Intraoperative hypotension advanced early warning and vasoactive drug automatic titration method and system

Abstract

The invention discloses an intraoperative hypotension advanced early warning and vasoactive drug automatic titration method. The method comprises the steps that S1, multi-mode physiological signal collection, conveying and time synchronization are conducted on a patient in an operation; s2, performing real-time hypotension prediction based on the multi-head attention network, and outputting three predicted values: a hypotension risk probability, an average arterial pressure predicted value and a short-term average arterial pressure change rate; judging whether to step into a hypotension risk interval and early warning; s3, automatic titration of the closed-loop vasoactive drug: S3.1, closed-loop triggering or exiting is realized according to different conditions; s3.2, performing medicine titration rate increment calculation in a closed-loop mode; and S3.3, carrying out automatic titration and safety protection on the vasoactive drug according to the increment of the drug titration rate. Real-time operation and advanced early warning of the model end side are achieved, a prediction result and vasoactive drug titration form a closed loop, and the system is suitable for various surgical operation scenes needing fine blood pressure management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biomedical engineering, in particular to a method and system for intraoperative hypotension early warning and vasoactive drug automatic titration. BACKGROUND

[0002] For the hemodynamic management during the operation, the prior art has the following shortcomings: 1. Post-threshold alarm mode delays intervention and cannot provide early warning before the real occurrence of hypotension. Most existing monitors use fixed threshold triggering and only alarm when the mean arterial pressure is lower than the set value, which cannot provide minutes-level early warning. The time for the anesthesiologist to intervene is often less than 1 minute, the patient is exposed to low perfusion risk, and the alarm sensitivity is highly related to the specific threshold, which cannot take into account different types of surgery and individual differences.

[0003] 2. Single physiological signal dependence leads to frequent false alarms and missed alarms. Most commercially available early prediction modules only analyze the arterial pressure waveform shape, ignoring complementary information such as cuff blood pressure and skin temperature. When the arterial catheter is flushed, zero drift or waveform distortion, the algorithm performance drops sharply, and false alarms and missed alarms are prominent.

[0004] 3. The prediction result has not yet formed a closed loop with vasoactive drug infusion. The infusion rate of vasoactive drugs such as phenylephrine is still manually adjusted, which has cognitive load, dose jumping and reaction lag, and it is difficult to immediately correct at the initial stage of blood pressure fluctuation. Existing closed-loop systems mostly use fixed gain feedback regulation or preset logic table, which lack dynamic adaptive ability to different patients' pharmacokinetic differences, often leading to sudden dose changes, blood pressure overshoot or oscillation, and it is difficult to maintain smooth and stable hemodynamics.

[0005] 4. Multi-source data is difficult to fuse in real time and lacks interpretable output The sampling frequencies of arterial pressure, cuff pressure and skin temperature are different, and the interface standards are different. Existing solutions mostly rely on offline synchronization or external hardware boxes to complete fusion. The end-side real-time operation cost is high, the sample size of the public algorithm is limited, and there is a lack of feature contribution visualization, which is not conducive to clinical trust and immediate physician takeover.

[0006] The statements herein merely provide background technology related to the present application, and do not necessarily constitute the prior art. SUMMARY

[0007] The purpose of the present application is to provide a method and system for intraoperative hypotension early warning and vasoactive drug automatic titration to solve the problems of inability to provide early warning, single physiological signal dependence leading to frequent false alarms and missed alarms, prediction result not yet forming a closed loop with vasoactive drug infusion, and lack of real-time operation and interpretability of deep models, suitable for various surgical scenarios that require fine blood pressure management.

[0008] In order to achieve the above object, the present application provides an intraoperative hypotension early warning and automatic vasopressor titration method, comprising: S1, patient intraoperative multi-modal physiological signal acquisition, transmission and time synchronization; S2, real-time hypotension prediction based on multi-head attention network, outputting three prediction values: hypotension risk probability, mean arterial pressure prediction value and short-term mean arterial pressure change rate; determining whether to step into a hypotension risk interval and warning; S3, closed-loop automatic vasopressor titration, comprising the following steps: S3.1, closed-loop triggering or exiting according to different conditions; S3.2, drug titration rate increment calculation in closed-loop mode; S3.3, automatic vasopressor titration and safety protection according to drug titration rate increment.

[0009] Optionally, the step S1 specifically comprises the following steps: S1.1, collecting arterial pressure waveform morphology, cuff blood pressure and skin temperature three physiological signals; S1.2, analog-to-digital conversion of the three physiological signals; S1.3, time synchronization and integrity check; specifically, aligning the three physiological signals with hardware timestamps, ensuring that the time synchronization error of the three types of data is not more than 5 ms; calculating the hash check value of the three types of data for integrity check.

[0010] Optionally, the step S2 specifically comprises the following steps: S2.1, data is sent to the multi-head attention network after online preprocessing; S2.2, multi-head attention network real-time inference arterial pressure value; S2.3, multi-task joint training and verification of multi-head attention network; S2.4, explainable output; S2.5, setting an interface with closed-loop automatic vasopressor titration, determining whether to step into a hypotension risk interval and warning.

