Extracorporeal circulation data analysis method and system based on deep learning

Abstract

The invention relates to the technical field of deep learning, in particular to an extracorporeal circulation data analysis method and system based on deep learning, and the method comprises the following steps: collecting pump rotating speed, pump flow, arterial perfusion pressure, arteriovenous blood oxygen saturation, hemoglobin and lactic acid, carrying out the time alignment, completing the time sequence coding, and forming state characterization; the method comprises the following steps: correcting a state evolution direction by combining a relationship between oxygen delivery and oxygen extraction, introducing a lactic acid change trend to constrain time sequence consistency, constructing a circulating resistance matching perfusion capability according to flow and pressure, continuously correcting a state change direction, and generating an operation state analysis result. Oxygen delivery and oxygen extraction constraints are introduced to correct state evolution, a lactic acid trend is combined to strengthen time sequence consistency, resistance constraints are formed by perfusion pressure and flow, perfusion capacity changes are continuously corrected, and nonlinear state expression stability and overall interpretation coherence are improved.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method and system for analyzing extracorporeal circulation data based on deep learning. Background Technology

[0002] The field of deep learning technology involves data modeling and feature learning methods with multi-layer neural networks as the core. It typically includes the collection and labeling of large-scale sample data, feature representation construction, network structure design, parameter training and inference processes. Its core purpose is to model the nonlinear correlation in complex data through hierarchical neuron connections. It is widely used in scenarios such as medical data analysis, pattern recognition and predictive computing.

[0003] Traditional extracorporeal circulation data analysis methods and systems refer to the continuous collection of operational data such as blood flow, pump speed, tubing pressure, blood temperature, and oxygenation parameters during the operation of extracorporeal circulation equipment. These methods typically employ manually set thresholds, time series statistics, linear or polynomial regression calculations, and rule-based comparison and judgment to analyze the relationship between various parameters and changes over time. The implementation generally includes time synchronization processing of raw monitoring data, outlier removal, key parameter combination calculations, and result output based on empirical rules to complete the data analysis and judgment of the extracorporeal circulation operation status.

[0004] Current extracorporeal circulation data analysis relies on manual threshold setting and rule comparison. Parameter relationship processing is mainly based on static statistics and linear correlation, which makes it difficult to characterize the coupling characteristics of multiple indicators evolving over time. It also results in a delayed response to the linkage changes between hemodynamics and metabolic state, and abnormal identification is easily affected by experience differences. Parameter fluctuations under complex physiological conditions are simplified, leading to insufficient stability in judging the operational status, discontinuous trend reflection, and low sensitivity to short-term disturbances and gradual changes. This affects the reliability of the analysis conclusions and their reference value for clinical decision-making. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a deep learning-based method for analyzing extracorporeal circulation data, comprising the following steps: S1: Collect pump speed, pump flow rate, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin and lactate concentrations of the extracorporeal circulation machine, align and vectorize them in time, input them into the long short-term memory network temporal encoding, and obtain the initial hidden state vector; S2: Obtain pump flow rate and hemoglobin concentration to calculate oxygen delivery, analyze oxygen extraction amount by arterial and venous blood oxygen saturation, and adjust the direction of the initial hidden state vector according to the ratio of oxygen delivery to extraction amount to obtain the corrected hidden state vector. S3: Monitor the continuous time change of lactic acid concentration, combine it with the oxygen extraction amount of the corrected hidden state vector to judge the trend consistency, use the judgment result to constrain the hidden state of the long short-term memory network and reconstruct it to obtain the time-series corrected hidden state vector. S4: Obtain the pump flow rate and arterial perfusion pressure to calculate the equivalent circulatory resistance. Combine the pump speed to perform resistance matching and adjustment on the hidden states in the time-series correction hidden state vector that are associated with perfusion capacity, and obtain the hemodynamic correction hidden state vector. S5: Based on the equivalent circulatory resistance, determine the consistency of the time change direction of the hemodynamic correction latent state vector and correct the latent state direction to generate circulatory data analysis results.

[0006] As a further aspect of the present invention, the initial hidden state vector includes pump speed time-series features, pump flow rate time-series features, and arterial perfusion pressure time-series features; the corrected hidden state vector includes oxygen delivery ratio features and oxygen extraction ratio features; the time-series corrected hidden state vector includes lactate change trend labels, oxygen extraction trend constraint labels, and hidden state time series after consistency constraints; the hemodynamic corrected hidden state vector includes equivalent circulatory resistance mapping features, perfusion capacity-related hidden state components, and kinetic state representation after resistance matching adjustment; and the circulatory data analysis results include hidden state time direction correction identifiers, resistance consistency constraint sequences, and circulatory state comprehensive judgment vectors.

[0007] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect pump speed, pump flow, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin concentration and lactate concentration generated during the operation of the extracorporeal circulation machine, rearrange the execution order according to a unified timestamp, perform linear interpolation correction on the multi-value sequence, and generate a multi-parameter time-aligned data sequence. S102: Based on the multi-parameter time-aligned data sequence, perform numerical normalization mapping on the multi-parameter sequence, continuously slice it according to a fixed time window length, and concatenate the parameter vectors within the window in chronological order while maintaining index continuity to obtain a time-series feature input sequence; S103: Call the time-series feature input sequence and inject it into the long short-term memory network in time step order. Perform gating state update and state transition operations on the feature vector of each time step, and extract the state output corresponding to the last time step to generate the initial hidden state vector.

[0008] As a further aspect of the present invention, the Long Short-Term Memory network consists of an input gate, a forget gate, an output gate, a cell state, a hidden state, and a set of gating weight parameters.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the initial hidden state vector, obtain the pump flow rate, arterial blood oxygen saturation and hemoglobin concentration, perform time-by-time multiplication on all values ​​under the same time index, arrange the product results in time order and maintain index consistency, and generate an oxygen delivery sequence. S202: Based on the oxygen delivery sequence and the initial hidden state vector, obtain the arterial blood oxygen saturation value and the venous blood oxygen saturation value and perform difference calculation, filter out the time index corresponding values ​​that have not reached the oxygen saturation difference threshold, arrange the retained values ​​in chronological order and maintain continuity, and generate the oxygen extraction sequence. S203: Based on the oxygen delivery sequence and the oxygen extraction sequence, perform ratio calculation on the two types of values ​​under the same time index to form a proportional scalar sequence, call the initial hidden state vector, perform directional weighted adjustment operation on the multi-dimensional components, and generate a corrected hidden state vector.

[0010] As a further aspect of the present invention, the oxygen saturation difference threshold is determined by statistically analyzing the difference sequence formed by arterial blood oxygen saturation values ​​and venous blood oxygen saturation values ​​within the same monitoring period, obtaining the continuous temporal distribution of the difference sequence, performing frequency statistics on multiple difference values ​​in the distribution, and calculating the median value of the difference.

[0011] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Obtain lactic acid concentration values ​​corresponding to multiple time indices within the same monitoring period, perform difference calculation on lactic acid concentrations under adjacent time indices, and arrange the difference results in chronological order to generate a lactic acid change trend sequence; S302: Based on the lactic acid change trend sequence, call the oxygen extraction amount value corresponding to the time index in the correction hidden state vector, perform adjacent time index difference calculation on the oxygen extraction amount to obtain the oxygen extraction change sequence, perform consistency judgment on the two types of change sequences, and establish a trend consistency constraint identification sequence. S303: Based on the trend consistency constraint identifier sequence, perform constrained state transition calculations for the multi-time index components of the corrected hidden state vector, perform numerical reconstruction for components that do not meet the constraint identifier, maintain numerical continuity for components that meet the constraint identifier, and generate a time-series corrected hidden state vector.

