Liver blood flow and anesthesia depth combined monitoring system in liver operation

Abstract

The invention provides a liver blood flow and anesthesia depth combined monitoring system in a liver operation. The liver blood flow and anesthesia depth combined monitoring system comprises seven modules including a multi-source data collector, a signal preprocessor, a physiological coupling model calculator, an operation recognizer, a brain-liver information flow analyzer, a risk prediction early warning device and a clinical decision assist device. According to the system, liver blood flow parameters and anesthesia depth parameters are collected and analyzed in real time, a dynamic coupling relation model between the liver blood flow parameters and the anesthesia depth parameters is established, and in combination with automatic recognition of surgical operation and quantitative analysis of brain-liver information flow, early prediction and early warning of surgical risks are achieved, and individualized clinical decision support is provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical monitoring, and in particular relates to a liver blood flow and anesthesia depth combined monitoring system during liver surgery. Background Art

[0002] Hepatobiliary surgery is characterized by significant tissue damage, high bleeding risk, and significant hemodynamic fluctuations. The liver receives a dual blood supply (80% from the portal vein and 20% from the hepatic artery), with a total blood flow of approximately 1.5 liters per minute. This makes its blood flow regulation complex and highly sensitive to surgical manipulation and anesthetics.

[0003] Blood flow occlusion techniques (such as the Pringle maneuver) are often used during liver surgery to minimize bleeding, but they can result in a decrease in cardiac output of approximately 10% and a 20-30% increase in left ventricular afterload, causing significant hemodynamic fluctuations. Different anesthetics have varying effects on hepatic blood flow: propofol primarily increases total hepatic blood flow by increasing portal venous flow, while sevoflurane has a minimal effect on total hepatic blood flow.

[0004] The main problems with current clinical monitoring technology are as follows: liver blood flow monitoring and anesthesia depth monitoring systems are independent of each other; existing anesthesia depth monitoring has a data processing delay of 15-30 seconds; there is a lack of blood flow and anesthesia depth correlation analysis system specific to the characteristics of liver surgery; the existing system uses standardized reference ranges without considering individual patient differences and surgical type specificity; and artificial intelligence technology is not fully utilized for multi-source data fusion and risk intelligent warning.

[0005] Therefore, it is urgent to develop a new system that integrates liver blood flow monitoring and anesthesia depth monitoring to meet the clinical needs of hepatobiliary surgery. Summary of the Invention

[0006] The present invention provides a combined monitoring system for liver blood flow and anesthesia depth during liver surgery, the system comprising the following components:

[0007] Multi-source data collector: Receives and collects patient liver blood flow-related signals (hepatic artery blood flow, portal vein blood flow, total hepatic blood flow, hepatic artery resistance index, portal vein resistance index) and anesthesia depth-related signals (bispectral index, entropy index, anesthetic drug concentration, EEG raw data), as well as systemic hemodynamic parameters (central venous pressure, mean arterial pressure, cardiac output). The collector digitizes data at a sampling rate of 1000Hz, ensuring a time synchronization accuracy of ±0.1ms. It also supports the ISO / IEEE11073 medical device communication standard, enabling unified acquisition and synchronization of data from multiple heterogeneous medical devices. The data collector transmits the collected data to the signal preprocessor for further processing.

[0008] Signal preprocessor: Performs filtering, denoising and data verification on the original signals transmitted from the multi-source data collector. A 0.5-100Hz bandpass filtering algorithm is applied to the EEG signal to remove baseline drift and high-frequency noise; a 0.01-10Hz bandpass filtering algorithm is applied to the blood flow signal to retain effective physiological fluctuations; and a wavelet transform algorithm is used to eliminate environmental and physiological artifacts. The signal quality index (SQI) is calculated for the processed signal, and signal segments with an SQI lower than 60 are marked or repaired. Time domain features (mean, coefficient of variation, etc.), frequency domain features (frequency band energy distribution, etc.) and time-frequency features (wavelet coefficients, etc.) are extracted from the preprocessed signal. The preprocessed signal features are transmitted to the physiological coupling model calculator and surgical operation identifier for further analysis.

[0009] Physiological coupling model calculator: Based on the signal features provided by the preprocessor, a correlation model between liver hemodynamics and anesthesia depth is constructed. The monitoring parameters are organized into three state vectors: liver blood flow state vector, anesthesia state vector and systemic hemodynamic state vector, and the dynamic correlation between the three vectors is described by a set of differential equations. The liver blood flow-anesthesia depth coupling index (HAACI) is calculated by a weighted nonlinear function, and the weight coefficient is dynamically adjusted according to the individual characteristics of the patient, the stage of surgery and the type of anesthetic drug. The calculator updates the calculation results once per second, provides basic parameters for the risk prediction and early warning device, and supports the dynamic coupling of liver blood flow and anesthesia depth.

[0010] Surgical Procedure Identifier: Based on signal characteristics transmitted by the preprocessor, it identifies the type of surgical procedure being performed. A convolutional neural network extracts spatial signal features and combines this with a long-short-term memory network to analyze temporal variation patterns. This automatically identifies 13 typical hepatobiliary surgical procedures, including Pringle clipping and segmentectomy. The identifier uses a 10-second signal analysis window and outputs recognition results every 2 seconds. These results are then fed into the risk prediction and early warning system and the clinical decision support system as surgical environment parameters.

[0011] Brain-Liver Information Flow Analyzer: Synchronously analyzes EEG and liver blood flow data and calculates the information transfer relationship between the two. The EEG signal is decomposed into five frequency bands: δ, θ, α, β, and γ, and the energy characteristics of each frequency band are extracted. The frequency characteristics of the liver blood flow data are also analyzed. Based on the transfer entropy theory, the information flow intensity in both directions, from EEG activity to liver blood flow and from liver blood flow to EEG activity, is calculated to synthesize the Brain-Liver Information Transfer Index (BLITI). The analyzer updates the calculation results every 10 seconds, and the results are transmitted to the risk prediction and warning device and the clinical decision aid, supporting the innovative quantitative analysis of bidirectional brain-liver information flow.

[0012] The Risk Prediction and Early Warning System integrates the outputs of the Physiological Coupling Model Calculator, the Surgical Procedure Identifier, and the Brain-Liver Information Flow Analyzer to predict potential risks. Based on current physiological status data and the identified surgical procedure type, a dynamic Bayesian network model is applied to predict system state changes within the next 30-60 seconds. The system calculates the Mahalanobis distance between the predicted state and the safe domain to quantify the degree of deviation. Based on the Mahalanobis distance value, the risk level is categorized into four levels: green (safe), yellow (caution), orange (warning), and red (danger). The early warning system updates its risk assessment results every 5 seconds. When the risk level reaches warning or danger, a warning signal is triggered and transmitted to the clinical decision support system.

[0013] Clinical Decision Aid: Generates clinical intervention recommendations based on the output of the risk prediction and early warning system, combined with individual patient characteristics and surgical stage information. The system has a built-in hepatobiliary surgery knowledge base and anesthesia management rule base, automatically matching corresponding anesthesia depth adjustment recommendations, anesthetic drug adjustment recommendations, or blood flow protection measures for different risk types and levels. The generated recommendations include a description of the intervention direction, adjustment range, and expected effects, and are presented to clinicians through structured text and visual graphics. Updated every minute, urgent recommendations (risk level ≥ warning) are pushed instantly.