[0011] Optionally, the step S2.2 specifically comprises: The multi-head attention network adopts a multi-task design of six-layer self-attention encoder and double-branch decoder, each layer of self-attention encoder contains 8 attention heads and 256-dimensional hidden dimensions, and outputs global features after residual connection and layer normalization, then the general features enter two parallel decoding branches; Two parallel decoding branches include a classification branch and a regression branch; the classification branch is activated by two layers of feedforward network and Sigmoid function, and outputs the probability prediction of future 180-second mean arterial pressure value below the threshold, i.e. the probability of hypotension risk ; the regression branch outputs the 180-second mean arterial pressure value and the change rate of future 30-second mean arterial pressure value .

[0012] Optionally, the step S2.3 is specifically: The multi-head attention network adopts a previously collected perioperative physiological signal dataset for model training and verification, the dataset covers different types of operations and various patient characteristics, and the entire dataset is divided into a training set, a verification set and a test set, and in the training process, a weighted multi-task loss function is used:

[0013] Among them, is the total loss value; is the real label of future short-term hypotension risk, is the hypotension risk probability output by the model; is the real value of mean arterial pressure 180 seconds from the current time, is the predicted value of mean arterial pressure 180 seconds from the current time; represents the real value of the change rate of mean arterial pressure 30 seconds from the current time, is the predicted value of the change rate of mean arterial pressure 30 seconds from the current time; is the cross-entropy used to optimize the classification branch; MSE is the mean square error used to optimize the regression branch; α, β, γ>0 are dimensionless weights used to balance the relative contribution of the prediction task, which can be adjusted according to the performance of the verification set; The step S2.4 is specifically: In the same millisecond cycle in which the inference is completed, the weighting coefficients of all attention heads in the six-layer self-attention encoder are summarized to obtain a channel weight vector, and the channel weight vector is visually displayed.

[0014] Optionally, the step S2.5 is specifically: Three core results are written into a shared memory queue at a frequency of refreshing once per second and are visually displayed; the three core results are: the hypotension risk probability ; the target pressure difference ; and the short-term trend ; wherein, is the target value, is the mean arterial pressure value 180 seconds from the current time, Rate of change of mean arterial pressure in the next 30 seconds; A closed-loop controller is set to monitor the shared memory queue, when If both of the two consecutive values are higher than 0.60 within one second interval and the target pressure difference exceeds the threshold, the system determines that the patient has entered the low blood pressure risk interval, the closed-loop controller enters the active intervention state, and a pre-warning is given through visual display.

[0015] Optionally, the step S3.1 specifically includes: A closed-loop controller is set to read the low blood pressure risk probability every second and the predicted value of mean arterial pressure; The closed-loop triggering condition is that when If both of the two consecutive values are higher than 0.60 within one second interval and the target pressure difference exceeds the threshold, the system determines that the patient has entered the low blood pressure risk interval, and automatically switches to the closed-loop mode; the target pressure difference is the difference between the target value and the predicted value of mean arterial pressure; A gray zone observation period is set: if is between 0.30 and 0.60, enter the 15-second observation state, and if rises to ≥ 0.60, the closed-loop is entered immediately; if it falls to < 0.30, the closed-loop is not triggered; There are two conditions for exiting the closed-loop: one is falls to 0.30 or below and lasts for 30 seconds; the other is that the predicted value of mean arterial pressure is 5 mmHg higher than the target value for 30 consecutive seconds; after either condition is met, the closed-loop controller is frozen, and manual monitoring is restored.

[0016] Optionally, the step S3.2 specifically includes the following steps: S3.2.1, in the closed-loop mode, the closed-loop controller calculates the rate increment of vasoactive drugs using a proportional-differential speed regulation formula:

[0017] wherein, is the rate increment of vasoactive drugs, is the current target pressure difference, is the rate of change of mean arterial pressure in the next 30 seconds, and the proportional coefficient k p is linearly interpolated between 0.01 and 0.03 according to the patient's age, weight, and baseline vascular tone; the differential coefficient k d decreases with the absolute value of the target pressure difference; S3.2.2, the is smoothed, and then compared with the double threshold that the single rate change does not exceed ±20% and the cumulative change of thirty consecutive seconds does not exceed ±50%; if it exceeds the limit, it is truncated or executed in segments according to the limit rule.

[0018] Optionally, the step S3.3 is specifically: According to the rate increment of the vasoactive drug, whether the dose exceeds the upper limit of the dose; if it exceeds the limit, refuse to execute and return an overdose error code, and alarm at the same time; if it does not exceed the limit, execute the vasoactive drug titration operation; wherein the upper limit of the dose for adults is 0.5 µg·kg-¹·min-¹, and for children, it is proportionally and linearly reduced according to the weight; At the same time, the closed loop link sets three tripping conditions: first, if the new prediction is not updated within five seconds, it is determined that the communication is lost, and the closed loop controller immediately freezes the rate and alarms; second, if any input channel is marked as serious distortion for more than fifteen seconds, the system automatically switches to manual mode; third, the operator can force the closed loop to exit by pressing the pause closed loop button for two seconds. After exiting the closed loop, the titration rate remains the last stable value.