[0012] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Obtain pump flow rate and arterial perfusion pressure, perform ratio calculation on the arterial perfusion pressure value and pump flow rate value under the same time index, arrange the ratio results in time order, perform consistency check on the ratio under continuous time index, and generate equivalent circulation resistance sequence. S402: Based on the equivalent circulating resistance sequence, obtain pump speed change data, perform difference calculation on pump speed values ​​under adjacent time indices to obtain speed change sequence, map speed change sequence and equivalent circulating resistance sequence in time and perform deviation judgment to generate resistance matching constraint parameter set; S403: Based on the resistance matching constraint parameter set, call the hidden state parameters associated with perfusion capacity in the time-series correction hidden state vector, perform a restricted adjustment operation on the corresponding parameters, and rewrite the adjusted parameters into the hidden state vector structure to obtain the hemodynamic correction hidden state vector.

[0013] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the hemodynamic correction hidden state vector, obtain the hidden state parameter sequence corresponding to multiple time indices in the vector, perform sign change detection on the parameter values ​​under adjacent time indices, sequentially mark the parameter increase / decrease directions, and generate a hidden state time direction sequence. S502: Based on the hidden state time direction sequence, obtain the equivalent cyclic resistance, perform adjacent difference operation on the resistance values ​​under the same time index, extract the resistance change direction, and perform consistency judgment on the resistance change direction and the hidden state time direction sequence item by item to generate a direction consistency judgment sequence. S503: Based on the direction consistency determination sequence, call the corresponding index hidden state parameter in the hemodynamic correction hidden state vector, perform direction correction calculation on the parameter at the recorded index, and re-aggregate the corrected parameters into a time series structure to generate cyclic data analysis results.

[0014] A deep learning-based extracorporeal circulation data analysis system includes: The timing coding module collects the pump speed, pump flow rate, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin and lactate concentrations of the extracorporeal circulation machine, and aligns and vectorizes them in time. It then inputs the data into the long short-term memory network timing coding to obtain the initial hidden state vector and passes it to the oxygen supply correction module. The oxygen supply correction module acquires pump flow rate and hemoglobin concentration to calculate oxygen delivery, analyzes oxygen extraction amount based on arterial and venous blood oxygen saturation, and adjusts the direction of the initial hidden state vector according to the ratio of oxygen delivery to extraction amount to obtain the corrected hidden state vector and transmit it to the metabolic constraint module. The metabolic constraint module monitors the continuous time change of lactate concentration, judges the trend consistency by combining the oxygen extraction amount of the corrected hidden state vector, and uses the judgment result to constrain the hidden state of the long short-term memory network and reconstruct it to obtain the time-series corrected hidden state vector and pass it to the blood flow matching module. The blood flow matching module obtains the pump flow rate and arterial perfusion pressure to calculate the equivalent circulatory resistance. It then combines the pump speed to perform resistance matching and adjustment on the hidden states in the time-series correction hidden state vector that are associated with perfusion capacity, thereby obtaining the hemodynamic correction hidden state vector and transmitting it to the circulation analysis module. The circulation analysis module determines the consistency of the time change direction of the hemodynamic correction latent state vector based on the equivalent circulation resistance and corrects the latent state direction, generating circulation data analysis results.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by aligning and temporally encoding the operating parameters of multi-source extracorporeal circulation, an intrinsic constraint relationship between oxygen delivery and oxygen extraction is introduced. The direction of implicit state evolution is dynamically corrected, ensuring that the state characterization is consistent with the actual metabolic load. The consistency judgment of temporal sequence is strengthened by combining the trend of lactate change. Furthermore, the relationship between perfusion pressure and flow rate is transformed into a circulatory resistance constraint, enabling continuous correction of changes in perfusion capacity. This improves the ability of operating state analysis to express nonlinear changes and overall trends, and enhances the stability and interpretability of the results. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] Please see Figure 1 This invention provides a deep learning-based method for analyzing extracorporeal circulation data, comprising the following steps: S1: Collect pump speed, pump flow rate, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin and lactate concentrations of the extracorporeal circulation machine, align and vectorize them in time, input them into the long short-term memory network temporal encoding, and obtain the initial hidden state vector; S2: Obtain pump flow rate and hemoglobin concentration to calculate oxygen delivery, analyze oxygen extraction amount by arterial and venous blood oxygen saturation, and adjust the direction of the initial hidden state vector according to the ratio of oxygen delivery to extraction amount to obtain the corrected hidden state vector. S3: Monitor the continuous time change of lactic acid concentration, combine the oxygen extraction amount of the corrected hidden state vector to judge the trend consistency, use the judgment result to constrain the hidden state of the long short-term memory network and reconstruct it to obtain the time-series corrected hidden state vector. S4: Obtain pump flow rate and arterial perfusion pressure to calculate equivalent circulatory resistance. Combine pump speed to perform resistance matching and adjustment on the hidden states in the time-series correction hidden state vector that are associated with perfusion capacity, and obtain the hemodynamic correction hidden state vector. S5: Based on the equivalent circulatory resistance, the consistency of the time change direction of the hemodynamic correction latent state vector is judged and the latent state direction is corrected to generate circulatory data analysis results.

[0021] The initial hidden state vector includes pump speed time-series features, pump flow rate time-series features, and arterial perfusion pressure time-series features. The corrected hidden state vector includes oxygen delivery ratio features and oxygen extraction ratio features. The time-corrected hidden state vector includes lactate change trend labels, oxygen extraction trend constraint labels, and hidden state time series after consistency constraints. The hemodynamically corrected hidden state vector includes equivalent circulatory resistance mapping features, perfusion capacity-related hidden state components, and kinetic state representation after resistance matching adjustment. The circulatory data analysis results include hidden state time direction correction labels, resistance consistency constraint sequences, and circulatory state comprehensive judgment vectors.