[0014] Beneficial effects of the present invention:

[0015] 1. Improve surgical safety: The liver blood flow-anesthesia depth coupling model enables early prediction of liver blood flow abnormalities, providing a critical time window for clinical intervention and reducing the risk of complications.

[0016] 2. Optimize anesthetic drug management: The brain-liver information transfer index supports precise management of anesthesia depth, ensuring anesthetic effectiveness while rationally using anesthetic drugs.

[0017] 3. Reduce false alarm interference: Automatic recognition technology for surgical operations distinguishes normal parameter fluctuations from pathological changes, improving the accuracy of early warnings.

[0018] 4. Support individualized monitoring: adjust safety thresholds and warning standards according to patient characteristics and provide customized monitoring solutions.

[0019] 5. Optimize clinical workflow: Multi-source data integration and automatic analysis reduce doctors' workload and provide structured intervention recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 shows a system architecture diagram of the present invention;

[0021] Figure 2 Shown is a system workflow diagram of the present invention. DETAILED DESCRIPTION

[0022] Exemplary embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] like Figure 1 As shown, the present invention provides a combined monitoring system for liver blood flow and anesthesia depth during liver surgery, the system composition is as follows:

[0024] 1. Multi-source data collector

[0025] Designed to receive and collect a variety of physiological signals during liver surgery, the multi-source data collector uses a 1000Hz sampling rate for data digitization, ensuring a time synchronization accuracy of ±0.1ms. It supports the ISO / IEEE 11073 medical device communication standard, enabling unified acquisition and synchronization of data from multiple heterogeneous medical devices.

[0026] Liver blood flow-related signals are acquired by connecting to an ultrasonic Doppler device, including hepatic artery blood flow, portal vein blood flow, total hepatic blood flow, hepatic artery resistance index, and portal vein resistance index. The multi-source data collector is connected to the ultrasonic Doppler device via an RS-232 serial interface, with a baud rate set to 115200 bps, using 8 data bits, 1 stop bit, and no parity check. Acquisition control utilizes a polling method, with the system sending a data request command to the ultrasonic Doppler device every 10 ms. The command format is "SDATA + parameter ID + checksum." Upon receiving the command, the ultrasonic Doppler device returns the corresponding parameter value. The measurement range of hepatic artery blood flow is 100-500 ml / min, with a resolution of 1 ml / min; the measurement range of portal vein blood flow is 500-1500 ml / min, with a resolution of 2 ml / min; the measurement range of total hepatic blood flow is 600-2000 ml / min, with a resolution of 2 ml / min; the measurement range of hepatic artery resistance index is 0.4-1.0, with a resolution of 0.01; the measurement range of portal vein resistance index is 0.1-0.5, with a resolution of 0.01.

[0027] Anesthesia depth-related signals are acquired by connecting an EEG monitor and an anesthesia machine, including the bispectral index, entropy index, anesthetic drug concentration, and EEG raw data. The multi-source data collector is connected to the EEG monitor via a USB interface, and a dedicated driver is used to achieve data communication. Data requests are triggered by interrupts. The system sends a data request interrupt to the EEG monitor, and the EEG monitor returns the EEG data via batch transmission mode after receiving the interrupt. The measurement range of the bispectral index and entropy index are both 0-100, with a sampling rate of 1Hz; the measurement range of anesthetic drug concentration is 0-10μg / ml (intravenous anesthesia) or 0-10% (inhalation anesthesia), with a sampling rate of 1Hz; the sampling rate of EEG raw data is 500Hz, and the bandwidth is 0.5-70Hz.

[0028] Systemic hemodynamic parameters, including central venous pressure, mean arterial pressure, and cardiac output, are acquired by connecting to a multi-parameter monitor. The multi-source data collector is connected to the multi-parameter monitor via an Ethernet interface and communicates using the TCP / IP protocol. The system sends a data request to the multi-parameter monitor every 100 ms, and the request format complies with the HL7 standard. The measurement range of central venous pressure is 0-30 mmHg, with a resolution of 1 mmHg and a sampling rate of 100 Hz; the measurement range of mean arterial pressure is 0-300 mmHg, with a resolution of 1 mmHg and a sampling rate of 100 Hz; the measurement range of cardiac output is 1-20 L / min, with a resolution of 0.1 L / min and a sampling rate of 1 Hz.

[0029] The multi-source data collector uses time-division multiplexing to collect signals with different sampling rates, and the priority of signal collection is: EEG raw data > hemodynamic parameters > liver blood flow parameters > anesthesia depth parameters. The sampling control adopts an interrupt-driven method. The system timer generates an interrupt signal according to the preset sampling rate to trigger the sampling operation of the corresponding channel. A unified timestamp is attached to all collected signals, and the time synchronization adopts the IEEE 1588PTP protocol. The main clock source is the system's built-in temperature-compensated crystal oscillator with a frequency stability of ±0.5ppm. The time synchronization algorithm includes two parts: network delay compensation and temperature drift compensation. Network delay compensation calculates the one-way delay by measuring the request-response round-trip time, and temperature drift compensation adjusts the clock frequency by monitoring the ambient temperature changes, comprehensively achieving a synchronization accuracy of ±0.1ms.

[0030] The collected signals first undergo preliminary processing, including removal of significant artifacts, range checking, and unit conversion. Artifact detection uses a sliding window method with a window length of 10 sampling points. The median and median absolute difference (MAD) are calculated for the data within the window. When the difference between a point and the median exceeds 3 times the MAD, it is identified as an artifact point and corrected using linear interpolation. Range checking is based on the physiological normal range of each parameter, and data outside the preset range is marked. Unit conversion standardizes the measurement units of different devices to standard units as needed, such as pressure in mmHg and flow in ml / min.

[0031] The processed data is encapsulated in a unified data format, and the data packet structure is:

[0032] The packet header identifier (4 bytes) is a fixed value of 0xAA55AA55 and is used to mark the beginning of the data packet. The timestamp (8 bytes) is a 64-bit microsecond count that records the data acquisition time. The device ID (2 bytes) identifies the data source device. The channel ID (2 bytes) identifies the specific parameter type. The data length (4 bytes) identifies the length of the data payload. The data payload (variable length) is the actual parameter value and uses the IEEE 754 floating-point format. The CRC32 checksum (4 bytes) uses the standard CRC32 algorithm to calculate the checksum value of the entire data packet.

[0033] The multi-source data collector implements a dual buffering mechanism internally, including a high-speed buffer (32MB) and a large-capacity buffer (256MB). The high-speed buffer adopts a ring buffer structure, and the data writing and reading operations are controlled by a write pointer and a read pointer. When the distance between the write pointer and the read pointer exceeds 75% of the buffer capacity, a data overflow warning is triggered. The large-capacity buffer adopts a block linked list structure to facilitate dynamic expansion. The buffer management strategy is based on data priority, and EEG and hemodynamic data have higher priority than other data. When the buffer usage rate reaches 80%, the system automatically adjusts the sampling rate and reduces the sampling frequency of low-priority data to ensure that critical data is not lost.