[0019] The application also provides an intraoperative hypotension early warning and vasoactive drug automatic titration system for realizing the intraoperative hypotension early warning and vasoactive drug automatic titration method. The system comprises a workstation, a data hub, an acquisition module, a drug titration module and a terminal. The acquisition module is connected with the data hub and is used for acquiring multiple physiological signals and transmitting the physiological signals to the data hub. The data hub is also connected with the workstation and is used for performing analog-to-digital conversion on the physiological signals acquired by the acquisition module and then transmitting the converted data to the workstation. The drug titration module is connected with the workstation and comprises a micro pump. The micro pump receives a vasoactive drug rate increment instruction, checks whether the dose exceeds the upper limit of the dose, executes a vasoactive drug titration operation according to the vasoactive drug rate increment instruction if the dose does not exceed the upper limit of the dose, and is frozen at the last rate after exiting the closed loop. The terminal is a touch terminal and is connected with the workstation and is used for data visualization display and human-computer interaction. The workstation is provided with a central processing unit, a neural network acceleration chip and an encrypted solid state disk. A multi-head attention network model is constructed in the workstation. The workstation is used for processing data, predicting blood pressure and controlling drug titration.

[0020] Compared with the prior art, the application has at least the following beneficial effects: Through the multi-head attention network and the closed loop control, intraoperative hypotension can be predicted in advance, the exposure time of hypotension can be significantly shortened through instant speed regulation, and real preventive hemodynamic management can be realized.

[0021] By synchronously inputting the invasive arterial pressure, the cuff intermittent blood pressure and the peripheral skin temperature three signals into the same multi-head attention network and simultaneously giving the low blood pressure risk probability, the future 180-second mean arterial pressure prediction value and the future 30-second pressure value change rate three results in one frame of inference, the prediction accuracy is greatly improved, and the false alarm and missed alarm frequency phenomenon is avoided.

[0022] By directly embedding the target differential pressure predicted by the model into the proportional-differential formula, and allowing the proportional coefficient and differential coefficient to be dynamically interpolated according to the patient's weight, age, vascular tension and real-time differential pressure, AI feedforward and online adaptive integrated drug delivery are realized for the first time in clinical practice. Compared with fixed gain, it can significantly reduce differential pressure overshoot and rate oscillation.

[0023] Inference is performed on local hardware without relying on the cloud. The entire model is compressed to less than 6 MB, the edge inference latency is 60 ms, the power consumption of the workstation is less than 60 W, it can run directly in a regular operating room socket, meet the real-time requirements of the operating room, and avoid the risks of hospital network and data privacy.

[0024] By setting dual limits at the drug delivery level (single dose rate ±20% and cumulative dose ±50% over 30 seconds) and at the system level (5 seconds of no prediction, 15 seconds of signal distortion, and 2 seconds of manual press-and-hold), it ensures that no communication, interruption, or hardware failure will lead to continuous overdose, forming a multi-layered insurance system that combines software and hardware, and meets the latest medical device safety regulations. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the intraoperative hypotension early warning and vasoactive drug automatic titration system of the present invention; Figure 2 This is a flowchart of the intraoperative hypotension early warning and vasoactive drug automatic titration method of the present invention. Detailed Implementation

[0026] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates the intraoperative hypotension early warning and automated titration method for vasoactive drugs proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the embodiments of this invention. Please refer to the accompanying drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes to aid those skilled in the art and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0027] This invention provides a method for intraoperative hypotension early warning and automated titration of vasoactive drugs. This method is implemented based on an intraoperative hypotension early warning and automated titration system for vasoactive drugs, such as... Figure 1As shown, the system includes a workstation, a data transceiver, a data acquisition module, a drug titration module, and a terminal. The data acquisition module, connected to the data transceiver, includes a pressure sensor, a monitor, and a small infrared skin temperature probe, used to acquire physiological signals and transmit them to the data transceiver. The data transceiver is also connected to the workstation, used to perform analog-to-digital conversion on the physiological signals acquired by the data acquisition module, and then transmit the converted data to the workstation. The drug titration module, connected to the workstation, includes a micro-pump for injecting vasoactive drugs into the patient and allowing for adjustment of the drug titration rate at any time. The terminal is a touch-screen terminal connected to the workstation, used for data visualization and human-computer interaction. The workstation is equipped with a central processing unit, a neural network acceleration chip, and an encrypted solid-state drive. The neural network acceleration chip is specifically designed to accelerate the computational power of deep neural networks, and a multi-head attention network model is constructed in the workstation. The workstation is used for data processing, blood pressure prediction, and drug titration control.

[0028] The method for intraoperative hypotension early warning and automated titration of vasoactive drugs includes the following steps: S1. Intraoperative multimodal physiological signal acquisition, transmission, and time synchronization for the patient, specifically including the following steps: S1.1 Positioning and placement of the acquisition module: Acquire three physiological signals: arterial pressure waveform morphology, cuff blood pressure, and skin temperature, and transmit them to the data transfer unit.

