ensure the accuracy of subsequent BPA allocation and evidence fusion.

[0120] S72. Based on the BPA allocation principle of rules and data-driven, obtain the comprehensive decision confidence using the evidence body.

[0121] Based on ECMO clinical practice, combined with the core indicators of evidence body E1 (autonomic nerve balance characteristics) and E2 (pulsation characteristics), formulate multiple groups of refined BPA allocation rules to cover different clinical scenarios and ensure the comprehensiveness and operability of the rules.

[0122] The rule formulation follows the principles of giving priority to clinical priority and collaborative judgment of multiple indicators, and conducts supplementary constraints in combination with the characteristics of the cardiovascular and cerebrovascular status (such as perfusion risk level). Some core rules are as follows: [[ID=Ill]]1. E1 indicates sympathetic dominance (ANI > 0.5) and E2 indicates poor pulsation effect (pulsation amplitude < 15 mmHg, cumulative pulsation energy < 500) → BPA for perfusion parameter up-regulation = 0.8, remain unchanged = 0.1, down-regulation = 0.1 (such scenarios indicate that the patient has an obvious stress response and current perfusion insufficiency, and parameters need to be up-regulated first); 2. E1 indicates parasympathetic dominance (ANI ≤ -0.2) and E2 indicates excessive pulsation effect (pulsation amplitude > 25 mmHg, cumulative pulsation energy > 1000) → BPA for perfusion parameter down-regulation = 0.75, remain unchanged = 0.2, up-regulation = 0.05 (such scenarios indicate that the patient has low neuroregulation demand, and over-perfusion may increase the cardiac burden, and parameters need to be moderately down-regulated); 3. E1 indicates a balanced state (-0.2 < ANI < 0.2) and E2 indicates normal pulsation effect (pulsation amplitude 15 - 25 mmHg, cumulative pulsation energy 500 - 1000) → BPA for perfusion parameter remaining unchanged = 0.85, up-regulation = 0.07, down-regulation = 0.08 (such scenarios indicate that the neural demand matches the perfusion effect and no parameter adjustment is required); 4. E1 indicates sympathetic dominance (ANI > 0.5), but E2 indicates normal pulsation effect → For BPA where perfusion parameters need to be up - regulated = 0.5, remain unchanged = 0.4, down - regulated = 0.1 (This scenario indicates high neural demand, but the current perfusion effect can meet the basic demand, and only a moderate up - regulation is needed); 5. E1 indicates parasympathetic dominance (ANI ≤ - 0.2), but E2 indicates poor pulsation effect → For BPA where perfusion parameters remain unchanged = 0.6, up - regulated = 0.3, down - regulated = 0.1 (This scenario needs to balance neural demand and perfusion insufficiency, temporarily do not down - regulate, and can be adjusted after appropriate observation).

[0123] For the evidence bodies E1 and E2 respectively, calculate their BPA values for the three propositions in the recognition framework according to the above rules, ensuring the standardization and consistency of the calculation process.

[0124] First, extract the core features of E1 at the current time step (ANI index, status label, change rate) and the core features of E2 (pulsation amplitude, frequency, cumulative energy, coefficient of variation), and对照 the BPA assignment rules to determine the corresponding rule entries.

[0125] Secondly,结合 the perfusion risk level in the cardiovascular and cerebrovascular status features, fine - tune the basic BPA values in the rules (for example, when the risk is high, the up - regulated / down - regulated BPA value increases by 0.1, and the BPA value that remains unchanged decreases by 0.1).

[0126] Finally, ensure that the sum of the BPA values of each evidence body is 1 (that is, for E1, m1(up - regulated)+m1(remain unchanged)+m1(down - regulated)=1; for E2, m2(up - regulated)+m2(remain unchanged)+m2(down - regulated)=1). If there is a deviation in the sum, use normalization processing to correct it to ensure the rationality of the BPA values.

[0127] Regarding the possible information complementarity and conflict between the two evidence bodies E1 and E2 (such as E1 indicates up - regulation and E2 indicates remaining unchanged), use the Dempster combination rule for evidence fusion, and at the same time introduce a conflict evidence correction mechanism to quantify the degree of collaborative support of multi - source evidence, and finally generate a reliable comprehensive decision confidence level. The specific operations are as follows: Dempster combination rule fusion: First, calculate the conflict coefficient K of the two evidence bodies E1 and E2 to judge the degree of conflict between the evidences. The larger the K value, the more obvious the contradiction between the two evidence bodies. The calculation formula for the K value is K = 1 - Σ[m1(A)×m2(A)] (A is each proposition in the recognition framework). If K = 0, it means the evidences are completely consistent and can be directly fused; if 0 < K < 1, it means the evidences have partial conflicts and use the Dempster combination rule for fusion; if K = 1, it means the evidences are completely in conflict and the conflict correction mechanism needs to be started.