[0022] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect pump speed, pump flow, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin concentration and lactate concentration generated during the operation of the extracorporeal circulation machine, rearrange the execution order according to a unified timestamp, perform linear interpolation correction on the multi-value sequence, and generate a multi-parameter time-aligned data sequence. A data transmission channel is established with each sensor through the high-frequency data acquisition interface of the extracorporeal circulation device. This captures in real time the pump speed sensor output, the pump flow rate data from the electromagnetic flowmeter, the arterial perfusion pressure monitored by the pressure transmitter, and the arterial and venous oxygen saturation, hemoglobin concentration, and lactate concentration data acquired by the optical sensor. First, the heterogeneous data streams from these six dimensions are timestamped, extracting the Unix timestamp information carried in the header of each data packet. A quicksort logic is then used to rearrange all data frames in ascending order of timestamp values ​​to ensure the monotonically increasing temporal order of the data. For data gaps caused by inconsistent sensor sampling frequencies or transmission delays, linear interpolation correction is performed. The specific execution logic of linear interpolation correction is as follows: First, locate the valid observation values ​​of the previous and subsequent time steps for the missing data point. Calculate the difference between the value of the subsequent time step and the value of the previous time step, and simultaneously calculate the difference between the timestamp of the subsequent time step and the timestamp of the previous time step. Divide the numerical difference by the timestamp difference to obtain the rate of change per unit time. Then, multiply this rate of change by the time interval between the timestamp of the missing point and the timestamp of the previous time step, and sum the result with the valid observation value of the previous time step to derive the filling value for the missing time step. For example, when monitoring arterial perfusion pressure parameters, the pressure value collected at time 1000 ms is 80 mmHg, the pressure value at time 1002 ms is 82 mmHg, and the data at time 1001 ms is missing. According to the above logic, first calculate the pressure difference as 2 mmHg and the time difference as 2 ms, dividing them to obtain the rate of change per unit time of 1 mmHg. The time interval between 1001 milliseconds and 1000 milliseconds is then calculated to be 1 millisecond. The rate of change 1 is multiplied by the interval 1 to obtain the increment 1. Finally, the increment 1 is added to the baseline value of 80 to obtain the corrected pressure value of 81 mmHg at time 1001 milliseconds. After performing the above operations on all parameter sequences, a multi-parameter data sequence with strictly aligned time axes in each dimension is generated. Some data fragments after specific acquisition and processing are shown in Table 1.

[0023] Table 1. Example of time-aligned data for multiple parameters of a cardiopulmonary bypass machine; S102: Based on the multi-parameter time-aligned data sequence, perform numerical normalization mapping on the multi-parameter sequence, continuously slice it according to a fixed time window length, and concatenate the parameter vectors within the window in chronological order while maintaining index continuity to obtain the time-series feature input sequence; The generated aligned data is used to first perform numerical normalization mapping operations on six dimensions of parameters: pump speed, pump flow rate, arterial perfusion pressure, blood oxygen saturation, hemoglobin, and lactate concentration. The execution logic of the normalization mapping is as follows: traverse the complete time series of a parameter, filter out the maximum and minimum values ​​in the series, calculate the difference between the measured value and the minimum value of the parameter at the current time step, and simultaneously calculate the difference between the maximum and minimum values. Divide the former by the latter to obtain the normalized value falling within the closed interval of 0 to 1. For example, for the pump flow rate parameter, the maximum flow rate in historical observations is set to 10 liters per minute, and the minimum flow rate is set to 0 liters per minute. When the measured pump flow rate at time 1001 milliseconds in step S101 is 4.52 liters per minute, first calculate 4.52-0=4.52, calculate 10-0=10, and then 4.52 / 10=0.452, obtaining the normalized flow rate characteristic value of 0.452. This computational logic eliminates the order-of-magnitude differences between different physical dimensions, accelerating the convergence of subsequent neural network models. After normalization, continuous slicing is performed according to a preset fixed time window length. The time window length parameter is set based on grid search experiments, specifically, traversing and testing within a time step range of 30 to 100 with a step size of 10, selecting the value with the lowest model prediction error rate as the optimal window length. For example, experiments show that the model has the strongest ability to capture abnormal states of extracorporeal circulation when the window length is set to 50 time steps, therefore the window length is fixed at 50. The slicing operation adopts a sliding window mechanism, setting the sliding step size to 1, and sequentially extracting continuous vector groups of length 50 starting from the starting point of the normalized sequence. For each slice window, the values ​​of the 6 parameters contained in the window at 50 time steps are concatenated in chronological order, maintaining the continuity of the indices, forming a temporal feature matrix with a dimension of 50 multiplied by 6. Finally, all the matrices generated by the slices are stacked sequentially to obtain the temporal feature input sequence used for model input.

[0024] S103: Call the temporal feature input sequence, inject it into the long short-term memory network in the order of time steps, perform gating state update and state transition operations on the feature vector of each time step, and extract the state output corresponding to the last time step to generate the initial hidden state vector. The generated temporal feature input sequences are sequentially injected into a Long Short-Term Memory (LSTM) network. This LSTM network consists of an input layer, hidden layers, and an output layer. The input layer has 6 neurons, corresponding to the 6 physiological parameter features. The hidden layer contains 128 neurons to capture high-dimensional temporal dependencies, and the output layer's dimension matches the task objective. Fully connected layers are used for signal transmission, and a Dropout mechanism is introduced with a Dropout ratio of 0.2. This means that 20% of the neuron connections are randomly disconnected during training to prevent overfitting. This Dropout ratio was determined based on multiple comparative experiments. Experimental data shows that when the ratio is set to 0.2, the model's accuracy on the validation set improves by 4.5% compared to when Dropout is not implemented. The specific process of performing gated state update and state transition operations on the feature vector at each time step is as follows: First, the forget gate is calculated. The hidden state vector from the previous time step and the input feature vector from the current time step are weighted and summed using the forget gate weight matrix, and a forget gate bias term is added. The sum is then input into the Sigmoid activation function. The Sigmoid activation function maps the input value to between 0 and 1, and the output determines the proportion of cell state information retained from the previous time step. Specifically, the calculation logic is to multiply the hidden state value from the previous time step by the corresponding weight of the forget gate, multiply the current input feature value by the input weight, add the two together, accumulate the bias value, and then perform a Sigmoid transformation. For example, if a component of the hidden state from the previous time step is set to 0.5, the corresponding weight of the forget gate is 0.8, the normalized value of the current input feature is 0.452, the input weight is 0.6, and the bias value is 0.1. First, calculate 0.5 * 0.8 = 0.4, then calculate 0.452 * 0.6 = 0.2712. Add these two together and then add 0.1 to get a total of 0.7712. Substituting 0.7712 into the Sigmoid function logic (1 divided by 1 plus the natural constant e to the power of -0.7712), the result is approximately 0.684. This value of 0.684 is the forgetting coefficient, indicating that approximately 68.4% of the old state information is retained. Next, the input gate and candidate state calculations are performed. New candidate memory information is generated using the Tanh activation function and combined with the input gate output to update the cell state. Finally, the output gate controls the output of the current hidden state. During the training phase, this network model uses the mean squared error (MSE) as the loss function. This loss function quantifies the error by calculating the square of the difference between the model's predicted value and the true label value, and averaging it over all samples. The optimization algorithm employs the Adam optimizer with an initial learning rate of 0.001. The learning step size for each parameter is dynamically adjusted by calculating the first and second moment estimates of the gradient. After iterative calculations through forward and backward propagation, the state output corresponding to the first time step is extracted as the initial hidden state vector that can characterize the features of the current initiation state of extracorporeal circulation.

[0025] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on the initial hidden state vector, obtain pump flow rate, arterial blood oxygen saturation and hemoglobin concentration, perform time-by-time multiplication operation on all values ​​under the same time index, arrange the product results in time order and maintain index consistency, and generate oxygen delivery sequence. Based on the time step index corresponding to the generated initial hidden state vector, the pump flow rate and hemoglobin concentration sequences are accurately extracted from the generated multi-parameter time-aligned data sequence. A time-by-time multiplication operation is performed, that is, at each time step, the measured pump flow rate is multiplied by the hemoglobin concentration at the same time, thereby quantifying the blood's oxygen delivery potential at that moment. For example, at time 1002 milliseconds during extracorporeal circulation, the pump flow rate is read as 4.55 liters per minute and the hemoglobin concentration as 121 grams per liter. At this point, a multiplication operation is directly performed, 4.55 * 121 = 550.55, calculating the oxygen delivery characteristic value at that moment as 550.55 grams per minute. This calculation process is repeated for all time points in the sequence, generating a series of oxygen delivery characteristic values. Subsequently, these calculated values ​​are arranged according to the original timestamp order, ensuring that the index is strictly aligned with the original multi-parameter sequence, forming an oxygen delivery sequence. The advantage of this operational logic is that by introducing the coupling characteristics of flow rate and oxygen-carrying substance concentration, it can more intuitively reflect the body's oxygen supply basis in the current circulatory state, overcoming the deficiency that a single parameter cannot characterize oxygen supply capacity.