[0029] Specifically, after entering the operating room, the patient lies supine with both arms slightly abducted at 15 degrees to facilitate non-invasive cuff fixation and radial artery puncture. First, a disposable invasive pressure catheter is inserted into the radial artery and connected to a zero-drift calibrated pressure sensor in the acquisition module, with the sampling frequency set to 200 Hz to fully record the details of the arterial pressure waveform. Then, an automatically inflatable cuff is wrapped around the upper arm, and the cuff blood pressure is tested. The monitor in the acquisition module is set to measure pressure every three minutes by default, which can be adjusted to one to five minutes depending on the characteristics of the surgery. Finally, a small infrared skin temperature probe in the acquisition module is fixed to the skin surface of the center of the distal phalanx of the index finger to collect skin temperature, with the sampling frequency set to 1 Hz.

[0030] Optionally, if no invasive arterial access is established in the operating room, a high-sampling-rate photoplethysmography (PPG) probe can be attached to the subject's fingertip or earlobe. The PPG scheme can be implemented in two ways, one of which can be chosen: First, using a PPG probe with an integrated temperature channel (e.g., a thermistor or miniature infrared sensor unit housed within the PPG housing). This integrated PPG probe can simultaneously acquire pulse conduction time (PTT), pulse waveform profile, and skin temperature, thus simultaneously replacing both the pressure sensor and the small infrared skin temperature probe. Second, using a PPG probe containing only an optical channel, used solely to replace the pressure sensor for acquiring PTT and pulse waveform profile; skin temperature is still acquired independently by the small infrared skin temperature probe. Feature mapping of the acquired PTT and pulse waveform profile allows for real-time recalculation of the equivalent mean arterial pressure curve, which, along with skin temperature and cuff blood pressure recorded synchronously via an automatically inflating cuff, is then fed into a data concatenation unit.

[0031] S1.2 Data transmission.

[0032] Specifically, the data transceiver performs analog-to-digital conversion on the three physiological signals acquired in step S1.1, and then outputs the converted three types of data to the bedside workstation through a high-speed data interface.

[0033] The workstation is equipped with a central processing unit, a neural network acceleration chip, and an encrypted solid-state drive, with a total power consumption controlled below 60 watts, allowing direct connection to the standard power supply in the operating room. The workstation runs a security-hardened real-time operating system that automatically performs a self-test of all acquisition channels upon startup. If any acquisition channel loses connection for more than 5 seconds, an alarm will sound simultaneously on the terminal and in the corridor, prompting technical personnel to intervene.

[0034] S1.3 Time synchronization and integrity verification.

[0035] Specifically, the time synchronization involves ensuring millisecond-level alignment of multi-source data. The data transceiver timestamps the data packets of each of the three physiological signals the instant it receives them. Upon receiving the data, the workstation rewrites the data to the system clock and calculates the difference between the system clock and the hardware timestamp. If the time deviation of any acquisition channel exceeds 5 ms, the terminal displays a yellow pop-up indicating a synchronization error and marks the frame as of secondary quality. If synchronization cannot be restored for 30 consecutive seconds, the system pauses and requires manual intervention.

[0036] Specifically, the integrity verification involves writing all collected raw data into the circular buffer of the encrypted solid-state drive in real time, and the central processing unit calculating the hash verification value every ten minutes to prevent data tampering or frame loss. S2. Real-time hypotension prediction based on multi-head attention network, specifically including the following steps: S2.1 Online data preprocessing, specifically including the following steps: S2.1.1 After completing time synchronization and integrity verification, the data is reconstructed into a sliding window with a duration of 20 seconds and written to the onboard high-speed memory (such as video memory or equivalent high-bandwidth memory) of the neural network acceleration chip for subsequent processing. The sliding window is updated in 1-second increments to ensure the response and continuity to the latest physiological signals.

[0037] S2.1.2 Process the three types of data separately; Specifically, within each sliding window, the arterial pressure waveform is resampled to 100 Hz and subjected to a second-order Butterworth low-pass filter to remove high-frequency noise; the cuff blood pressure retains the most recent reading in the unmeasured segment and adds a missing measurement mask so that the multi-head attention network can identify time gaps; the skin temperature signal is smoothed by a first-order moving average and then incorporated into the feature vector along with its first-order difference.

[0038] The three types of raw data were time-aligned and processed as described above, then decomposed into 6 raw channels. Based on these, 6 derived features were calculated, forming a total of 12 input channels. All channels were zero-mean normalized before entering the model and packaged within the computing unit of the neural network acceleration chip. The preprocessing time for a single batch was approximately 15 ms. The 12 input channels consist of: 1. Six raw channel items: continuous invasive arterial pressure raw waveform P(t); cuff systolic pressure SBP_NIBP, cuff diastolic pressure DBP_NIBP, cuff mean pressure MAP_NIBP (last measurement held at zero order); skin temperature T_skin; cuff measurement status flag NIBP_flag (0 / 1); 2. Six derived features: pulsatile mean arterial pressure MAP_bb based on continuous arterial pressure; pulsatile pressure PP_bb = SBP_AP − DBP_AP; pulsatile heart rate HR_bb (calculated from the interval between adjacent pulsatile beats); MAP short window slope dMAP / dt_20s (slope of linear regression on a 20-second window); maximum systolic rate of rise dP / dt_max; enhancement index AIx (ratio of dicrotic amplitude to main wave amplitude).