[0034] Data transmission utilizes a reliable TCP / IP-based transport protocol, supporting packet acknowledgment and retransmission mechanisms. Transmission control utilizes a sliding window algorithm, with the window size dynamically adjusted based on network conditions, initially 64 packets. The transmission rate is adaptively adjusted based on a congestion control algorithm, starting at 100 Mbps and reaching a maximum of 800 Mbps. If a packet loss rate exceeding 1% is detected, the system reduces the transmission rate and increases the number of retransmissions. If a network outage lasts longer than 5 seconds, the system automatically activates local storage, saving the data to non-volatile storage for transmission upon network recovery.

[0035] 2. Signal Preprocessor

[0036] The signal preprocessor receives liver blood flow signals, EEG signals and hemodynamic signals, and uses corresponding preprocessing algorithms to improve signal quality and information content.

[0037] The signal preprocessor uses a digital filter to filter various signals. For EEG signals, a fourth-order Butterworth bandpass filter is applied with a passband of 0.5-100 Hz and a roll-off characteristic of 24 dB / octave. The differential equation of the bandpass filter is:

[0038] y(n)=b0x(n)+b1x(n-1)+b2x(n-2)+b3x(n-3)+b4x(n-4)-a1y(n

[0039] -1)-a2y(n-2)-a3y(n-3)-a4y(n-4)

[0040] Where n represents the discrete time point number, x(n) represents the input signal value at time point n, y(n) represents the output signal value at time point n, x(ni) represents the value of the input signal delayed by time points i, and y(ni) represents the value of the output signal delayed by time points i. The filter coefficients are: b0 = 0.0009, b1 = 0.0036, b2 = 0.0054, b3 = 0.0036, b4 = 0.0009 (feedforward coefficients), a1 = -3.8906, a2 = 5.6769, a3 = -3.6848, a4 = 0.8986 (feedback coefficients). The filtering process uses a forward-backward filtering technique, where the signal is first filtered from left to right and then from right to left to eliminate phase distortion and ensure that the waveform characteristics of the physiological signal are not distorted.

[0041] For liver blood flow signals, a fourth-order Butterworth bandpass filter with a passband of 0.01-10 Hz was applied to retain effective physiological fluctuations and remove high-frequency interference and baseline drift. For hemodynamic signals, a fourth-order Butterworth low-pass filter with a cutoff frequency of 5 Hz was applied to remove high-frequency noise. All signals were also subjected to an adaptive notch filter to remove 50 Hz power frequency interference. The transfer function of the notch filter is:

[0042]

[0043] Where H(z) represents the Z-domain transfer function of the system, z -1 represents the unit delay operator, ω0=2π·50 / f s is the digital frequency (radians / sample), 50 represents the power frequency (Hz), f s is the sampling frequency (Hz), r = 0.995 is the pole radius, which controls the notch bandwidth. The closer the r value is to 1, the narrower the notch bandwidth and the better the suppression effect. However, too narrow a notch bandwidth may cause signal distortion. cos represents the cosine function.

[0044] The signal preprocessor uses wavelet transform technology to denoise the signal. First, multi-scale wavelet decomposition is performed, using the db4 wavelet basis function and 5 decomposition layers. Then, soft threshold processing is applied to the wavelet coefficients, and the threshold T is calculated using the formula:

[0045]

[0046] Where T is the threshold, σ is the noise standard deviation, N is the signal length (number of sample points), and ln is the natural logarithm function. The noise standard deviation σ is estimated by the formula:

[0047] σ=MAD / 0.6745

[0048] Wherein, MAD is the absolute median difference of high-frequency coefficients and is calculated as MAD = median(|d-median(d)|), where d represents the high-frequency coefficients of the first wavelet decomposition layer, median represents the median operator, and |·| represents the absolute value operator. The constant 0.6745 is a normalization factor based on the characteristics of the Gaussian distribution and is used to convert the MAD into a standard deviation estimate. Threshold processing uses the soft threshold function:

[0049] η(d)=sgn(d)·max(|d|-T,0)

[0050] Where η(d) is the thresholded wavelet coefficient, d is the original wavelet coefficient, and sgn is the sign function: sgn(d) = 1 when d > 0, -1 when d < 0, and 0 when d = 0. max represents the maximum value operator, and |d| represents the absolute value of d. The processed coefficients are then reconstructed using an inverse wavelet transform, preserving important features while reducing noise.

[0051] The signal preprocessor calculates the signal quality index (SQI) for real-time evaluation. The SQI calculation is based on three dimensions: signal-to-noise ratio, stability, and artifacts:

[0052] SQI=c1·S1+c2·S2+c3·(100-A)

[0053] Where SQI is the signal quality index (range 0-100), c1=0.4, c2=0.4, c3=0.2 are weight coefficients, S1 is the normalized signal-to-noise ratio, S2 is the normalized stability index, and A is the normalized artifact. The calculation formula for S1 is:

[0054]

[0055] Where SNR is the signal-to-noise ratio, expressed in decibels (dB). The formula for calculating S2 is:

[0056]

[0057] Where σ is the standard deviation of the signal, μ is the mean of the signal, σ / μ is the coefficient of variation, which indicates the relative fluctuation of the signal, and min represents the minimum operator. The calculation formula for A is:

[0058]

[0059] Among them, N a is the number of detected artifact points, N is the total number of samples, N a / N represents the ratio of pseudo-error points to the total samples. When SQI < 60, the corresponding signal segment is marked as low-quality data.

[0060] The signal preprocessor uses repair processing for low-quality data segments. For short-term (<1 second) low-quality data, cubic spline interpolation is used for repair, and the interpolation function is:

[0061] S(x)=a i (xx i ) 3 +b i (xx i ) 2 +c i (xx i )+d i ,x i ≤x≤x i+1

[0062] Among them, S(x) is the interpolation function, x is the independent variable (time point), x i and x i+1 is the endpoint of the interpolation interval, a i 、b i 、c i d i is the polynomial coefficient of the i-th segment, which is determined by the values ​​of adjacent data points and the derivative continuity condition. For long-term (>1 second) low-quality data, the autoregressive model prediction method is used to repair it:

[0063]

[0064] Among them, x(n) is the signal value at time point n, p is the model order (usually 10), α i is the i-th order autoregressive coefficient, calculated using the Levinson-Durbin algorithm based on the normal signal segment. x(ni) is the signal value delayed by time points i. ε(n) is the random noise term at time point n, following a Gaussian distribution with mean zero. ∑ represents the summation operator, with the summation ranging from i = 1 to i = p. The repaired data segments are labeled for reference in subsequent analysis.