[0026] S202: Based on the oxygen delivery sequence and the initial hidden state vector, obtain the arterial blood oxygen saturation value and venous blood oxygen saturation value and perform difference calculation. Filter out the time index corresponding values ​​that have not reached the oxygen saturation difference threshold, arrange the retained values ​​in chronological order and maintain continuity, and generate the oxygen extraction sequence. The generated oxygen delivery sequence and the output initial hidden state vector are used to synchronously index the arterial and venous oxygen saturation values ​​in the multi-parameter data sequence. First, the difference between the arterial and venous oxygen saturation at the same time point is calculated. Specifically, the venous oxygen saturation is subtracted from the arterial oxygen saturation to obtain the characteristic value of the arteriovenous oxygen partial pressure difference. To eliminate invalid data due to sensor jitter or non-physiological factors, an oxygen saturation difference threshold is introduced for filtering. This threshold is set based on a retrospective analysis of 500 historical cardiopulmonary bypass surgeries, statistically analyzing the distribution range of the arteriovenous oxygen difference under normal metabolic conditions, and setting the lower limit of the distribution range as the threshold. Statistical analysis sets this threshold to 0.15, or 15%. If the calculated difference is less than 0.15, the data point is considered an invalid metabolic feature and is discarded; if the difference is greater than or equal to 0.15, the value is retained. For example, at time 1002 milliseconds, the arterial oxygen saturation is measured as 0.979 (97.9%), and the venous oxygen saturation is measured as 0.725 (72.5%). A subtraction operation is performed to calculate the difference between 0.979 and 0.725, resulting in 0.254. This result, 0.254, is then compared to a preset threshold of 0.15. Since 0.254 is greater than 0.15, this value is considered valid and retained. The retained values ​​are then rearranged according to the original time step order, maintaining the continuity of the time index, to generate an oxygen extraction sequence characterizing the body's oxygen utilization efficiency. The specific calculations and data comparisons are shown in Table 2.

[0027] Table 2. Calculation data table of oxygen transport and oxygen extraction characteristics; As shown in Table 2, through the above calculations, oxygen transport characteristic values ​​reflecting oxygen supply capacity and oxygen extraction characteristic values ​​reflecting oxygen consumption level were obtained, providing a quantitative basis for subsequent physiological correction of the latent state.

[0028] S203: Based on the oxygen delivery sequence and the oxygen extraction sequence, perform ratio calculation on the two types of values ​​under the same time index to form a proportional scalar sequence, call the initial hidden state vector, perform directional weighted adjustment operation on the multi-dimensional components, and generate a corrected hidden state vector. Based on the generated oxygen delivery sequence and the generated oxygen extraction sequence, a ratio calculation is performed on the values ​​at the same time index. The calculation logic involves dividing the oxygen extraction feature value as the numerator and the oxygen delivery feature value as the denominator to obtain a dimensionless supply-demand balance ratio coefficient. This coefficient is then arranged into a scalar ratio sequence in chronological order. Subsequently, the generated initial hidden state vector is invoked, and this scalar ratio sequence is used to perform a directional weighted adjustment operation on the multi-dimensional components of the hidden state vector. The specific logic of the adjustment operation is as follows: First, the scalar ratio is multiplied by a preset sensitivity adjustment coefficient, which is experimentally determined to be 100, aiming to amplify small proportional changes to an order of magnitude capable of affecting the neural network state. Then, the amplified proportional value is incremented by 1 to obtain the adjustment factor. Finally, this adjustment factor is multiplied element-wise by the corresponding component in the initial hidden state vector. For example, taking data at time 1002 milliseconds, the oxygen extraction feature value is 0.254, and the oxygen delivery feature value is 550.55. Performing a division operation, 0.254 / 55 = 0.55, yields a proportional scalar of approximately 0.000461. Next, the adjustment factor is calculated: 0.000461 multiplied by the sensitivity adjustment coefficient of 100, resulting in 0.0461, which is then added to 1, yielding 1.0461. Setting the eigenvalue of one dimension of the initial hidden state vector at this moment to 0.5, a multiplicative correction is performed: 0.5 * 1.0461 = 0.52305, resulting in a corrected hidden state component value of 0.52305. This weighting operation is then performed on all dimensions of the hidden state vector, ultimately generating a corrected hidden state vector containing information on the body's metabolic supply and demand. The experimental results demonstrate that, after introducing physical constraint correction based on oxygen supply and demand balance, the model's accuracy in predicting postoperative complications after cardiopulmonary bypass is improved compared to the traditional pure data-driven LSTM model, effectively enhancing the model's interpretability and robustness at the physiological level.

[0029] Please see Figure 4 The specific steps of S3 are as follows: S301: Obtain lactic acid concentration values ​​corresponding to multiple time indices within the same monitoring period, perform difference calculation on lactic acid concentrations under adjacent time indices, and arrange the difference results in chronological order to generate a lactic acid change trend sequence; First, the storage space of the multi-parameter time-aligned data sequence is accessed, and the full time series data corresponding to the lactic acid concentration parameter is accurately extracted. A time traversal pointer is set, and starting from the second time index position of the sequence, the values ​​at the previous adjacent time index positions are traced back one by one. The difference operation between adjacent time steps is performed: the measured lactic acid concentration value at the current time index is obtained, and the measured lactic acid concentration value at the previous time index is obtained simultaneously. The value at the current time index is used as the minuend, and the value at the previous time index is used as the subtrahend, performing a subtraction operation to obtain the instantaneous change in lactic acid at this time step relative to the previous time step. This operation is performed sequentially on all data points in the sequence except the starting time point, thereby generating a set of differences containing positive and negative signs and numerical values. Subsequently, all the calculated difference results are serialized and arranged strictly according to the timestamp order of the original data to construct a lactic acid change trend sequence. Each numerical point in this sequence precisely characterizes the rate of formation or clearance of anaerobic metabolites within a very short time window, where positive values ​​represent an increase in concentration, negative values ​​represent a decrease in concentration, and zero values ​​represent a constant concentration. For example, according to monitoring records, the lactate concentration was 1.200 mmol / L at 1000 ms, increased to 1.210 mmol / L at 1001 ms, and further increased to 1.220 mmol / L at 1002 ms. To calculate the difference, the data at 1001 ms was processed first, subtracting the value at 1000 ms from 1.200, resulting in a change of 0.010 mmol / L; then, the data at 1002 ms was processed, subtracting the value at 1001 ms from 1.210, also resulting in a change of 0.010 mmol / L. If the lactic acid concentration drops back to 1.215 mmol / L at time 1003 milliseconds, then subtracting 1.220 from 1.215 yields a change of -0.005 mmol / L. The numerical sequence 0.010, 0.010, and -0.005 generated by the above calculation represents the specific constituent elements of the lactic acid change trend sequence within the corresponding time period.