[0039] The tensors of the 12 input channels multiplied by 20 seconds and 100 Hz are divided into 100 time steps and 12 feature dimensions in the onboard high-speed memory, and then fed into the downstream multi-head attention network at once to balance short-term dynamic capture and edge resource constraints.

[0040] All of the above processing is driven by a neural network acceleration chip; Data streams with a total latency controlled within a latency threshold of 100ms can be sent to a multi-head attention network for real-time hypotension prediction. The total latency refers to the sum of time taken from the acquisition of raw data to time synchronization and integrity verification and then to preprocessing. The latency threshold (100ms) is strictly set based on clinical needs and physiological response mechanisms.

[0041] S2.2, Multi-head attention network for real-time inference of arterial pressure values.

[0042] Specifically, the multi-head attention network adopts a multi-task design with a six-layer self-attention encoder and a dual-branch decoder. Each layer of the self-attention encoder contains 8 attention heads and 256 hidden dimensions. After residual connection and layer normalization, it outputs global features, and then the general features enter two parallel decoding branches.

[0043] The two parallel decoding branches include a classification branch and a regression branch. The classification branch is activated by a two-layer feedforward network and a sigmoid function, outputting a probability prediction of the mean arterial pressure value being below a threshold in the next 180 seconds, i.e., the probability of low blood pressure risk. The prediction and regression branches then output the mean arterial pressure value after 180 seconds via two layers of linear networks. Rate of change of mean arterial pressure over the next 30 seconds After 8-bit quantization, the file size of this multi-head attention network model can be compressed to less than 6 MB, and the time for a single inference on the edge is about 60 ms, which fully meets the real-time prediction requirements of the operating room.

[0044] Optionally, the six-layer self-attention encoder can be equivalently replaced by a set of depthwise separable convolutional gated recurrent units or sparse Transformers, provided that the edge inference time is no more than 60ms and simultaneously outputs a triplet of risk probability, 180s average pressure value, and 30s pressure value change rate. Any model structure that satisfies the above multi-task output and latency within 60ms falls within the protection scope of this invention.

[0045] S2.3 Multi-task joint training and validation of multi-head attention networks.

[0046] Specifically, in an offline environment, the multi-head attention network of this invention uses a previously collected perioperative physiological signal dataset for model training and validation. This dataset covers different surgical types and various patient characteristics. The entire dataset is divided into training, validation, and test sets, split according to common proportions, and cross-validation is used to evaluate the model's robustness. During training, a weighted multi-task loss function is used:

[0047] in, This represents the total loss value. To provide a true label for the risk of short-term hypotension in the future P represents the probability of low blood pressure output by the model; P is the true mean arterial pressure (MAP, mmHg) over the next 180 seconds from the current moment. This is a predicted value for the mean arterial pressure over the next 180 seconds. This represents the true value of the average rate of change of arterial pressure (mmHg / s) over the next 30 seconds. This is a predicted value for the rate of change of mean arterial pressure over the next 30 seconds. Cross-entropy is used to optimize the classification branch; mean squared error is used to optimize the regression branch; α, β, γ>0 are dimensionless weights used to balance the relative contributions of the prediction task and can be adjusted according to the performance on the validation set.

[0048] Model training incorporates an early stopping strategy and learning rate decay to prevent overfitting. During online deployment, the system's prediction performance under different surgical scenarios is dynamically monitored and feedback is provided to further fine-tune the model's hyperparameters, ensuring generalization ability and prediction reliability in real clinical settings.

[0049] S2.4 Interpretable output, specifically including the following steps: S2.4.1 Within the same millisecond cycle of completing forward inference, the central processing unit summarizes the weighted coefficients of all attention heads in the six-layer self-attention encoder so that the operator can clearly understand which physiological information the model emphasizes in each decision.

[0050] Specifically, the two-dimensional weight matrix of each attention head on the 12 input channels is first summed along the time dimension, and then the percentage distribution is normalized within the same layer and the same head. Then, the total of 48 one-dimensional weight vectors obtained from the eight attention heads of the six layers are accumulated item by item and normalized to the 0-1 interval again, so as to obtain a channel weight vector of length twelve.

[0051] S2.4.2 The channel weight vector is sent to the visualization module of the touch terminal and presented as a static bar chart with gradient color bars. Specifically, channels with weights higher than 0.15 are automatically darkened, and a percentage label is given at the top of the high-weight bars. If the operator taps any channel on the screen, the interface will display the raw waveform of the corresponding signal for the most recent two minutes. If the weight of a channel drops sharply (more than 50%) in two consecutive predictions, a yellow prompt will pop up at the top of the terminal: "Please check the signal quality of channel XX," helping the operator to promptly troubleshoot common problems such as misinsertion, cuff leakage, or probe detachment.

[0052] All weight vectors and prompts are synchronously written to the log file, which can be used for signal quality auditing and algorithm iteration training after the operation.

[0053] S2.5. Set up an interface for automatic titration of closed-loop vasoactive drugs to determine whether the patient has entered the low blood pressure risk zone and issue an early warning.