[0065] The signal preprocessor extracts time domain, frequency domain and time-frequency domain features. The time domain features include mean (μ), standard deviation (σ), coefficient of variation (CV = σ / μ), kurtosis (κ = E[(x-μ) 4 ] / (E[(x-μ) 2 ]) 2 ), skewness (s=E[(x-μ) 3 ] / (E[(x-μ) 2 ]) 3 / 2), etc., where E[·] represents the mathematical expectation operator. Frequency domain features are calculated by fast Fourier transform, including the main frequency (f p =argmaxP(f)), band energy ratio Where P(f) is the power spectrum density, f is the frequency variable, B i is the i-th frequency band, argmax represents the independent variable that makes the function reach its maximum value. The time-frequency characteristics are calculated by continuous wavelet transform:

[0066]

[0067] Where W(a,b) is the wavelet coefficient, a>0 is the scale parameter (inversely proportional to the frequency), b is the translation parameter (corresponding to the time position), x(t) is the signal function, t is the time variable, ψ is the wavelet basis function (using the complex Morlet wavelet), ψ * represents the complex conjugate of ψ, Represents an integral operation from negative infinity to positive infinity. is the normalization factor, tb / a represents the scaling and translation of the wavelet function. The feature extraction adopts the sliding window method with a window length of 10 seconds and an overlap rate of 50%.

[0068] 3. Physiological Coupling Model Calculator

[0069] It is used to construct a correlation model between liver hemodynamics and anesthesia depth, calculate the hepatic blood flow-anesthesia depth coupling index (HAACI), and achieve quantitative evaluation of the effects of anesthetic drugs and surgical operations on liver blood flow.

[0070] The physiological coupling model calculator describes the relationship between liver blood flow and anesthesia depth based on a multi-level state space model. The model contains three state vectors: the liver blood flow state vector F, the anesthesia state vector A, and the systemic hemodynamic state vector H. The liver blood flow state vector is defined as:

[0071] F=[F1,F2,F3,F4,F5]

[0072] Among them, F1 represents hepatic artery blood flow (HAF), the unit is ml / min, and the normal range is 100-400 ml / min; F2 represents portal vein blood flow (PVF), the unit is ml / min, and the normal range is 600-1200 ml / min; F3 represents total hepatic blood flow (THBF), the unit is ml / min, and the normal range is 800-1500 ml / min; F4 represents hepatic artery resistance index (HARI), which is dimensionless and the normal range is 0.55-0.80; F5 represents portal vein resistance index (PVRI), which is dimensionless and the normal range is 0.15-0.30.

[0073] The anesthesia state vector is defined as:

[0074] A=[A1,A2,A3,A4,A5]

[0075] Among them, A1 represents the bispectral index (BIS), ranging from 0 to 100, and the surgical anesthesia target range is 40-60; A2 represents the entropy index (SE / RE), ranging from 0 to 100; A3 represents the power ratio of EEG delta waves to theta waves (δ / θ), which increases with increasing anesthesia depth; A4 represents the power ratio of EEG alpha waves to beta waves (α / β), which decreases with increasing anesthesia depth; A5 represents the anesthetic drug concentration (DrugConc), with propofol units in μg / ml and sevoflurane units in vol%.

[0076] The whole-body hemodynamic state vector is defined as:

[0077] H=[H1,H2,H3,H4,H5]

[0078] Among them, H1 is the mean arterial pressure (MAP), the unit is mmHg, the normal range is 70-105 mmHg; H2 is the cardiac output (CO), the unit is L / min, the normal range is 4-8 L / min; H3 is the central venous pressure (CVP), the unit is mmHg, the normal range is 2-8 mmHg; H4 is the systemic vascular resistance (SVR), the unit is dynes·sec·cm -5 , normal range is 800-1200; H5 represents heart rate (HR), the unit is bpm, and the normal range is 60-100bpm.

[0079] The physiological coupling model calculator uses a set of differential equations to describe the dynamic relationship between three state vectors:

[0080]

[0081] Where t represents the time variable, F(t), A(t), and H(t) represent the hepatic blood flow state vector, anesthesia state vector, and systemic hemodynamic state vector at time t, respectively. O(t) represents the surgical operation state at time t. d / dt represents the differential operator with respect to time. G1, G2, and G3 are state transition functions. The state transition functions are implemented using a recurrent neural network. The network structure consists of two layers of LSTM units, each containing 128 and 64 neurons, respectively. The activation function uses the hyperbolic tangent function (tanh). The network input is the concatenation of the current state vectors and the surgical operation state, and the output is the rate of change of each state vector.

[0082] The physiological coupling model calculator uses the fourth-order Runge-Kutta method to numerically solve the above differential equations. Assuming X = [F, A, H] is the complete system state and G = [G1, G2, G3] is the system dynamics function, the calculation steps of the fourth-order Runge-Kutta method are:

[0083] k1=hG(X n ,O n )

[0084]

[0085]

[0086] k4=hG(X n +k3,O n )

[0087]

[0088] Among them, h represents the solution step size, which is set to 0.1 seconds, X n represents the system state at step n, X n+1 represents the system state at step n+1, and k1, k2, k3, and k4 are intermediate calculation variables.

[0089] The physiological coupling model calculator calculates the HAACI index based on the nonlinear mapping relationship between liver blood flow parameters and anesthesia depth parameters. First, the basic coupling index is calculated:

[0090] HAACI base =w1·f1(F1,A1)+w2·f2(F2,A1)+w3·f3(F3,A5)

[0091] Among them, w1, w2 and w3 are weight coefficients, f1, f2 and f3 are correlation functions. The correlation function between hepatic artery blood flow and BIS is:

[0092] f1(F1,A1)=a1·exp(-b1·|F1-c1·A1-d1|)

[0093] Where a1 = 1.0 is the normalization coefficient, b1 = 0.01 controls the function's sensitivity, c1 = 5.0 and d1 = 100.0 are the linear relationship parameters, exp represents the exponential function, and |·| represents the absolute value operator. The function reaches its maximum value at F1 = c1·A1 + d1, reflecting the ideal relationship between hepatic artery blood flow and BIS.

[0094] The correlation function between portal vein blood flow and BIS is:

[0095] f2(F2,A1)=a2·exp(-b2·|F2-c2·A1-d2|)

[0096] Where a2 = 1.0, b2 = 0.002, c2 = 15.0, and d2 = 300.0 are model parameters. This function describes the ideal relationship between portal vein blood flow and BIS value.

[0097] The correlation function between total hepatic blood flow and anesthetic drug concentration is:

[0098] f3(F3,A5)=a3·(1-exp(-b3·F3 / A5))

[0099] Where a3 = 1.0 and b3 = 0.005 are model parameters. This function describes the nonlinear relationship between total hepatic blood flow and anesthetic drug concentration, reflecting the changing trend of total hepatic blood flow as anesthetic drug concentration increases.

[0100] The weight coefficient is dynamically adjusted according to the type of anesthetic drug and individual patient characteristics:

[0101]

[0102] in, F j (t) including factors such as patient age, body mass index, liver function score, surgical stage code, and type of anesthetic drugs, β ij are the corresponding weight coefficients. This formula is based on the softmax function, ensuring that the sum of the weight coefficients is 1. For propofol anesthesia, the initial weights are set to w1 = 0.4, w2 = 0.4, and w3 = 0.2; for sevoflurane anesthesia, the initial weights are set to w1 = 0.5, w2 = 0.3, and w3 = 0.2.