[0128] During the fusion process, the calculation of the comprehensive BPA value m(·) follows Dempster's combination rule: for each proposition A in the recognition frame, m(A)=[Σ{m1(B)×m2(C)}] / (1-K), where B and C are propositions in the recognition frame, and B∩C=A.

[0129] For scenarios with completely conflicting evidence (K=1) or highly conflicting evidence (K>0.7), a conflict correction is introduced by prioritizing ECMO clinical decision-making. Evidence related to cerebral perfusion is given priority (such as pulsation amplitude and cumulative energy in E2, or indirect features related to cerebral oxygenation in E1) to avoid decision-making bias due to conflicting evidence. For example, if E1 indicates parasympathetic dominance (needs to be downregulated), E2 indicates extremely low pulsation amplitude (needs to be upregulated), and cerebral perfusion pressure is <55mmHg, then the evidence from E2 is given priority, and the BPA value is corrected before fusion.

[0130] After fusing to obtain the comprehensive BPA value, the confidence level (Bel) of each decision proposition is calculated based on the comprehensive BPA. The confidence level is defined as the sum of the BPA values ​​of the proposition and all its subsets. Since the three propositions in the recognition framework are independent of each other, the confidence level of each proposition is equal to its corresponding comprehensive BPA value.

[0131] Finally, a comprehensive decision confidence score is generated and output in the form of increased confidence score, maintained confidence score, and decreased confidence score, which clarifies the degree of support for each decision proposition. For example, the confidence score is increased by 0.75, maintained by 0.2, and decreased by 0.05. The proposition with the highest confidence score is the most reasonable decision direction at present.

[0132] The dynamic regulation baseline generation module 8 is used to generate a dynamic regulation baseline based on historical regulation experience data through multi-agent deep reinforcement learning and knowledge distillation, including the following steps: S81. Construct multiple collaborative agents and define the state space, action space, and reward function. Train and generate an offline MARL model using historical regulation experience datasets.

[0133] To provide high-quality training data support for multi-agent reinforcement learning (MARL) and ensure that the trained model can fit the actual clinical decision-making logic of ECMO, the dataset construction follows the principles of comprehensiveness, timeliness, and clinical relevance. The specific operation is as follows: First, we integrate historical data from multiple sources. The core data sources include three categories: First, real-time physiological monitoring data of ECMO patients, covering multimodal signals such as ABP, HR, rSO2, and FBF after cleaning and reconstruction, as well as extracted cardiovascular and cerebrovascular status characteristics, autonomic nervous balance characteristics, and pulsation characteristics; Second, clinical parameter adjustment records, which collect detailed information on every adjustment made by medical staff to the core ECMO parameters (blood flow, pulsation frequency, pulsation amplitude, rotation speed, and oxygen concentration), including adjustment time, parameter values ​​before adjustment, parameter values ​​after adjustment, and reasons for adjustment (such as decreased cerebral oxygenation, abnormal heart rate, etc.); Third, clinical outcome data, covering key clinical indicators such as patient ECMO on-machine duration, 24-hour / 72-hour survival rate, duration of normal cerebral oxygenation, occurrence of complications (such as cerebral infarction, arrhythmia, hemorrhage, etc.), and weaning success rate.

[0134] Secondly, data preprocessing and labeling: the collected historical data is standardized to remove invalid samples with equipment failure, severe data loss (more than 10 sampling points missing in a single instance), and incomplete clinical records. Finally, complete historical data of ECMO patients is selected and the samples are split according to a 5-second time step to form a historical regulation experience dataset containing a sufficient number of samples.

[0135] Finally, the reward and penalty rules for parameter adjustment and status feedback were formulated. Combining ECMO clinical guidelines and expert experience, a refined reward and penalty mechanism was developed to avoid model bias caused by a single reward and penalty standard. The core rules are as follows: Positive Rewards (Encouraging Reasonable Adjustments) -- If rSO2 increases by ≥3% and remains elevated for more than 5 seconds after adjustment, reward +1; if cerebral perfusion pressure recovers to the optimal range of 55-80 mmHg after adjustment, reward +1.2; if myocardial oxygen consumption drops to the normal range (25-35 ml / min·100g) after adjustment, reward +0.8; if there are no serious complications within 24 hours, additional reward +5.