[0030] S302: Based on the lactic acid change trend sequence, call the oxygen extraction amount value corresponding to the time index in the correction hidden state vector, perform adjacent time index difference calculation on the oxygen extraction amount to obtain the oxygen extraction change sequence, perform consistency judgment on the two types of change sequences, and establish a trend consistency constraint identification sequence. The generated lactate change trend sequence is retrieved, and the oxygen extraction amount sequence is simultaneously obtained from the upstream processing stage. First, the difference between the oxygen extraction amount sequence and the identical adjacent time index is calculated; that is, the oxygen extraction feature value of the current time index is subtracted from the oxygen extraction feature value of the previous time index, thus generating the oxygen extraction change sequence. Then, the lactate change trend sequence and the oxygen extraction change sequence under the same time dimension are substituted into the consistency judgment logic. The specific execution process of the consistency judgment is as follows: First, the lactate change value and the oxygen extraction change value under the same time index are obtained separately, and the absolute values ​​of these two values ​​are calculated respectively; then, the two absolute values ​​are added to obtain the total variation amplitude value at that moment; simultaneously, the original lactate change value and the oxygen extraction change value are multiplied, and the correlation between the two change directions is evaluated based on the sign attribute of the product result. Based on this, a minimum variation threshold parameter is introduced, and the calculated total variation amplitude value is compared with the minimum variation threshold. If the total variation amplitude is less than the minimum variation threshold, the data at that moment is determined to be in a low signal-to-noise ratio range, and its corresponding trend consistency constraint flag is set to 0. If the total variation amplitude is greater than or equal to the minimum variation threshold, and the direction of change conforms to the preset physiological logic, its corresponding trend consistency constraint flag is set to 1, and the minimum variation threshold is set to 0.005. For example, at 1001 milliseconds, the lactate change value is 0.010, and at the same time, the oxygen extraction amount changes from 0.260 at 1000 milliseconds to 0.258 at 1001 milliseconds, so the calculated oxygen extraction change value is -0.002. At this time, a consistency judgment is performed: first, the absolute value of 0.010 and the absolute value of -0.002 are calculated, i.e., 0.010 + 0.002 = 0.012, resulting in a total variation amplitude of 0.012. Since 0.012 is greater than the minimum variation threshold of 0.005, the amplitude condition is satisfied, and the trend consistency constraint flag at that moment is set to 1. If, at time 1004 milliseconds, the change in lactate is 0, the change in oxygen extraction is 0, and the total variation is 0, which is less than the threshold of 0.005, then the marker for that time is set to 0. Specific consistency analysis data are shown in Table 3.

[0031] Table 3. Analysis of Consistency Constraints on Metabolic Trends; S303: Based on the trend consistency constraint identifier sequence, perform constrained state transition calculations for the multi-time index components of the correction hidden state vector, perform numerical reconstruction for components that do not meet the constraint identifier, maintain numerical continuity for components that meet the constraint identifier, and generate a time-series correction hidden state vector. Based on the output trend consistency constraint identifier sequence, constrained state transition calculations and numerical reconstruction operations are performed on each time index component in the corrected hidden state vector. Each time step covered by the corrected hidden state vector is traversed one by one, and the constraint identifier value corresponding to that time step is read. When a constraint identifier of 1 is detected at a certain time step, a state preservation strategy is executed, directly using the original corrected hidden state vector value at that moment as the final output without any numerical modification, in order to maintain the original continuity of physiological characteristics. When a constraint identifier of 0 is detected at a certain time step, a numerical reconstruction strategy is executed, calling the weighted moving average logic to smooth the hidden state value at that moment. The specific operation logic of numerical reconstruction is as follows: obtain the finally confirmed hidden state vector value of the previous time step and the corrected hidden state vector value to be processed at the current time step; multiply the value of the previous time step by a preset historical inertia weight, and multiply the value of the current time step by a preset current observation weight; finally, perform an addition operation on these two products, and the sum is the reconstructed hidden state vector value at that moment. The above judgment and reconstruction operations are performed in parallel on all dimensions of the hidden state vector. The historical inertia weight is set to 0.9, and the current observation weight is set to 0.1. This parameter setting has been verified through extensive backtesting with historical data and can effectively suppress high-frequency noise. For example, at time 1004 milliseconds, referring to Table 3, we know that the constraint flag at this time is 0, and it is determined that numerical reconstruction needs to be performed. The value of a certain dimension in the input correction hidden state vector at this time is set to 0.550, while the final output hidden state value of this dimension at time 1003 milliseconds (i.e., the previous time) is 0.520. At this time, the reconstruction operation is performed: first, the value of 0.520 at the previous time is multiplied by the historical inertia weight of 0.9, resulting in 0.468; at the same time, the value of 0.550 at the current time is multiplied by the current observation weight of 0.1, resulting in 0.055; finally, 0.468 + 0.055 = 0.523, and the reconstructed value is calculated to be 0.523. The result 0.523 is closer to the state of the previous time step than the original input 0.550, thus achieving smooth suppression of invalid fluctuations. Conversely, for time step 1002, the constraint flag is 1. If the value of the corrected hidden state input at this time is 0.525, then 0.525 is directly output.

[0032] Please see Figure 5 The specific steps of S4 are as follows: S401: Obtain pump flow rate and arterial perfusion pressure, perform ratio calculation on the arterial perfusion pressure value and pump flow rate value under the same time index, arrange the ratio results in time order, perform consistency check on the ratio under continuous time index, and generate equivalent circulation resistance sequence. First, the generated time-aligned multi-parameter data sequence of the extracorporeal circulation machine is accessed via the data bus to accurately extract the time-series data of the pump flow rate and arterial perfusion pressure parameters. A time-step hydrodynamic impedance conversion operation is then performed. Specifically, for each identical timestamp index, the measured arterial perfusion pressure value at that moment is used as the divisor, and the measured pump flow rate value at that moment is used as the divisor. A division operation is then performed to obtain the equivalent circulation resistance value at that moment. For example, at 1002 milliseconds, if the arterial perfusion pressure is 82.0 mmHg and the pump flow rate is 4.55 L / min, then 82.0 / 4.55 = 18.022, and the equivalent circulation resistance at that moment is calculated to be 18.022 mmHg / L / min. This division operation is repeated for all valid time points in the sequence to generate an original resistance sequence composed of continuous resistance values. Subsequently, a consistency smoothing check is performed on the original resistance sequence to eliminate artifact fluctuations caused by high-frequency noise from the sensors. The specific execution logic of the consistency check is as follows: A sliding window with a length of 5 time steps is set, and the arithmetic mean of the resistance values ​​within the window is calculated. First, the 5 resistance values ​​within the window are summed, and then the sum is divided by 5 to obtain the smoothed equivalent cyclic resistance value corresponding to the center point of the time window. For example, if the original resistance values ​​at 5 consecutive time points are 17.920, 18.022, 18.100, 17.950, and 18.050, these 5 values ​​are first added together to obtain 90.042, and then divided by 5 to calculate the smoothed resistance value of 18.008. This calculation logic eliminates the interference of single-point abrupt changes on the overall trend judgment. The smoothed values ​​are then rearranged according to the original timestamp order to finally generate the equivalent cyclic resistance sequence.