[0054] Specifically, three core results are written to a shared memory queue at a refresh rate of once per second. The workstation is equipped with memory, and the shared memory queue exists in memory. The three core results are: probability of low blood pressure risk. Target pressure difference ( For the target value, (Mean arterial pressure value over the next 180 seconds); short-term trend (Mean arterial pressure change rate over the next 30 seconds); these three core results are simultaneously sent to the terminal's visualization module. A closed-loop controller is installed within the workstation, which monitors the shared memory queue. If the mean arterial pressure is above 0.60 twice consecutively with a one-second interval, and the mean arterial pressure is more than 5 mmHg lower than the target value in the next 180 seconds, the system determines that the patient has entered the low blood pressure risk zone, and the closed-loop controller enters the active intervention state, that is, the system enters the closed-loop mode; at the same time, the terminal pops up a red warning to issue an alert.

[0055] S3. The drug titration module, under the control of the closed-loop controller, realizes automatic titration of vasoactive drugs in a closed-loop manner, specifically including the following steps: S3.1. Implement closed-loop triggering or exit based on different conditions; Specifically, in monitoring mode, the closed-loop controller reads the probability of low blood pressure risk from the shared memory queue once per second. Compared with the predicted value of mean arterial pressure, when If the mean arterial pressure is above 0.60 twice in a row with a one-second interval, and the mean arterial pressure is more than 5 mmHg lower than the target value in the next 180 seconds, the system will determine in real time that the patient has entered the low blood pressure risk zone and automatically switch to closed-loop mode.

[0056] To avoid frequent start-stop cycles triggered by noise spikes, the system is configured with a gray-zone observation period: if If the value is between 0.30 and 0.60, first enter a 15-second observation state. During this period, if... If the value rises back to ≥ 0.60, the loop will immediately be closed; if it falls back to < 0.30, the counter will be reset to zero and the loop will not be triggered.

[0057] There are two types of conditions for exiting the closed loop: one is... The target differential pressure threshold is determined by two conditions: first, the pressure drops below 0.30 and remains below it for 30 seconds; second, the mean arterial pressure remains above the target value of 5 mmHg for 30 consecutive seconds. Once either condition is met, the closed-loop controller freezes the micro-pump in the drug titration module at its final rate and displays a green banner at the top of the terminal indicating that manual monitoring has resumed. In this embodiment, the target differential pressure threshold is set to 5 mmHg by default. The surgeon can also change the target differential pressure threshold to 3 or 8 mmHg in the terminal's menu, or directly input a precise value to meet different surgical needs.

[0058] S3.2, Drug titration rate control and parameter adaptation, specifically includes the following steps: S3.2.1 In closed-loop mode, the closed-loop controller uses an improved proportional-derivative speed control formula to calculate the rate increment of the vasoactive drug in the micro-infusion pump for the next step. The formula is as follows:

[0059] In this embodiment, the vasoactive drug is phenylephrine. In the above formula, It is the rate increment of vasoactive drugs. For the current target pressure difference, It is a prediction of the rate of change of mean arterial pressure over the next 30 seconds, with a proportionality coefficient k. p Linear interpolation was performed based on patient age, weight, and baseline vascular tone, within the range of 0.01 to 0.03; differential coefficient k d (t) then decreases with the absolute value of the target pressure difference, when | When | < 3 mmHg, it automatically drops to 0.5 to avoid overcompensating for small fluctuations.

[0060] S3.2.2, Closed-loop controller will The input is processed by a 10-second moving average filter set in the workstation. Then it is compared with a dual threshold of no more than ±20% for a single rate change and no more than ±50% for a cumulative change over 30 seconds. If the limit is exceeded, the output is truncated or segmented according to the amplitude limiting rule, thereby preventing blood pressure fluctuations caused by sudden rate changes.

[0061] S3.3 Micro-pump command reception and safety protection; The final vasoactive drug rate increment command is sent to the stepper motor-driven micro-infusion pump via an isolated 485 bus. Upon receiving the command, the micro-infusion pump firmware immediately checks if the dose exceeds the upper limit. If it does, the pump refuses to execute and returns an overdose error code, while a red warning pops up on the terminal. If the dose does not exceed the limit, the micro-infusion pump performs the vasoactive drug titration operation according to the vasoactive drug rate increment command. The upper limit for adults is 0.5 µg·kg⁻¹·min⁻¹, while for children it is linearly reduced proportionally to body weight.

[0062] In addition to dosage limits, to ensure safety in case of out-of-sync, the closed-loop link is configured with three disconnection conditions: First, if the shared memory queue does not update with a new prediction within five seconds, communication is considered lost, the closed-loop controller immediately freezes the rate and issues an alarm; second, if any input channel is marked red (severe distortion) by the data integrity module for more than fifteen seconds, the system automatically switches to manual mode; third, the operator can forcibly exit the closed loop by pressing and holding the pause closed-loop button on the terminal for two seconds. After exiting the closed loop, the micro-pump speed remains at the last stable value, and a flashing message indicates that the closed loop has stopped until manually restarted. This logic ensures that the prediction and control chain can both smoothly adjust medication using prospective information and instantly return to manual mode when signals are abnormal or clinical needs arise, maximizing both safety and practicality.