[0136] Negative penalties (suppressing irrational adjustments) – A decrease in blood pressure ≥10 mmHg after adjustment lasting more than 3 seconds results in a penalty of -0.5; a decrease in cerebral oxygen saturation ≥3% after adjustment results in a penalty of -1.5; the occurrence of arrhythmias (such as atrial fibrillation or ventricular tachycardia) after adjustment results in a penalty of -2; and improper parameter adjustments leading to patient weaning failure or worsening of condition results in a penalty of -10. These reward and penalty rules have been reviewed and confirmed by three ECMO clinicians to ensure they align with clinical decision-making logic and provide clear optimization guidance for the training of the MARL model.

[0137] Furthermore, a multidisciplinary collaborative decision-making scenario for ECMO clinical practice was simulated (cardiologists were responsible for cardiac function regulation, neurologists were responsible for ensuring cerebral perfusion, and ECMO operators were responsible for parameter execution). Through multi-agent collaborative learning, long-term clinical outcomes were optimized, avoiding the limitations of single-agent decision-making. The specific training process is as follows: Three collaborative agents are constructed, each corresponding to the decision-making logic of different clinical roles, to achieve precise adaptation to multi-dimensional regulatory needs. A collaborative decision-making layer is also set up to coordinate the actions of each agent and avoid decision-making conflicts. First, there is the cardiovascular regulatory intelligent agent, which is responsible for the regulation and decision-making of cardiac function. It focuses on cardiac function indicators such as myocardial oxygen consumption, cardiac output, and heart rate. The decision-making goal is to maintain stable cardiac function and avoid excessive myocardial oxygen consumption or insufficient cardiac output.

[0138] Second, there is a cerebrovascular regulatory intelligent agent, which is responsible for the regulation and decision-making of cerebrovascular status. It focuses on cerebrovascular indicators such as cerebral perfusion pressure, rSO2, and cerebral blood flow velocity. The decision-making goal is to ensure cerebral perfusion safety and avoid insufficient or excessive cerebral perfusion.

[0139] Third is the operation execution agent, which is primarily responsible for the execution and feedback of parameter adjustments. It receives decision suggestions from the first two agents, combines them with the feasibility of clinical parameter adjustments (such as step size limits for parameter adjustments and equipment operation constraints), outputs specific parameter adjustment actions, and provides feedback on the state changes after adjustment, providing a basis for subsequent decision optimization.

[0140] Each agent adopts a GRU+fully connected layer structure to ensure that it can capture temporal features and decision correlations, and the agents can communicate in real time through an information interaction layer to share the current patient status and decision intentions.

[0141] Furthermore, the state space, action space, and reward function are defined: the state space comprehensively covers the patient's multi-dimensional state, ensuring that the agent can accurately perceive the current regulatory needs, specifically including: cardiovascular and cerebrovascular status characteristics (cardiac function, cerebrovascular status, perfusion risk level), autonomic nervous balance characteristics (ANI index, sympathetic / parasympathetic state), pulsation characteristics (single-cycle characteristics and cumulative characteristics), and comprehensive decision confidence (up-adjustment / maintenance / down-adjustment confidence). At the same time, the patient's baseline information (age, weight, underlying diseases) is supplemented. The state space dimension is consistent with the fusion input feature matrix to ensure the consistency of data format.

[0142] The adjustment range follows the ECMO clinical parameter adjustment guidelines, clearly defining the adjustment range and step size limits for each core parameter to avoid exceeding the clinical safety range: blood flow adjustment range 4-8 L / min, adjustment step size 0.1-0.5 L / min; pulse rate adjustment range 60-120 beats / min, adjustment step size 1-5 beats / min; pulse amplitude adjustment range 10-30 mmHg, adjustment step size 1-2 mmHg; centrifuge pump speed adjustment range 1500-3000 r / min, adjustment step size 50-100 r / min; oxygen concentration adjustment range 50%-100%, adjustment step size 5%-10%.

[0143] The reward function adopts a fusion model of immediate reward and long-term reward. The immediate reward is based on the parameter adjustment-state feedback reward and punishment rule, reflecting the effect of a single parameter adjustment. The long-term reward focuses on the long-term clinical outcomes of ECMO patients, such as the survival rate within 24 hours (survival is counted as +10, death as -20), the duration of normal brain oxygenation (for every hour maintained, +2), and the success rate of weaning (successful weaning is counted as +15). Through long-term rewards, the agent is guided to break away from the limitations of short-term indicators and focus on the core clinical goals.