[0103] The final HAACI index is calculated as:

[0104] HAACI=HAACI base ×CF

[0105] Where CF is the compensation factor, and the calculation formula is:

[0106] CF=1+α·sin(ω·t+φ)·exp(-γ·t)

[0107] Here, α = 0.2 controls the amplitude of the fluctuation, ω = 0.1 controls the frequency of the fluctuation, φ = 0 is the phase, γ = 0.01 controls the attenuation rate, and sin represents the sine function. This compensation factor simulates the attenuation effect of periodic fluctuations during surgery, and t represents the duration of the surgery.

[0108] The physiological coupling model calculator uses a particle filter algorithm to estimate and update model parameters in real time. The particle filter algorithm is based on the principle of sequential importance sampling and represents the posterior probability distribution through a set of weighted samples (particles). The algorithm steps include: particle initialization, state propagation, weight update, importance resampling, and state estimation. The number of particles is set to 1000, and each particle represents a set of possible model parameters. The particle weights are calculated using a likelihood function based on the error between the predicted value and the actual observed value:

[0109]

[0110] in, represents the weight of the i-th particle at the k-th moment, z k is the actual observed value, is the observed value predicted based on the i-th particle, σ is the standard deviation of the observation noise, and ||·|| represents the Euclidean norm. When the particle weight distribution is uneven, resulting in the number of effective particles falling below a threshold (usually half of the total number of particles), a resampling operation is performed, sampling with replacement based on the particle weights to obtain a new set of particles with weights reset to a uniform distribution.

[0111] The physiological coupling model calculator establishes a safety domain for individual patients based on their preoperative baseline data and standard reference values. The safety domain is represented by a multidimensional Gaussian distribution:

[0112]

[0113] Where p(x) is the probability density function, x is the state vector, d is the vector dimension, μ is the mean vector, Σ is the covariance matrix, |Σ| represents the determinant of Σ, (·) T represents the matrix transpose, (·) -1 Represents matrix inversion. The mean vector calculation formula is:

[0114] μ=α·μ baseline +(1-α)·μ standard

[0115] Among them, μ baseline is the patient's preoperative baseline value, μ standard is the standard reference value, and α = 0.7 is the individualization coefficient, which controls the weight of individual characteristics and general standards. The covariance matrix calculation formula is:

[0116] Σ=R·D·R T

[0117] Where R is the correlation matrix, describing the correlation between parameters, and D is a diagonal matrix, with the diagonal elements representing the variance of each parameter. The safety zone parameters are dynamically adjusted based on the surgical stage and patient response:

[0118] μ(t)=μ0+Δμ(t)

[0119] ∑(t)=∑0+Δ∑(t)

[0120] Among them, μ0 and Σ0 are initial parameters, Δμ(t) and ΔΣ(t) are adjustment quantities that change with time.

[0121] 4. Surgical operation identifier

[0122] The system automatically identifies key hepatobiliary surgical procedures by analyzing the temporal variation patterns of multi-source physiological signals, providing the system with surgical environment awareness capabilities. The surgical procedure identifier can identify 13 typical hepatobiliary surgical procedures, including liver mobilization, portal vessel exposure, Pringle clamp initiation, Pringle clamp removal, selective hepatic artery occlusion, middle hepatic vein occlusion, left / right hepatic vein occlusion, liver sectioning, segmental segmentation completion, bile duct resection / reconstruction, major vessel repair, liver section hemostasis, and terminal surgical procedures.

[0123] The surgical operation identifier is based on a hybrid architecture of a two-stream convolutional neural network and a long short-term memory network (CNN-LSTM). The network structure of this hybrid architecture includes four parts: a signal input layer, a feature extraction layer, a sequence analysis layer, and an operation classification layer.

[0124] The signal input layer receives the multi-source physiological signal features extracted by the signal preprocessor. The input data is windowed with a window length of 10 seconds and contains the following feature vectors:

[0125]

[0126] Among them, X t represents the input feature vector at time point t, represents the liver blood flow related eigenvector, represents the anesthesia-related feature vector, Represents the hemodynamic eigenvector. Each sub-eigenvector is defined as follows:

[0127]

[0128]

[0129] Among them, F i (t), A i (t) and H i (t) represent the liver blood flow, anesthesia and hemodynamic parameters at time point t, ΔF i (t) = F i (t)-F i (t-Δt), ΔA i(t) = A i (t)-A i (t-Δt) and ΔH i (t) = H i (t)-H i (t-Δt) represents the change in the corresponding parameter, where Δt = 1 second. The dimension of the input feature vector is 30. Time series data is formed by combining consecutive sample points within a 10-second window. The input data shape is (30×10).

[0130] The feature extraction layer uses a two-stream convolutional neural network structure to process time-domain features and frequency-domain features separately, extracting higher-level feature representations. The time-domain stream processes the original signal features, while the frequency-domain stream processes the signal's spectral features. Spectral features are obtained by applying a short-time Fourier transform to the original features. The convolutional network consists of three convolutional layers, each configured as follows:

[0131] The first convolution layer uses 64 3×7 convolution kernels with a stride of 1, padding of "same", and activation function of ReLU. The first pooling layer uses 2×2 maximum pooling with a stride of 2. The second convolution layer uses 128 3×5 convolution kernels with a stride of 1, padding of "same", and activation function of ReLU. The second pooling layer uses 2×2 maximum pooling with a stride of 2. The third convolution layer uses 256 3×3 convolution kernels with a stride of 1, padding of "same", and activation function of ReLU. The third pooling layer uses 2×2 maximum pooling with a stride of 2. The kernel size is expressed as height × width, and the ReLU activation function is defined as f(x) = max(0, x). The mathematical expression of the convolution operation is:

[0132]

[0133] Among them, Y i,j,k represents the value of the kth channel at position (i, j) of the output feature map, X represents the input feature map, W represents the convolution kernel weight, b represents the bias term, and f h and f w Represent the height and width of the convolution kernel, C in Represents the number of input channels, and s represents the step size. The maximum pooling operation is defined as:

[0134]

[0135] Among them, P i,j,k represents the pooling output, p h and p w Represent the height and width of the pooling window respectively.

[0136] The convolutional features of the time domain stream and the frequency domain stream are merged through the feature fusion layer, and the fusion adopts the channel splicing method:

[0137] F fusion =Concat(F time ,F freq )

[0138] Among them, Concat represents the concatenation operation along the channel dimension, F time and F freq Represent the output features of time domain stream and frequency domain stream respectively.

[0139] The sequence analysis layer uses a bidirectional long short-term memory (BiLSTM) network structure to capture the long-term and short-term dependencies of feature sequences. The BiLSTM consists of two sub-networks: the forward LSTM and the backward LSTM, which process the forward and backward information of the sequence, respectively. The output of the BiLSTM is composed of the concatenation of the hidden states of the forward LSTM and the backward LSTM:

[0140]

[0141] in, represents the hidden state of the forward LSTM, Represents the hidden state of the backward LSTM.

[0142] The sequence analysis layer is configured as a two-layer BiLSTM with 64 LSTM units in each direction. A Dropout layer is added in the middle with a dropout rate of 0.5 to prevent overfitting.