[0063] Considering extreme situations, the micro-infusion pump is equipped with a backup lithium battery, which can maintain the rate reduction process for 2 minutes in the event of a mains power outage; when the battery level drops below 20%, the rate will linearly decrease to 0 and an alarm will sound indicating that the battery is depleted, preventing uncontrolled continuous drug administration. If the system runs continuously in a closed loop for 30 minutes and the cumulative drug dosage exceeds 30 µg·kg⁻¹, the system will forcibly exit the closed loop according to quality control rules and require the physician to reassess.

[0064] Optionally, the proportional-derivative speed control formula can be replaced by model predictive control (MPC), adaptive fuzzy logic, or deep reinforcement learning strategies, as long as the closed-loop controller still receives data in real time. and By implementing dual limits on the micro-infusion pump rate—no more than ±20% per dose and no more than ±50% cumulatively over 30 seconds—and setting an upper dose limit, the same smooth drug delivery and safety protection effects as the above-mentioned scheme can be achieved.

[0065] Optionally, the touch terminal can be replaced with a head-up display (HUD), a 4K pendant display, or a combination of voice broadcast and foot confirmation. The system only needs to present the low blood pressure probability, drug administration status, and channel weight heatmap in any perceptible manner within 200 ms, and retain both yellow alerts and red forced exit alarms to be considered an equivalent implementation. Sound, light, or haptic motors can all be used as alternative alarm media without affecting the implementation of the claims of this invention.

[0066] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0067] In the description of this invention, it should be understood that the terms "center," "height," "thickness," "upper," "lower," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0068] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0069] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0070] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for intraoperative hypotension early warning and automated titration of vasoactive drugs, characterized in that, include: S1. Acquisition, transmission and time synchronization of multimodal physiological signals of patients during surgery; S2. Real-time hypotension prediction based on multi-head attention network, outputting three predicted values: hypotension risk probability, mean arterial pressure prediction value, and short-term mean arterial pressure change rate; determining whether the patient has entered the hypotension risk range and issuing an early warning. S3. Automated titration of closed-loop vasoactive drugs, including the following steps: S3.

1. Implement closed-loop triggering or exit based on different conditions; S3.2 Calculate the incremental drug titration rate in closed-loop mode; S3.3 Automatic titration and safety protection of vasoactive drugs based on the incremental titration rate.

2. The method for intraoperative hypotension early warning and automated titration of vasoactive drugs as described in claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1 Acquire three physiological signals: arterial pressure waveform morphology, cuff blood pressure, and skin temperature; S1.2 Perform analog-to-digital conversion on the three physiological signals; S1.3 Time Synchronization and Integrity Verification: Specifically, the three physiological signals are time-stamped using hardware to ensure that the time synchronization error of the three types of data does not exceed 5 ms; hash check values ​​are calculated for the three types of data to perform integrity verification.

3. The method for intraoperative hypotension early warning and automated titration of vasoactive drugs as described in claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1 The data is preprocessed online and then fed into the multi-head attention network; S2.2, Multi-head attention network for real-time inference of arterial pressure values; S2.3 Multi-task joint training and validation of multi-head attention networks; S2.4 Interpretable output; S2.

5. Set up an interface for automatic titration of closed-loop vasoactive drugs to determine whether the patient has entered the low blood pressure risk zone and issue an early warning.

4. The method for intraoperative hypotension early warning and automated titration of vasoactive drugs as described in claim 3, characterized in that, Step S2.2 specifically includes: The multi-head attention network adopts a multi-task design with a six-layer self-attention encoder and a dual-branch decoder. Each layer of self-attention encoder contains 8 attention heads and 256 hidden dimensions. After residual connection and layer normalization, it outputs global features, and then the general features enter two parallel decoding branches. The two parallel decoding branches include a classification branch and a regression branch; the classification branch is activated by a two-layer feedforward network and a sigmoid function, outputting a probability prediction of the mean arterial pressure value being below a threshold in the next 180 seconds, i.e., the probability of low blood pressure risk. The regression branch outputs the mean arterial pressure value after 180 seconds through two layers of linear networks. Rate of change of mean arterial pressure over the next 30 seconds .

5. The method for intraoperative hypotension early warning and automated titration of vasoactive drugs as described in claim 4, characterized in that, Step S2.3 specifically includes: The multi-head attention network is trained and validated using a previously collected perioperative physiological signal dataset. This dataset covers different surgical types and various patient characteristics. The entire dataset is divided into training, validation, and test sets. During training, a weighted multi-task loss function is used. in, This represents the total loss value. To provide a true label for the risk of short-term hypotension in the future The probability of low blood pressure output by the model; The true value of the mean arterial pressure over the next 180 seconds from the current moment. This is a predicted value for the mean arterial pressure over the next 180 seconds. This represents the true value of the rate of change of mean arterial pressure over the next 30 seconds. This is a predicted value for the rate of change of mean arterial pressure over the next 30 seconds. Cross-entropy is used to optimize the classification branch; MSE is the mean squared error used to optimize the regression branch; α, β, γ>0 are dimensionless weights used to balance the relative contributions of the prediction task and can be adjusted according to the performance on the validation set. Step S2.4 specifically includes: Within the same millisecond cycle of inference completion, the weighted coefficients of all attention heads in the six-layer self-attention encoder are summarized to obtain the channel weight vector, and the channel weight vector is then visualized.