[0144] The training employed an offline training + iterative optimization model to ensure the model could fully learn from historical regulatory experience while adapting to individual differences among patients. First, the constructed historical regulatory experience dataset was divided into a training set (80%), a validation set (10%), and a test set (10%). The training set was input into the MARL model, and the parameters of each agent were initialized. The training iteration count was set to 1000 rounds. In each iteration, each agent autonomously outputs a decision action based on the current state space information. The reward function was used to calculate the action's benefit, and the agent parameters were updated using gradient descent. Simultaneously, a collaborative decision layer coordinated the actions of each agent to resolve decision conflicts (e.g., the cardiovascular agent suggested increasing the pulsation amplitude, while the cerebrovascular agent suggested maintaining it; the collaborative layer combined the comprehensive decision confidence with the current brain perfusion state to select the optimal action). Every 100 iterations, the model performance is validated using a validation set. The fit between the model's parameter adjustment suggestions and actual clinical adjustment records is calculated, as well as the deviation between the model's predicted long-term clinical outcomes and the actual outcomes. If the fit is below 85% or the deviation is too large, the reward function weights are adjusted (e.g., increasing the weight of long-term rewards), the agent structure is optimized, and training is iterated again. After training, the performance is evaluated using a test set to ensure that the model can still output reasonable decision suggestions on unseen patient data. Finally, an offline MARL model is generated. This model incorporates multi-agent collaborative decision-making logic and a large amount of historical regulatory experience, providing reliable decision support for online regulation.

[0145] S82. Using knowledge distillation technology, the decision knowledge of the offline MARL model is transferred to the lightweight online student model.

[0146] While offline MARL models can achieve high-precision decision-making, their complex structure and high computational cost cannot meet the computing power requirements of real-time ECMO control. Therefore, knowledge distillation technology is used to transfer the decision-making knowledge of complex offline models to lightweight online models, achieving a balance between high precision and low latency. The specific operation is as follows: The offline MARL model, which has been trained, was used as the teacher model. This model incorporates multi-agent collaborative decision-making logic, extensive historical control experience, and high-precision decision-making capabilities, and can output optimal parameter adjustment suggestions. However, the model has a large number of parameters (approximately 10 million parameters) and high computational latency (approximately 300ms), making it unsuitable for online deployment. A lightweight online student model was constructed to meet the computational demands of real-time ECMO control. The model structure adopts a simplified CNN+GRU lightweight design, reducing the number of network layers and parameters (approximately 500,000 parameters), removing redundant collaborative decision-making layers from the offline model, and retaining the core decision-making reasoning logic.

[0147] Simultaneously, batch normalization layers and lightweight activation functions (such as ReLU6) are introduced to further reduce computational latency, ensuring that the student model can complete inference output within 50ms. The input and output formats of the student model are consistent with those of the teacher model, with real-time features as input and parameter adjustment suggestions and baseline range as output, ensuring compatibility of knowledge transfer.

[0148] Then, samples from the historical regulation experience dataset were selected and input into the teacher model to obtain the output results of the teacher model (including parameter adjustment suggestions, decision confidence, and action selection probability), which served as the label data for distillation training. The same samples were input into the student model to obtain the preliminary output of the student model. A hybrid loss function of prediction loss and distillation loss was constructed to achieve dual optimization: the prediction loss was used to measure the deviation between the student model output and the real clinical parameter adjustment records to ensure that the student model fits the clinical reality; the distillation loss was used to measure the deviation between the student model output and the teacher model output to ensure that the student model learns the decision logic of the teacher model (such as prioritizing the increase of pulsation amplitude when cerebral oxygen is below 50% and moderately increasing blood flow when sympathetic dominance). The distillation loss was calculated using KL divergence to quantify the difference in the output distribution of the two models.

[0149] During training, gradient descent was used to iteratively optimize the student model parameters, adjusting the weight ratio of prediction loss to distillation loss (validated at 1:1.2) to ensure the student model could both learn the high-precision decision-making logic of the teacher model and closely match actual clinical data. Every 50 iterations, the student model's performance (decision accuracy, computational latency) was verified. If the decision deviation between the student model and the teacher model exceeded 5%, the weight of distillation loss was increased; if the computational latency exceeded 50ms, the model structure was further simplified until the student model met the requirements of decision accuracy ≥90% (close to the teacher model) and computational latency ≤50ms. Knowledge distillation was then completed, and the student model parameters were saved for online dynamic adjustment of baseline generation.

[0150] S83. A lightweight online student model based on knowledge distillation is used to generate a dynamic control baseline by combining the patient's real-time physiological state.

[0151] The student model receives all core features in real time (cardiovascular status, autonomic balance, pulsation characteristics, and comprehensive decision confidence), combines them with the patient's baseline information (age, weight, and underlying diseases), and outputs recommended adjustment values ​​plus baseline ranges for each core ECMO parameter through lightweight inference.