[0143] The operation classification layer uses a fully connected network structure to map the output of BiLSTM to surgical operation categories. The classification network includes:

[0144] First fully connected layer: 256 neurons, ReLU activation function; Dropout layer: dropout rate 0.5;

[0145] The second fully connected layer has 14 neurons (13 surgical operations + 1 non-operation state), and a softmax activation function. The softmax activation function is defined as:

[0146]

[0147] Among them, z i represents the raw score of the i-th category, K represents the total number of categories, and e is the base of the natural logarithm. The softmax function converts the raw score into a probability distribution that satisfies and softmax(z i )∈(0,1). The output of the classification network is the probability of each operation category:

[0148] P(O i |X)=softmax(Wh T +b)

[0149] Among them, P(O i |X) represents the surgical operation under the given signal sequence X i The probability of h T represents the output of the last time step of BiLSTM, W and b are the weight matrix and bias vector of the fully connected layer, respectively.

[0150] The training of the surgical operation identifier adopts a supervised learning method, and the loss function is the cross entropy loss:

[0151]

[0152] Among them, y i represents the true label, Denotes the predicted probability, and n denotes the number of samples. The optimization algorithm uses the Adam optimizer, with an initial learning rate of 0.001, parameters β1 = 0.9, β2 = 0.999, and a learning rate decay of 0.9 times the original rate every 30 cycles. The training cycle is 100. To address the class imbalance problem, a class weighting strategy is adopted, with the weight inversely proportional to the number of class samples:

[0153] w c =N / C·N c

[0154] Among them, w c represents the weight of category c, N represents the total number of samples, C represents the number of categories, N c represents the number of samples of category c. Data augmentation techniques are also used during training, including signal amplitude adjustment (±20%), adding Gaussian noise (SNR 20-30dB), and time shift (±1 second) to enhance the generalization ability of the model.

[0155] The real-time recognition process of the surgical operation identifier is as follows: a sliding window method is used with a window length of 10 seconds and a window sliding step of 0.5 seconds to continuously obtain the latest signal features; the feature vector of each window is preprocessed, including normalization and noise filtering; the processed features are input into a pre-trained CNN-LSTM hybrid network to calculate the probability distribution of each operation category; a decision rule is applied to determine the current surgical operation state, and the recognition threshold θ is used. rec =0.85, when the probability of an operation exceeds the threshold, it is preliminarily identified as the operation; the operation state must last for at least 2 seconds before it is confirmed to avoid misidentification caused by instantaneous fluctuations; there is at least 3 seconds between adjacent different operations to prevent frequent state switching.

[0156] The surgical procedure identifier also automatically divides the surgical process into five stages: preoperative preparation, anesthesia induction, commencement of surgery, major procedures, and conclusion of surgery, based on the identified surgical procedure sequence. This stage division utilizes a state machine model, determining the current surgical stage based on the sequence and time intervals of key procedures. This stage information is fed back to the surgical procedure identifier as contextual features to improve recognition accuracy.

[0157] 5. Analysis of Brain-Liver Information Flow

[0158] It is used to synchronously analyze EEG data and liver blood flow data, quantify the information transmission relationship between the two, calculate the brain-liver information transmission index (BLITI), and provide a theoretical basis and technical support for personalized dosing strategies of anesthetic drugs.

[0159] The brain-liver information flow analyzer receives the EEG and liver blood flow data processed by the signal preprocessor through the data acquisition interface. The EEG data is subjected to time-frequency analysis using the wavelet packet transform method. The wavelet packet transform formula is:

[0160] W j,k,l (t) = 2 j / 2 ψ l (2 j tk)

[0161] Among them, W j,k,l (t) represents the wavelet packet function of the jth layer, with translation parameter k and node number l, ψ l Represents an oscillatory function, t represents a time variable, and j, k, and l are integers. The wavelet packet transform uses the db4 wavelet basis function, with a decomposition layer of 5 layers, resulting in 32 sub-bands. These sub-bands are reorganized into 5 physiologically significant frequency bands: δ (0.5-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (30-100 Hz). The relative power (RP) is calculated for each frequency band, and the power ratios between frequency bands are also calculated, including δ / θ, θ / α, α / β, and (δ+θ) / (α+β). These ratios reflect the different characteristics of the depth of anesthesia.

[0162] Liver blood flow data is adaptively decomposed using the empirical mode decomposition (EMD) method. EMD decomposes the signal into a series of intrinsic mode functions (IMFs) and a residual component:

[0163]

[0164] Among them, x(t) represents the original signal, IMF i (t) represents the i-th IMF component, r n(t) represents the residual component, and n represents the number of IMFs. The first three IMFs are usually selected as eigenmodes, corresponding to the high-frequency, medium-frequency, and low-frequency components, respectively. The energy and energy ratio are calculated for each IMF, and the instantaneous frequency of the IMF is calculated by Hilbert transform, which is defined as:

[0165]

[0166] Among them, H[IMF i (t)] represents IMF i The Hilbert transform of (t) is integrated as the Cauchy principal value integral. The instantaneous phase and frequency are calculated using the time derivative of the phase.

[0167] The brain-liver information flow analyzer converts the continuous signal into a symbol sequence. Using the sorting pattern method, the embedding dimension m = 3 and the time delay τ = 1 are selected to construct the delay vector and sort it by size. The sorting pattern is obtained and mapped into symbols. The symbolized EEG feature sequence and liver blood flow feature sequence are represented as S respectively. Brain (t) and S Liver (t).

[0168] Based on the symbolic sequence, the brain-liver information flow analyzer calculates the transfer entropy from brain to liver and from liver to brain. The basic definition of transfer entropy is:

[0169]

[0170] Among them, TE(X→Y) represents the transfer entropy from variable X to variable Y, H represents the conditional entropy, and Y t+τ represents the value of variable Y at time t+τ, represents the k-order history of Y at time t, represents the l-order history of X at time t, and τ is the time delay parameter. In actual calculations, the k-nearest neighbor (k-NN) estimation method is used to calculate the conditional entropy, and the k-NN estimator is defined as:

[0171]

[0172] Among them, ψ represents the digamma function, N represents the number of samples, n i represents the number of points in Y space that are within the same distance as the k nearest neighbors of the i-th data point in X space. In the k-NN estimation method, k = 4, and the maximum norm is used as the distance metric. Transfer entropy from brain to liver and from liver to brain was calculated separately. The optimal value of the time delay parameter τ was determined by scanning the 1-10 second range. To assess the statistical significance of transfer entropy, a scrambling test was used, with a significance level of p < 0.01.

[0173] The brain-liver information flow analyzer also computes a synergy metric, based on the mutual information:

[0174]

[0175] Where I(Brain;Liver) represents the mutual information between the brain and liver. The final BLITI index is calculated as follows:

[0176] BLITI=α·TE norm (Brain→Liver)+β·TE norm (Liver→Brain)+γ·Synergy norm

[0177] Among them, α, β, and γ are weight coefficients, and different values ​​are set for different types of anesthetic drugs: for propofol anesthesia, α = 0.6, β = 0.3, γ = 0.1; for sevoflurane anesthesia, α = 0.7, β = 0.2, γ = 0.1. TE norm and Synergy norm represents the normalized transfer entropy and synergy metric; Brain→Liver represents the direction of information transfer from EEG activity to liver blood flow; and Liver→Brain represents the direction of information transfer from liver blood flow to EEG activity. The final BLITI index was multiplied by 100 to adjust to the standard range of 0–100.