[0033] S402: Based on the equivalent circulating resistance sequence, obtain pump speed change data, perform difference calculation on pump speed values ​​under adjacent time indices to obtain speed change sequence, map speed change sequence and equivalent circulating resistance sequence in time and perform deviation judgment to generate resistance matching constraint parameter set; First, the generated smoothed equivalent circulating resistance sequence is invoked, and the pump speed parameter time series is simultaneously obtained from the source data. A difference calculation is performed on the pump speed data for adjacent time steps. Specifically, the pump speed value at the current time index and the pump speed value at the previous time index are obtained, and the current value is subtracted from the previous value to obtain the instantaneous change in pump speed, thus generating a speed change sequence. Subsequently, a resistance-speed matching deviation analysis model is constructed to quantify whether the current speed adjustment behavior conforms to the general laws of fluid mechanics. The matching deviation calculation logic is as follows: First, the resistance change value at adjacent time steps in the smoothed equivalent circulating resistance sequence is calculated. For example, at time 1002 milliseconds, the smoothed resistance value of 18.008 calculated by S401 is used, and the difference is calculated between this value and the smoothed resistance value of 17.958 (example value) at the previous time (1001 milliseconds). The resistance change value is 18.008 - 17.958 = 0.050. Then, the preset resistance-speed response coefficient (set to 50) is invoked. Multiplying the resistance change by a coefficient yields the theoretical speed adjustment: 0.050 * 50 = 2.50 rpm. If the actual pump speed increases from 2505 to 2508, the actual change is 3.00 rpm. The deviation is calculated as: 3.00 - 2.50 = 0.50, taking the absolute value as 0.50. A deviation normalization mapping logic is introduced (with a tolerance baseline set to 100): 1 - (0.50 / 100) = 0.995. The resistance matching constraint parameter at this moment is then determined to be 0.995. Point-by-point calculations are performed on the sequence to generate the resistance matching constraint parameter set. Specific calculation data is shown in Table 4.

[0034] Table 4. Resistance and Speed ​​Matching Verification Data Table; As shown in Table 4, the data in the table strictly corresponds to the smoothed results output by S401. By introducing fluid dynamics constraints into the data-driven model, the advantage of this operational logic is that it can effectively identify non-physiological speed fluctuations caused by pipe kinks or sensor drift.

[0035] S403: Based on the resistance matching constraint parameter set, call the hidden state parameters related to perfusion capacity in the time-series correction hidden state vector, perform constrained adjustment operations on the corresponding parameters, and rewrite the adjusted parameters into the hidden state vector structure to obtain the hemodynamic correction hidden state vector. The generated set of resistance matching constraint parameters is invoked, and the output time-corrected hidden state vector is obtained. First, the feature components representing perfusion capacity in the hidden state vector are located using a feature attention mapping mechanism. Since the hidden state feature distribution of the Long Short-Term Memory (LSTM) network has high-dimensional entanglement, a weighted projection is performed on the 128-dimensional hidden state vector using a pre-trained attention weight matrix or feature mask to identify the feature subspace strongly correlated with "perfusion capacity," rather than relying on fixed neuron indices. A restricted adjustment operation is performed on the state values ​​identified through attention weighting. Specifically, at each time step, the resistance matching constraint parameter corresponding to that moment is obtained as a confidence factor. A multiplicative weighting operation is performed, directly multiplying the hidden state value in the corresponding feature subspace by the resistance matching constraint parameter to obtain the corrected hidden state component value. For example, at time 1003 milliseconds, Table 4 shows that the resistance matching constraint parameter is 0.920, indicating a lag in the rotational speed response. The perfusion feature component value extracted through the attention mechanism at this point (e.g., 0.600) is obtained, and a correction calculation is performed: 0.600 * 0.920 = 0.552. This decrease directly reflects the reduction in effective perfusion capacity due to the drag-speed mismatch. The corrected value is then written back into the hidden state vector structure through inverse projection or masking to generate the final hemodynamically corrected hidden state vector. The experimental results show that after introducing constraint correction based on hydrodynamic deviation, the model's sensitivity in identifying hidden fault states such as high drag and low flow is significantly enhanced.

[0036] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the hemodynamic correction of the hidden state vector, obtain the hidden state parameter sequence corresponding to multiple time indices in the vector, perform sign change detection on the parameter values ​​under adjacent time indices, sequentially mark the direction of parameter increase or decrease, and generate a hidden state time direction sequence. Based on the generated hemodynamically corrected latent state vector, which stores multi-dimensional physiological characteristic data during extracorporeal circulation in the form of a time series, a time traversal pointer is set, starting from the second time index position of the sequence, and backtracking the parameter values ​​at the previous adjacent time index positions one by one. Sign change detection is performed independently for each dimension in the latent state vector. Specifically, the measured value of the latent state parameter at the current time index is obtained, and the value of the latent state parameter at the previous time index is also obtained. The value at the current time index is used as the minuend, and the value at the previous time index is used as the subtrahend, and a subtraction operation is performed to obtain the instantaneous change of that dimension at the current time step. A small fluctuation filtering logic is introduced, setting the zero-value judgment dead zone range to -0.000001 to +0.000001. If the calculated instantaneous change falls within the dead zone, the state of that dimension is determined to have not changed significantly, and the corresponding direction marker is set to 0. If the change is greater than positive 0.000001, it is determined to be a positive reinforcing trend, and the direction marker is set to 1. If the change is less than negative 0.000001, it is determined to be a negative decaying trend, and the direction marker is set to negative 1. The above detection is performed in parallel on all dimensions of the vector, thereby generating a hidden state time direction sequence of the same dimension as the original vector. For example, for the 12th dimension, the value is 0.550000 at 1002 milliseconds and changes to 0.550002 at 1003 milliseconds, a change of 0.000002, which is greater than the upper limit of the dead zone, so it is marked as 1. The change in the 13th dimension is 0.0000005, so it is marked as 0. Transforming continuous fluctuations at the numerical level into discrete trend features provides a standardized input format for subsequent logical consistency checks.

[0037] S502: Based on the hidden state time direction sequence, obtain the equivalent cyclic resistance, perform adjacent difference operation on the resistance values ​​under the same time index, extract the resistance change direction, and perform consistency judgment on the resistance change direction with the hidden state time direction sequence item by item to generate a direction consistency judgment sequence. The generated hidden state time direction sequence is invoked, and the equivalent cyclic resistance sequence is synchronously obtained from the upstream data stream. First, the difference between adjacent time steps is performed on the equivalent cyclic resistance sequence, i.e., the resistance value at the current moment is subtracted from the resistance value at the previous moment, and the direction of resistance change is extracted based on the positive or negative attribute of the difference (positive = 1, negative = -1, static = 0). Then, a consistency logic judgment is performed between the direction of resistance change and the direction marker of each dimension in the hidden state time direction sequence. The judgment logic is based on the negative feedback mechanism of fluid mechanics: when the resistance change direction is 1 (resistance increases) and the direction marker of the hidden state parameter (representing perfusion capacity) is -1 (capacity decreases), or when the resistance is -1 and the hidden state is 1, it is determined that the two conform to the physiological compensation law, i.e., "reverse change" is a consistent state; conversely, if the two direction signs are the same, it is determined to be a non-physiological abnormal coupling. Based on this judgment result, a direction consistency judgment sequence is generated, where time points conforming to physiological laws are marked as "consistent," and those not conforming are marked as "conflicting." For example, at time 1005 milliseconds, the circulation resistance increases from 18.0 to 18.2 (direction 1), while the parameter of the 64th dimension in the hidden state vector decreases from 0.65 to 0.62 (direction -1). These two are inversely related, therefore the directional consistency of this dimension is determined to be "consistent." By introducing physical constraints, the potential "spurious correlation" feature in the data-driven model is effectively identified.