[0152] The baseline range is set according to the principle of individualized adaptation and clinical safety constraints, combined with individualized physiological needs vectors, to develop exclusive baseline ranges for different patients. For example: elderly patients (≥65 years old) have poor vascular elasticity and weak cerebral perfusion regulation ability, so the baseline range for blood flow is set at 4.5-5.5 L / min, and the baseline range for pulsation amplitude is set at 18-25 mmHg; for patients with higher weight (≥80 kg), the baseline range for blood flow is set at 5.5-6.5 L / min to ensure sufficient perfusion; for patients with heart failure and weak myocardial function, the baseline range for pulsation amplitude is set at 15-22 mmHg to avoid excessively high pulsation amplitude increasing myocardial oxygen consumption. Specific output examples: Recommended adjustment value for blood flow rate is 5.2 L / min, with a baseline range of 5-5.5 L / min; recommended adjustment value for pulse rate is 78 beats / min, with a baseline range of 75-80 beats / min; recommended adjustment value for pulse amplitude is 22 mmHg, with a baseline range of 20-24 mmHg. The baseline confidence level (e.g., 0.85) is also output to reflect the reliability of the baseline.

[0153] Considering that the physiological state of ECMO patients is constantly changing (such as improvement in condition or occurrence of complications), the baseline needs to be updated in real time to ensure that it always matches the patient's current state. The update frequency should be combined with the routine adjustment cycle of ECMO parameters and the speed of changes in the patient's state.

[0154] The regulation command generation module 9 is used to generate the final pulsation parameter command by fusing collaborative perception features, comprehensive decision confidence, and dynamic regulation baseline through a graph neural network, including the following steps: S91. By combining collaborative perception features, comprehensive decision confidence, and dynamic control baseline, a dynamic graph structure network is constructed.

[0155] Based on the core requirement of generating ECMO pulsation parameter commands, a graph structure that dynamically adapts to the patient's real-time state is constructed. This ensures that graph nodes and edges can accurately characterize the complex relationships between multi-source features and perfusion parameters, providing a structural foundation that fits clinical practice for subsequent graph convolutional inference. The specific operation is as follows: Based on the characteristics and decision-making criteria of the aforementioned steps, nodes are divided into four categories to ensure that the nodes cover all dimensions of input for instruction generation, and that each node has clear clinical significance and quantitative characteristics. The specific classifications are as follows: First, there are collaborative perception feature nodes, which correspond to collaborative perception features and cover four major subcategories: cardiovascular adaptation features, cerebrovascular adaptation features, autonomic nervous system regulation adaptation features, and individualized demand matching features. Each subcategory node contains specific quantitative indicators (such as perfusion pressure for brain oxygen adaptation and pulsation amplitude for neuromodulation adaptation), accurately reflecting the adaptation status of multi-dimensional needs.

[0156] Second, there is the comprehensive decision confidence node, which corresponds to the comprehensive decision confidence level. It includes three sub-nodes: increasing confidence level, maintaining confidence level, and decreasing confidence level. This quantifies the support level for each decision direction and provides the core basis for node weight allocation.

[0157] Third, the baseline nodes are dynamically adjusted. The corresponding dynamic adjustment baseline is classified according to the core adjustable parameters of ECMO. Each parameter corresponds to a baseline node (such as the blood flow baseline node and the pulsation amplitude baseline node), which includes three core quantitative information: baseline recommended value, baseline range, and baseline confidence.

[0158] Fourthly, there are parameter type nodes, covering all core adjustable parameters of ECMO pulsatile perfusion, including five core nodes: blood flow, pulsation frequency, pulsation amplitude, centrifuge pump speed, and oxygen concentration. These serve as the output carrier for the final parameter commands, ensuring the targeted nature of the generated commands. All nodes are standardized and coded, with corresponding quantitative values ​​and clinical significance labeled to avoid inference biases caused by ambiguous nodes.

[0159] Edges between nodes are used to characterize the strength of the association between different nodes. The edge weights are dynamically assigned based on the infusion parameter-state causality relationship and parameter sensitivity coefficient to ensure that the edge weights can accurately reflect the physiological association and degree of influence between features and parameters, and between parameters. The specific rules are as follows: First, the edge types are divided into three categories: feature-parameter edge, confidence-parameter edge, and baseline-parameter edge, which correspond to the association between collaboratively perceived features, comprehensive decision confidence, and dynamically adjusted baseline and parameter type nodes, respectively.