[0178] The brain-liver information flow analyzer uses a sliding window method with a 60-second window length and a 50% overlap rate, and updates the BLITI index every 10 seconds. A local stationarity test is performed before analysis to ensure that the signal satisfies the weak stationarity assumption within the analysis window. An anesthetic drug delivery efficiency assessment is provided based on the BLITI index's changing trend. When the rate of change exceeds a preset threshold (±0.5 / second), the system determines that the anesthetic drug delivery efficiency has changed significantly and triggers a warning signal.

[0179] 6. Risk Prediction and Early Warning Device

[0180] It is used to integrate the output results of the physiological coupling model calculator, surgical operation identifier and brain-liver information flow analyzer to predict potential surgical risks and provide graded warnings, providing sufficient time window for clinical intervention.

[0181] The risk prediction and early warning device uses a dynamic Bayesian network (DBN) to construct a risk assessment model. The DBN model contains a three-layer network structure, representing the liver blood flow state layer, the anesthesia state layer, and the systemic hemodynamic state layer, totaling 34 nodes. The connections between nodes include intra-layer connections and inter-layer connections within the same time slice, as well as temporal connections between adjacent time slices. The model structure is represented as:

[0182] G=(V,E)

[0183] Where G represents the graph structure, V represents the node set, and E represents the edge set. The node set V includes state nodes and operation nodes:

[0184] V=V F ∪V A ∪V H ∪V O

[0185] Among them, V F Represents the liver blood flow status node set, including 5 nodes; V A Represents the anesthesia state node set, including 5 nodes; V H Represents the whole body hemodynamic state node set, including 5 nodes; V O It represents a set of surgical operation nodes, including 13 nodes, each of which corresponds to a surgical operation and takes a value of 0 or 1. In addition, it also includes 6 derived feature nodes, which represent the changing trends and ratio relationships of key parameters. The parameters of DBN include the conditional probability table (CPT), which is defined as:

[0186] P(V i |Pa(V i ))

[0187] Among them, V i represents the i-th node, Pa(V i ) represents V i The parent node set of . CPT learning adopts a combination of maximum likelihood estimation and Bayesian estimation:

[0188]

[0189] Among them, v ij Represents node V i The jth value of pa k Represents the parent node set Pa(V i )’s kth configuration, N ijk Indicates that V in the training data i =v ij And Pa(V i )=pa k The number of samples, N ik =∑ j N ijk , α ijk Represents the prior parameters. To handle sparse data problems, Laplace smoothing is used:

[0190] α ijk =1 / r i

[0191] Among them, r i Represents node V i The number of possible values.

[0192] The risk prediction and early warning device uses a particle filter algorithm to predict the state. Particle filtering is based on the principle of sequential importance sampling and represents the posterior probability distribution of the state through a set of particles and their weights. The algorithm flow is as follows:

[0193] A) Particle initialization: Generate N particles, N = 1000, each particle represents a possible system state, and the initial weight is set to 1 / N;

[0194] B) State propagation: The particle state is propagated according to the DBN conditional probability model. The formula is:

[0195]

[0196] in, represents the state of the i-th particle at the k-th moment, O k represents the operating state at the kth moment;

[0197] C) Weight update: Update particle weights based on observed data. The formula is:

[0198]

[0199] in, represents the weight of the i-th particle at the k-th moment, Z k represents the observation data at the kth moment, represents the likelihood function, which is calculated as:

[0200]

[0201] Where h represents the observation function, σ represents the standard deviation of the observation noise, and ||·|| represents the Euclidean norm;

[0202] D). Renormalization:

[0203]

[0204] E). Calculate the effective number of particles:

[0205]

[0206] F). Resampling: When N eff <N threshold Resampling is performed when N threshold = N / 2, and replace sampling is performed according to the normalized weight to obtain a new set of particles, and the weight is reset to

[0207] G). State estimation:

[0208]

[0209] The risk prediction and early warning system implements multi-step predictions, with prediction time windows of 30 seconds, 45 seconds, and 60 seconds. The prediction granularity is 5 seconds per step, and the prediction algorithm is a recursive application of single-step predictions, accumulating propagation particle states. The multi-step prediction results are expressed as the predicted mean and 95% confidence interval for each parameter. The confidence interval calculation formula is:

[0210]

[0211] in, represents the predicted mean, σ X represents the forecast standard deviation, calculated as:

[0212]

[0213] The risk prediction and early warning device evaluates the risk status based on the individualized safety domain. The safety domain is represented by a multi-dimensional Gaussian distribution:

[0214]

[0215] Where p(x) represents the probability density function, x represents the state vector, d represents the vector dimension, μ represents the mean vector, Σ represents the covariance matrix, |Σ| represents the determinant of Σ, (·) T represents the transposition operator, (·) -1 represents the matrix inversion operator, and exp represents the exponential function. The formula for calculating the mean vector is:

[0216] μ=α·μ baseline +(1-α)·μ standard

[0217] Among them, α = 0.7 represents the individualization coefficient, μ baseline represents the patient's preoperative baseline value, μ standard Represents the standard reference value. The covariance matrix calculation formula is:

[0218] Σ=R·D·R T

[0219] Where R represents the correlation matrix, D represents the diagonal matrix, and the diagonal elements are the variances of the parameters. The safety zone parameters are dynamically adjusted according to the surgical stage and patient response:

[0220] μ(t)=μ0+Δμ(t)

[0221] Σ(t)=Σ0+ΔΣ(t)

[0222] Among them, μ0 and Σ0 represent the initial parameters, Δμ(t) and ΔΣ(t) represent the adjustment amounts, which change with time.

[0223] The risk prediction and early warning device uses Mahalanobis distance to quantify the degree of deviation between the current state and the safe domain:

[0224]

[0225] Among them, D M represents the Mahalanobis distance, x represents the current state vector, S represents the security domain, and μ S and Σ S Sure, Represents the square root operator. Considering the influence of vector dimension, the normalized Mahalanobis distance is calculated as:

[0226]

[0227] Where d represents the dimension of the state vector. The distances of different prediction time windows are integrated:

[0228]

[0229] Among them, β1=0.5, β2=0.3, and β3=0.2 represent the weight coefficients of different time windows. Risk probability is calculated based on Mahalanobis distance:

[0230]

[0231] Where λ = 0.05 represents the risk coefficient. The risk change rate is also calculated:

[0232]

[0233] Where Δt = 5 seconds. The comprehensive risk score is calculated as:

[0234] RS=0.7·D M +0.3·min(10,5·R′(t))

[0235] Among them, min represents the minimum value operator.

[0236] The risk prediction and early warning device divides risks into four levels according to the Mahalanobis distance value:

[0237] Green (safety): D M ∈[0,2.0), low risk, no intervention required; yellow (caution): D M ∈[2.0,4.0), medium-low risk, need attention; orange (warning): D M ∈[4.0,6.0), medium to high risk, intervention recommended; red (dangerous): D M≥6.0, high risk, intervention is necessary.