6. The method for intraoperative hypotension early warning and automated titration of vasoactive drugs as described in claim 3, characterized in that, Step S2.5 specifically includes: Three core results are written to a shared memory queue at a refresh rate of once per second and then visualized; the three core results are: probability of low blood pressure risk. Target pressure difference Short-term trend ;in, For the target value, The mean arterial pressure value over the next 180 seconds. The rate of change of mean arterial pressure over the next 30 seconds; Configure a closed-loop controller to monitor the shared memory queue, when... If the blood pressure readings are both above 0.60 twice in a row with a one-second interval, and the target differential pressure exceeds the target differential pressure threshold, the system determines that the patient has entered the low blood pressure risk zone. The closed-loop controller enters active intervention mode and issues a warning through visual display.

7. The method for intraoperative hypotension early warning and automated titration of vasoactive drugs as described in claim 1, characterized in that, Step S3.1 specifically includes: A closed-loop controller is set up, which reads the probability of low blood pressure risk once per second. Compared with the predicted value of mean arterial pressure; The closed-loop trigger condition is: when If the pressure difference is higher than 0.60 twice consecutively at one-second intervals, and the target pressure difference exceeds the target pressure difference threshold, the system determines in real time that the patient has entered the low blood pressure risk zone and automatically switches to closed-loop mode; the target pressure difference is the difference between the target value and the predicted value of mean arterial pressure. Set a gray zone observation period: If If it is between 0.30 and 0.60, enter a 15-second observation state. During this period, if... If the value rises back to ≥ 0.60, the loop will immediately be closed; if it falls back to < 0.30, the loop will not be closed. There are two types of conditions for exiting the closed loop: one is... The following conditions must be met: 1) The mean arterial pressure drops below 0.30 and remains below 0.30 for 30 seconds; 2) The predicted mean arterial pressure is 5 mmHg higher than the target value for 30 consecutive seconds; 3) Once either condition is met, the closed-loop controller freezes and manual monitoring resumes.

8. The method for intraoperative hypotension early warning and automated titration of vasoactive drugs as described in claim 7, characterized in that, Step S3.2 specifically includes the following steps: S3.2.1 In closed-loop mode, the closed-loop controller uses a proportional-derivative speed regulation formula to calculate the rate increment of the vasoactive drug: in, It is the rate increment of vasoactive drugs. For the current target pressure difference, It is the rate of change of mean arterial pressure over the next 30 seconds, with the proportionality constant k. p Linear interpolation was performed based on patient age, weight, and baseline vascular tone, within the range of 0.01 to 0.03; differential coefficient k d (t) decreases as the absolute value of the target pressure difference decreases; S3.2.2, Regarding The data is smoothed and then compared with a dual threshold of no more than ±20% for a single rate change and no more than ±50% for a cumulative change over 30 seconds. If the threshold is exceeded, the data is truncated or segmented according to the amplitude limiting rules.

9. The method for intraoperative hypotension early warning and automated titration of vasoactive drugs as described in claim 8, characterized in that, Step S3.3 specifically includes: The dosage is checked against the rate increment of the vasoactive drug to see if it exceeds the upper limit. If it does, the operation is rejected and an overdose error code is returned, and an alarm is triggered. If it does not exceed the limit, the vasoactive drug titration operation is performed. The upper limit for adults is 0.5 µg·kg⁻¹·min⁻¹, while for children it is reduced linearly according to body weight. Meanwhile, the closed-loop link is set with three disconnection conditions: First, if the new prediction is not updated within five seconds, it is determined that the communication is lost, the closed-loop controller immediately freezes the rate and alarms; second, if any input channel is marked as severely distorted for more than fifteen seconds, the system automatically switches to manual mode; third, the operator can force exit the closed loop by pressing and holding the pause closed-loop button for two seconds; after exiting the closed loop, the titration rate remains at the last stable value.

10. A system for early warning of intraoperative hypotension and automated titration of vasoactive drugs, characterized in that, For implementing the method of intraoperative hypotension early warning and automatic titration of vasoactive drugs as described in any one of claims 1-9, the system includes: a workstation, a data conduit, a data acquisition module, a drug titration module, and a terminal; The acquisition module is connected to the data transceiver and is used to acquire multiple physiological signals and transmit them to the data transceiver. The data transceiver is also connected to the workstation and is used to perform analog-to-digital conversion on the physiological signals acquired by the acquisition module, and then transmit the converted data to the workstation. The drug titration module is connected to the workstation. The drug titration module includes a micro pump. The micro pump receives the rate increment command of the vasoactive drug and checks whether the dose exceeds the upper limit of the dose. If it does not exceed the limit, the vasoactive drug titration operation is performed according to the rate increment command of the vasoactive drug. After exiting the closed loop, the micro pump is frozen at the last rate. The terminal is a touch terminal, connected to the workstation, and used for data visualization and human-computer interaction. The workstation is equipped with a central processing unit, a neural network acceleration chip, and an encrypted solid-state drive. A multi-head attention network model is constructed in the workstation, which is used to process data, predict blood pressure, and control drug titration.

Images