[0160] Secondly, the edge weights range from 0 to 1. The larger the weight value, the stronger the correlation between the two nodes and the greater the impact on the final parameter instructions. For example, the weight of the cerebral perfusion pressure adaptation feature node → pulsation amplitude node is set with reference to the sensitivity coefficient of pulsation amplitude → cerebral perfusion pressure (e.g., 0.8), and is directly assigned a value of 0.8, reflecting the significant influence of pulsation amplitude on cerebral perfusion pressure. Another example is the weight of the confidence level adjustment node → blood flow node, which is dynamically adjusted according to the specific value of the confidence level adjustment (weight 0.9 when confidence level is 0.9, weight 0.6 when confidence level is 0.6), reflecting the supporting role of confidence level in the direction of parameter adjustment.

[0161] Finally, association edges are also set between parameter type nodes, such as blood flow node → pulsation amplitude node, with weights set based on clinical physiological mechanisms (e.g., 0.6), reflecting the synergistic regulatory relationship between blood flow and pulsation amplitude (when blood flow increases, pulsation amplitude needs to be adjusted appropriately to maintain perfusion balance), ensuring that the graph structure can fully depict the synergistic relationship between parameters.

[0162] S92. Based on the global feature vector generated by the dynamic graph structure network, preliminary parameter instructions are generated through inference in the fully connected layer.

[0163] Based on the constructed dynamic graph structure, deep fusion and global association learning of multi-source node features are achieved through multi-layer graph convolution operations. Combined with an attention mechanism, the influence of key nodes is highlighted, ensuring that the inference results align with the clinical decision-making logic of ECMO. The specific operations are as follows: We constructed a 3-layer GAT (Graph Attention Network). The core advantage of GAT is that it can automatically learn the importance of nodes through the attention mechanism, without the need for manual setting of node weights, and adapts to the dynamic changes in the state of ECMO patients.

[0164] The weight allocation of the attention mechanism follows the principle of clinical priority + confidence priority. For example, nodes with a comprehensive decision confidence > 0.7 (such as nodes with high upregulation confidence) and nodes with co-sensory features related to brain perfusion (such as nodes with brain oxygen adaptation features) will be assigned higher attention weights (1.2 times that of ordinary nodes) to ensure that the features of these key nodes can play a dominant role in the reasoning process. On the other hand, nodes corresponding to edges with low correlation strength (weight < 0.3) will be assigned lower attention weights to reduce interference from irrelevant information.

[0165] Meanwhile, a batch normalization layer is introduced after each graph convolution layer to accelerate network training convergence, avoid gradient vanishing, and ensure that the network can stably learn the complex relationship patterns between nodes, adapting to the individual differences of different patients.

[0166] The core function of graph convolutional layers is to extract global features by fusing the features of each node with those of its neighboring nodes through neighborhood aggregation operations. Specifically, the first graph convolutional layer focuses on local node associations, aggregating the features of each node with its directly adjacent nodes (e.g., fusing features of collaborative perception feature nodes with those of corresponding parameter type nodes), capturing local feature associations. The second graph convolutional layer focuses on cross-type feature fusion, further fusing the aggregated features of different types of nodes (e.g., confidence nodes, baseline nodes) to characterize the collaborative relationships of multi-dimensional inputs. The third graph convolutional layer focuses on global association optimization, globally integrating the fused features of the first two layers, correcting feature biases, and outputting a global feature vector with unified dimensions. The dimensions of the global feature vector correspond to the number of parameter type nodes, with each dimension corresponding to a fused feature of a core infused parameter, covering all core information of collaborative perception, decision confidence, and dynamic baseline, providing comprehensive feature support for subsequent parameter instruction generation.

[0167] Based on the global feature vector output by graph convolution, preliminary parameter instructions are generated through fully connected layers. These instructions are then revised and optimized in accordance with ECMO clinical safety guidelines and individualized needs to ensure their safety, accuracy, and personalization. The specific operations are as follows: The global feature vector output by the graph convolution is input into two fully connected layers. The first fully connected layer is responsible for feature compression and dimension adaptation, mapping the global feature vector to feature vectors corresponding to the core ECMO parameters. The second fully connected layer is responsible for command output, converting the feature vector into specific parameter values ​​through the Sigmoid activation function to generate preliminary pulsation parameter commands.

[0168] The initial instructions cover all core adjustable parameters, each with a specific value, such as blood flow rate 5.2 L / min, pulse rate 78 beats / min, pulse amplitude 22 mmHg, centrifugal pump speed 2000 r / min, and oxygen concentration 80%. It also outputs the adjustment direction (upward / downward / maintain) and adjustment range (e.g., increasing pulse amplitude by 2 mmHg) for each parameter, ensuring the completeness of the initial instructions.