[0038] S503: Based on the direction consistency determination sequence, call the corresponding index hidden state parameter in the hemodynamic correction hidden state vector, calculate the corrected hemodynamic hidden state vector for the parameters at the recorded index, and re-aggregate the corrected parameters into a time series structure to generate cyclic data analysis results. Based on the output direction consistency determination sequence, the corresponding hidden state parameters in the hemodynamic correction hidden state vector are called, and deep numerical correction and reconstruction operations are performed on the parameters at the recorded index. The aim is to utilize the joint constraints of global historical trajectory and local direction features to calculate a correction index that truly reflects the system's steady state. The formula used is: ; : indicates the first A hemodynamic latent state correction score at a time index is a scalar value used to quantify the overall deviation of the current state from the ideal physiological model. : Represents the index of the current processing time step, for example , The total dimension of the hidden state vectors is set to 128 based on the LSTM network structure. : indicates the first Under the i-th time index, the th hidden state vector The initial measured values ​​for each dimension, derived from the output of step S403. : indicates the first The target value for orientation correction in each dimension is obtained based on the following logic: if the orientation of that dimension is determined to be "consistent" in step S502, then... The value is equal to (That is, the current value is accepted); if it is determined to be a "conflict", then The value is the moving average of the past 5 time steps for this dimension, used as the correction benchmark. : indicates the first The importance weight coefficients for each dimension are pre-set based on feature sensitivity analysis. The key perfusion dimensions (such as dimensions 64-70) are set to 0.05, the other auxiliary dimensions are set to 0.005, and the sum of all weights is normalized to 1.0. The weight adjustment factor is modified to introduce a non-linear penalty mechanism as an exponential term. Here, it is set to 2.0, meaning the penalty for large deviations increases quadratically, thus highlighting the impact of significant anomalies. : Indicates the length of the historical time window used to calculate the global error, set to 50 milliseconds. : indicates the first The ideal target state trajectory value at a given time. This value is obtained by matching the optimal reference path value through a knowledge base containing data from 5000 standard surgical cases using the Dynamic Time Warping (DTW) algorithm. : indicates the first The average actual state at each moment (within the historical window), calculated in a practical example: setting the current time index to... To simplify the calculation, three representative dimensions of the hidden state vector are selected ( Demonstration will be conducted. 1. Parameter acquisition and assignment: Dimension 1 (physiological consistency): Because the directions are the same, Weight Dimension 2 (Physiological Conflict): Due to conflicting directions, the historical average is used. Weight (Key Dimension). Dimension 3 (Physiological Consistency): Because the directions are the same, Weight Parameter settings: Global Error Term: Set History Window Calculations show that 50 historical moments... The sum is 0.20. 2. Calculation of local deviation term (Sum part): For dimension 1: . Regarding dimension 2: After weighting, it becomes Squaring Regarding dimension 3: . The summation result is: 3. Global error term calculation (Sqrt part): Mean square calculation: , square root: Final result calculation: The advantage of this formula lies in constructing a spatiotemporal joint constraint mechanism by productizing and coupling "instantaneous logical conflicts" with "cumulative deviations from historical trajectories." The score is only corrected when a physiologically abnormal mutation occurs at the current moment (leading to an increase in the Sum term), and the overall trajectory has already deviated from the standard prognostic path (leading to an increase in the Sqrt term). This significantly increases the risk level, accurately identifying high-risk clinical conditions and avoiding false alarms caused by relying on a single threshold. The result of 0.000632 indicates that while there is a single-dimensional conflict at the current moment, the overall historical trajectory shows a very high degree of consistency, and the risk is within a controllable low range. Specific parameter configurations are shown in Table 5.

[0039] Table 5. Corrected calculation parameters for cyclic data analysis; As shown in Table 5, this calculation process provides a double-verified quantitative indicator for the final analysis of the cyclic data, ensuring the robustness and accuracy of the output results.

[0040] Please see Figure 7 A deep learning-based extracorporeal circulation data analysis system includes: The timing coding module collects the pump speed, pump flow rate, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin and lactate concentrations of the extracorporeal circulation machine, and aligns and vectorizes them in time. It then inputs the data into the long short-term memory network timing coding to obtain the initial hidden state vector and passes it to the oxygen supply correction module. The oxygen supply correction module obtains pump flow rate and hemoglobin concentration to calculate oxygen delivery, analyzes oxygen extraction amount based on arterial and venous blood oxygen saturation, adjusts the direction of the initial hidden state vector according to the ratio of oxygen delivery amount to extraction amount, obtains the corrected hidden state vector and transmits it to the metabolic constraint module. The metabolic constraint module monitors the continuous time change of lactate concentration, judges the trend consistency by combining the oxygen extraction amount of the corrected hidden state vector, and uses the judgment result to constrain the hidden state of the long short-term memory network and reconstruct it to obtain the temporal corrected hidden state vector and pass it to the blood flow matching module. The blood flow matching module obtains the pump flow rate and arterial perfusion pressure to calculate the equivalent circulatory resistance. It combines the pump speed to perform resistance matching and adjustment on the hidden states in the time-series correction hidden state vector that are associated with perfusion capacity, and obtains the hemodynamic correction hidden state vector and transmits it to the circulation analysis module. The circulation analysis module judges the consistency of the time change direction of the hemodynamic correction latent state vector based on the equivalent circulation resistance and corrects the latent state direction, generating circulation data analysis results.

[0041] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A deep learning-based method for analyzing extracorporeal circulation data, characterized in that, Includes the following steps: S1: Collect pump speed, pump flow rate, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin and lactate concentrations of the extracorporeal circulation machine, align and vectorize them in time, input them into the long short-term memory network temporal encoding, and obtain the initial hidden state vector; S2: Obtain pump flow rate and hemoglobin concentration to calculate oxygen delivery, analyze oxygen extraction amount by arterial and venous blood oxygen saturation, and adjust the direction of the initial hidden state vector according to the ratio of oxygen delivery to extraction amount to obtain the corrected hidden state vector. S3: Monitor the continuous time change of lactic acid concentration, combine it with the oxygen extraction amount of the corrected hidden state vector to judge the trend consistency, use the judgment result to constrain the hidden state of the long short-term memory network and reconstruct it to obtain the time-series corrected hidden state vector. S4: Obtain the pump flow rate and arterial perfusion pressure to calculate the equivalent circulatory resistance. Combine the pump speed to perform resistance matching and adjustment on the hidden states in the time-series correction hidden state vector that are associated with perfusion capacity, and obtain the hemodynamic correction hidden state vector. S5: Based on the equivalent circulatory resistance, determine the consistency of the time change direction of the hemodynamic correction latent state vector and correct the latent state direction to generate circulatory data analysis results.