[0169] The generation of preliminary instructions follows the correlation rules of the global feature vector. For example, if the weight of the brain oxygen adaptation feature in the global feature vector is high and indicates insufficient brain perfusion, the pulsation amplitude and blood flow in the preliminary instructions will tend to be increased to meet clinical physiological needs.

[0170] S93. Correct the preliminary parameter instructions according to the ECMO clinical safety constraints to generate the final pulsation parameter instructions.

[0171] ECMO clinical safety constraints were introduced to verify and revise preliminary parameter instructions one by one, ensuring the safety and clinical feasibility of the instructions. The core safety constraints, based on ECMO clinical guidelines, cover the safety ranges and adjustment limits for all core parameters, specifically including: blood flow rate safety range 4-8 L / min, with a single adjustment not exceeding 0.5 L / min; pulse rate safety range 60-120 beats / min, with a single adjustment not exceeding 5 beats / min; pulse amplitude safety range 10-30 mmHg, with a single adjustment not exceeding 2 mmHg; centrifuge pump speed safety range 1500-3000 r / min, with a single adjustment not exceeding 100 r / min; and oxygen concentration safety range 50%-100%, with a single adjustment not exceeding 10%.

[0172] During the verification process, if the parameters in the initial command exceed the safe range (e.g., blood flow rate of 8.2 L / min), they are automatically corrected to the critical value of the safe range (e.g., 8.0 L / min). If the single adjustment exceeds the limit (e.g., the pulsation amplitude is increased by 3 mmHg), it is corrected to the maximum safe adjustment range (e.g., 2 mmHg). At the same time, combined with the individualized physiological demand vector, the corrected command is further optimized to ensure that the command fits the individual patient's baseline characteristics (e.g., the blood flow rate of elderly patients is not lower than 4.5 L / min after correction), avoiding a one-size-fits-all correction method.

[0173] After safety constraints and individualized modifications, the parameter commands undergo final verification and formatting to generate final pulsation parameter commands containing complete information. This ensures that the commands can be directly used for clinical control of the ECMO device. Specific output includes: 1) Core parameter commands, specifying the final value of each adjustable parameter (e.g., blood flow 5.0 L / min, pulse rate 78 beats / min, pulse amplitude 22 mmHg, rotation speed 2000 r / min, oxygen concentration 80%); 2) Parameter adjustment information, specifying the adjustment direction (increase / decrease / maintain) and adjustment range for each parameter (e.g., increasing pulse amplitude from 20 mmHg to 22 mmHg). The adjustment criteria include: 1) increasing the blood pressure by 2 mmHg; 2) adjusting the blood pressure and monitoring requirements, combining the clinical routine for ECMO parameter adjustment, clarifying the timing of adjustment (e.g., immediate adjustment, gradual adjustment after 5 minutes), and the follow-up time and core monitoring indicators after adjustment (e.g., rechecking cerebral oxygen saturation (rSO2) and arterial blood pressure (ABP) 5 minutes after adjustment to observe the effect of parameter adjustment); 3) command confidence, the reliability of outputting the command (e.g., 0.88), calculated by combining the comprehensive decision confidence and baseline confidence, providing decision-making reference for clinical medical staff. If the confidence is lower than 0.7, it prompts medical staff to further judge based on clinical experience to ensure the safety and reliability of the command.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. An adaptive parameter control system for ECMO pulsatile perfusion systems, characterized in that, include: The multimodal signal reconstruction module is used to clean and reconstruct multimodal signals and extract physiological temporal features. The physiological state analysis module is used to extract the static state and dynamic trend of the cardiovascular system from physiological time-series features, and obtain the cardiovascular system state feature set. The heart-brain signal interaction module is used to generate autonomic nervous balance characteristics that change dynamically over time based on the information transmission intensity between heart and brain signals. The pulsation feature extraction module is used to obtain pulsation features using physical information neural networks and filter neural network accumulation algorithms. The physiological needs analysis module is used to identify the causal relationship between cardiovascular and cerebrovascular status and perfusion parameters, quantify the degree of influence of different parameters on individual status, and generate individualized physiological needs vectors. The collaborative perception feature extraction module is used to integrate the cardiovascular and cerebrovascular state feature set, autonomic nervous balance features, and individualized physiological demand vectors to obtain collaborative perception features. The comprehensive decision confidence analysis module is used to obtain comprehensive decision confidence by fusing autonomic nerve balance characteristics and pulsation characteristics through DS evidence theory; The dynamic regulation baseline generation module is used to generate dynamic regulation baselines based on historical regulation experience data through multi-agent deep reinforcement learning and knowledge distillation. The control command generation module is used to generate the final pulsation parameter command by fusing collaborative sensing features, comprehensive decision confidence and dynamic control baseline through graph neural network.

2. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The process of cleaning and reconstructing multimodal signals and extracting physiological temporal features includes the following steps: Acquire multimodal signals, including arterial blood pressure, heart rate, cerebral oxygen saturation, blood flow, and respiratory rate of ECMO patients; The multimodal signal is denoised by adaptive wavelet thresholding, and the denoised physiological signal is reconstructed by filter neural network. Physiological time-series feature matrix is ​​extracted using the physiological signals.

3. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The process of extracting the static state and dynamic trends of the cardiovascular and cerebrovascular systems from physiological time-series characteristics to obtain a cardiovascular and cerebrovascular state feature set includes the following steps: Temporal convolutional networks are used to extract local state features of the cardiovascular and cerebrovascular systems from physiological temporal features; A bidirectional gated recurrent unit network is introduced, and based on the local state features, a bidirectional temporal feature vector is generated through the cooperative action of forward and reverse directions; By integrating the local state features and the bidirectional temporal feature vector, a cardiovascular and cerebrovascular state feature set is obtained.

4. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The process of generating dynamically changing autonomic balance characteristics over time based on the information transmission intensity between heart and brain signals includes the following steps: Based on information theory, the intensity of bidirectional information transmission between the heart and brain is quantified to obtain the dynamic time-series curve of information transmission between the heart and brain. Construct a heart-brain coupling network and analyze the dynamic curve of heart-brain coupling strength through the heart-brain coupling network; By combining the dynamic time-series curve and the dynamic curve of heart-brain coupling strength, the autonomic nervous balance characteristics that change dynamically over time are generated.

5. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The method of obtaining pulsation features using physical information neural networks and filter neural network accumulation algorithms includes the following steps: A physical information neural network is constructed, and the single-cycle pulsation characteristics are obtained by combining the physical information neural network with the arterial blood pressure wave. A filter neural network accumulation algorithm is introduced to capture the temporal accumulation effect of the single-cycle pulsation characteristics; By combining the single-cycle pulsation characteristics and the cumulative pulsation characteristics, a pulsation feature set is obtained.

6. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The process of identifying the causal relationship between cardiovascular and cerebrovascular status and perfusion parameters, quantifying the influence of different parameters on individual status, and generating an individualized physiological demand vector includes the following steps: A dataset linking features and perfusion parameters was constructed using a cardiovascular and cerebrovascular state feature set, and a causal discovery algorithm was used to identify direct causal relationships between perfusion parameters and state. A parameter sensitivity network is constructed using the direct causal relationship between perfusion parameters and states as edges to quantify the sensitivity coefficient of perfusion parameters to cardiovascular and cerebrovascular states. By combining the aforementioned sensitivity coefficients and individual baseline characteristics, an individualized physiological demand vector is generated.

7. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The process of integrating cardiovascular and cerebrovascular state feature sets, autonomic nervous balance features, and individualized physiological demand vectors to obtain collaborative perception features includes the following steps: Build a dynamic routing mechanism; Based on the aforementioned dynamic routing mechanism, collaborative perception features are obtained by fusing cardiovascular and cerebrovascular state feature sets, autonomic nervous balance features, and individualized physiological demand vectors.

8. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The method of fusing autonomic balance and pulsation characteristics using DS evidence theory to obtain comprehensive decision confidence includes the following steps: Evidence was constructed using autonomic balance and pulsation characteristics; Based on the rule- and data-driven BPA allocation principle, the comprehensive decision confidence level is obtained using the evidence body.

9. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The process of generating a dynamic regulatory baseline based on historical regulatory experience data through multi-agent deep reinforcement learning and knowledge distillation includes the following steps: Construct multiple collaborative agents and define the state space, action space, and reward function. Train an offline MARL model using a historical regulation experience dataset. The decision knowledge of the offline MARL model is transferred to the lightweight online student model using knowledge distillation technology. Based on a lightweight online student model derived from knowledge distillation, and combined with the patient's real-time physiological state, a dynamic regulatory baseline is generated.

10. The adaptive parameter control system for ECMO pulsation perfusion system according to claim 1, characterized in that, The process of generating final pulsation parameter instructions by fusing collaborative sensing features, comprehensive decision confidence, and dynamic control baseline through graph neural networks includes the following steps: By combining collaborative perception features, comprehensive decision confidence, and dynamic control baseline, a dynamic graph structure network is constructed. Based on the global feature vector generated by the dynamic graph structure network, preliminary parameter instructions are generated through inference in the fully connected layer. The initial parameter instructions are corrected according to the ECMO clinical safety constraints to generate the final pulsation parameter instructions.

Images