2. The method for analyzing extracorporeal circulation data based on deep learning according to claim 1, characterized in that, The initial hidden state vector includes pump speed time-series features, pump flow rate time-series features, and arterial perfusion pressure time-series features. The corrected hidden state vector includes oxygen delivery ratio features and oxygen extraction ratio features. The time-series corrected hidden state vector includes lactate change trend labels, oxygen extraction trend constraint labels, and hidden state time series after consistency constraints. The hemodynamic corrected hidden state vector includes equivalent circulatory resistance mapping features, perfusion capacity-related hidden state components, and kinetic state representation after resistance matching adjustment. The circulatory data analysis results include hidden state time direction correction identifiers, resistance consistency constraint sequences, and circulatory state comprehensive judgment vectors.

3. The method for analyzing extracorporeal circulation data based on deep learning according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect pump speed, pump flow, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin concentration and lactate concentration generated during the operation of the extracorporeal circulation machine, perform sequential rearrangement according to a unified timestamp, perform linear interpolation correction on the multi-value sequence, and generate a multi-parameter time-aligned data sequence. S102: Based on the multi-parameter time-aligned data sequence, perform numerical normalization mapping on the multi-parameter sequence, continuously slice it according to a fixed time window length, and concatenate the parameter vectors within the window in chronological order while maintaining index continuity to obtain a time-series feature input sequence; S103: Call the time-series feature input sequence and inject it into the long short-term memory network in time step order. Perform gating state update and state transition operations on the feature vector of each time step, and extract the state output corresponding to the last time step to generate the initial hidden state vector.

4. The deep learning-based extracorporeal circulation data analysis method according to claim 3, characterized in that, The Long Short-Term Memory (LSTM) network consists of an input gate, a forget gate, an output gate, a unit state, a hidden state, and a set of gating weight parameters.

5. The deep learning-based extracorporeal circulation data analysis method according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Based on the initial hidden state vector, obtain the pump flow rate, arterial blood oxygen saturation and hemoglobin concentration, perform time-by-time multiplication on all values ​​under the same time index, arrange the product results in time order and maintain index consistency, and generate an oxygen delivery sequence. S202: Based on the oxygen delivery sequence and the initial hidden state vector, obtain the arterial blood oxygen saturation value and the venous blood oxygen saturation value and perform difference calculation, filter out the time index corresponding values ​​that have not reached the oxygen saturation difference threshold, arrange the retained values ​​in chronological order and maintain continuity, and generate the oxygen extraction sequence. S203: Based on the oxygen delivery sequence and the oxygen extraction sequence, perform ratio calculation on the two types of values ​​under the same time index to form a proportional scalar sequence, call the initial hidden state vector, perform directional weighted adjustment operation on the multi-dimensional components, and generate a corrected hidden state vector.

6. The deep learning-based extracorporeal circulation data analysis method according to claim 5, characterized in that, The oxygen saturation difference threshold is determined by statistically analyzing the difference sequence formed by arterial and venous blood oxygen saturation values ​​within the same monitoring period, obtaining the continuous temporal distribution of the difference sequence, performing frequency statistics on multiple difference values ​​in the distribution, and calculating the median value of the difference.

7. The method for analyzing extracorporeal circulation data based on deep learning according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Obtain lactic acid concentration values ​​corresponding to multiple time indices within the same monitoring period, perform difference calculation on lactic acid concentrations under adjacent time indices, and arrange the difference results in chronological order to generate a lactic acid change trend sequence; S302: Based on the lactic acid change trend sequence, call the oxygen extraction amount value corresponding to the time index in the correction hidden state vector, perform adjacent time index difference calculation on the oxygen extraction amount to obtain the oxygen extraction change sequence, perform consistency judgment on the two types of change sequences, and establish a trend consistency constraint identification sequence. S303: Based on the trend consistency constraint identifier sequence, perform constrained state transition calculations for the multi-time index components of the corrected hidden state vector, perform numerical reconstruction for components that do not meet the constraint identifier, maintain numerical continuity for components that meet the constraint identifier, and generate a time-series corrected hidden state vector.

8. The method for analyzing extracorporeal circulation data based on deep learning according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Obtain pump flow rate and arterial perfusion pressure, perform ratio calculation on the arterial perfusion pressure value and pump flow rate value under the same time index, arrange the ratio results in time order, perform consistency check on the ratio under continuous time index, and generate equivalent circulation resistance sequence. S402: Based on the equivalent circulating resistance sequence, obtain pump speed change data, perform difference calculation on pump speed values ​​under adjacent time indices to obtain speed change sequence, map speed change sequence and equivalent circulating resistance sequence in time and perform deviation judgment to generate resistance matching constraint parameter set; S403: Based on the resistance matching constraint parameter set, call the hidden state parameters associated with perfusion capacity in the time-series correction hidden state vector, perform a restricted adjustment operation on the corresponding parameters, and rewrite the adjusted parameters into the hidden state vector structure to obtain the hemodynamic correction hidden state vector.

9. The method for analyzing extracorporeal circulation data based on deep learning according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: Based on the hemodynamic correction hidden state vector, obtain the hidden state parameter sequence corresponding to multiple time indices in the vector, perform sign change detection on the parameter values ​​under adjacent time indices, sequentially mark the parameter increase / decrease directions, and generate a hidden state time direction sequence. S502: Based on the hidden state time direction sequence, obtain the equivalent cyclic resistance, perform adjacent difference operation on the resistance values ​​under the same time index, extract the resistance change direction, and perform consistency judgment on the resistance change direction and the hidden state time direction sequence item by item to generate a direction consistency judgment sequence. S503: Based on the direction consistency determination sequence, call the corresponding index hidden state parameter in the hemodynamic correction hidden state vector, perform direction correction calculation on the parameter at the recorded index, and re-aggregate the corrected parameters into a time series structure to generate cyclic data analysis results.

10. A deep learning-based extracorporeal circulation data analysis system, characterized in that, The system is used to implement the deep learning-based extracorporeal circulation data analysis method according to any one of claims 1-9, the system comprising: The timing coding module collects the pump speed, pump flow rate, arterial perfusion pressure, arterial and venous blood oxygen saturation, hemoglobin and lactate concentrations of the extracorporeal circulation machine, and aligns and vectorizes them in time. It then inputs the data into the long short-term memory network timing coding to obtain the initial hidden state vector and passes it to the oxygen supply correction module. The oxygen supply correction module acquires pump flow rate and hemoglobin concentration to calculate oxygen delivery, analyzes oxygen extraction amount based on arterial and venous blood oxygen saturation, and adjusts the direction of the initial hidden state vector according to the ratio of oxygen delivery to extraction amount to obtain the corrected hidden state vector and transmit it to the metabolic constraint module. The metabolic constraint module monitors the continuous time change of lactate concentration, judges the trend consistency by combining the oxygen extraction amount of the corrected hidden state vector, and uses the judgment result to constrain the hidden state of the long short-term memory network and reconstruct it to obtain the time-series corrected hidden state vector and pass it to the blood flow matching module. The blood flow matching module obtains the pump flow rate and arterial perfusion pressure to calculate the equivalent circulatory resistance. It then combines the pump speed to perform resistance matching and adjustment on the hidden states in the time-series correction hidden state vector that are associated with perfusion capacity, thereby obtaining the hemodynamic correction hidden state vector and transmitting it to the circulation analysis module. The circulation analysis module determines the consistency of the time change direction of the hemodynamic correction latent state vector based on the equivalent circulation resistance and corrects the latent state direction, generating circulation data analysis results.

Images