[0238] At the same time, the trend warning function is realized. When the risk rising rate exceeds the threshold θ rate = 0.4 / s, a trend warning is triggered and the warning level is raised by one level. When the risk status deteriorates sharply, that is, ΔD M If the risk is >2.0 and R′(t)>1.0 / s, a high-level warning is immediately triggered. Warning signals include visual, audio, and text warnings. For visual warnings, the risk indicator on the interface changes color and flashes to indicate the danger level. For audio warnings, different tones and rhythms correspond to different risk levels. For text warnings, a detailed description of the risk and possible causes are displayed.

[0239] The risk prediction and early warning system updates its risk status assessment every 5 seconds to ensure the system can promptly identify potential risks. The system assessment process includes data acquisition, status update, risk prediction, safety domain calculation, Mahalanobis distance calculation, risk level determination, and early warning signal generation. Data acquisition includes obtaining the HAACI index from the physiological coupling model calculator, the current operation status from the surgical operation identifier, and the BLITI index from the brain-liver information flow analyzer. System status updates are based on the latest observation data and the DBN model, and risk prediction is based on the system's current state and operation mode.

[0240] 7. Clinical Decision Aids

[0241] Based on the output results of the risk prediction and early warning device, combined with the patient's individual situation and surgical stage, targeted anesthesia management and surgical cooperation recommendations are generated to provide decision support for clinicians.

[0242] The clinical decision support system implements decision support functions based on a knowledge base system and a reasoning engine. The knowledge base consists of three components: a hepatobiliary surgery knowledge base, an anesthesia management rule base, and a case library. The hepatobiliary surgery knowledge base contains descriptions, risk points, and precautions for 13 common surgical procedures. It uses an ontology model to organize the knowledge structure, and OWL (Web Ontology Language) is used to describe knowledge entities and their relationships. The anesthesia management rule base contains anesthesia adjustment rules for different risk scenarios, expressed as production rules in the format of "IF <condition> THEN <action>". Rule priority is determined based on the level of evidence and the strength of recommendation. The level of evidence is divided into four levels: A (multicenter randomized controlled trial), B (single-center randomized controlled trial), C (non-randomized study), and D (expert opinion). The strength of recommendation is divided into three levels: I (strong recommendation), II (general recommendation), and III (suggested for consideration). The case library contains anesthesia records of historical surgical cases. Each case includes patient characteristics, surgical type, changes in physiological parameters, anesthesia management measures, and clinical outcomes.

[0243] The clinical decision aid adopts a hybrid reasoning architecture, including rule-based reasoning, case-based reasoning, and model-based reasoning. Rule-based reasoning uses a forward chaining reasoning strategy, starting from the conditions, applying matching rules, and deriving conclusions. The algorithm flow of forward chaining reasoning is as follows:

[0244] A) Initialize working memory to contain currently known facts;

[0245] B) Traverse the rule base and find the rules whose condition part matches the working memory;

[0246] C). If no matching rule is found, the reasoning process is terminated;

[0247] D) If multiple matching rules are found, a rule is selected for execution based on the rule priority;

[0248] E). Execute the action part of the selected rule and update the working memory;

[0249] G). Return to step B and continue reasoning.

[0250] Rule matching is implemented using the Rete algorithm, which reduces the computational complexity of rule matching and improves reasoning efficiency by building a rule network. When the rule is executed, the credibility of the conclusion is calculated based on the rule's certainty factor (CF). The calculation formula for the certainty factor is:

[0251] CF(H,E)=CF(E)×CF(rule)

[0252] Where CF(H,E) represents the certainty factor of hypothesis H given evidence E, CF(E) represents the certainty factor of evidence E, and CF(rule) represents the certainty factor of the rule. For multiple pieces of evidence for the same hypothesis, the combined certainty factor is calculated as:

[0253] CF(H,E1∧E2)=min(CF(H,E1),CF(H,E2))

[0254] CF(H,E1∨E2)=max(CF(H,E1),CF(H,E2))

[0255] Case-based reasoning uses a four-step model of "retrieval-reuse-correction-retention". Case retrieval is based on the k-nearest neighbor (k-NN) algorithm, which calculates the similarity between the current situation and the cases in the case library. The similarity calculation uses the weighted Euclidean distance:

[0256]

[0257] Among them, w i represents the weight of the i-th feature, and represents the i-th feature value of the two cases. Case correction adjusts the solution of the case based on the difference between the current situation and the retrieved case. Case retention adds the successful solution to the case library for future case retrieval.

[0258] Model-based reasoning uses physiological pharmacology models to predict intervention effects. The pharmacodynamic model uses the Emax model:

[0259]

[0260] Among them, E represents the effect, E max represents the maximum effect, C represents the drug concentration, and EC 50 represents the half-maximal effect concentration, and n represents the Hill coefficient. The pharmacokinetic model uses a multi-compartment model to describe the distribution and elimination process of the drug in the body.

[0261] The clinical decision aid uses fuzzy logic to process uncertain knowledge. The membership function of the fuzzy set uses a generalized bell-shaped function:

[0262]

[0263] Among them, μ A (x) represents the degree of membership of element x to set A. a, b, and c are parameters that control the width, slope, and center position of the function, respectively. Fuzzy reasoning uses the Mamdani method, which includes four steps: fuzzification, rule evaluation, aggregation, and defuzzification. Defuzzification uses the centroid method:

[0264]

[0265] Among them, z * represents the defuzzification result, μ A (z) represents the membership function of the aggregated fuzzy set.

[0266] The clinical decision aid supports multi-objective optimization, balancing anesthesia depth, blood flow protection, and surgical requirements. The objective function is defined as:

[0267] U=w1·U anesthesia +w2·U bloodflow +w3·U operation

[0268] Among them, w1, w2 and w3 represent the weight coefficients of each target, satisfying w1+w2+w3=1, U anesthesia 、U bloodflow and U operation The weight coefficients are dynamically adjusted based on the risk level and surgical stage. When the risk warning indicates abnormal liver blood flow, the weight of blood flow protection is increased; when the surgery is in the critical operation stage, the weight of surgical demand is increased.

[0269] The clinical decision aid also uses a reinforcement learning algorithm to continuously optimize the decision strategy. Reinforcement learning is based on the Q-learning algorithm, and the Q value update formula is:

[0270]

[0271] Where Q(s,a) represents the expected cumulative reward for taking action a in state s, α represents the learning rate, γ represents the discount factor, r represents the immediate reward, s′ represents the next state, and a′ represents the next possible action. The state represents the system's risk level and parameter status, the action represents the anesthesia adjustment measure, and the reward is defined as the weighted sum of the degree of risk reduction and the degree of improvement in physiological parameters.

[0272] The intervention recommendations generated by the clinical decision aid include three categories: anesthesia depth adjustment recommendations, anesthetic drug adjustment recommendations, and blood flow protection strategies. Anesthesia depth adjustment recommendations determine the adjustment direction and amplitude based on the deviation between the current BIS value and the target range. The target BIS range is set according to the surgical stage: 40-50 during the induction phase, 40-60 during the maintenance phase, and >70 during the awakening phase. The adjustment amplitude is divided into five levels (±5%, ±10%, ±15%, ±20%, ±25%), and the calculation formula is:

[0273] Adjustment=sign(BIS target -BIS current )·Level·Base

[0274] Where sign represents the sign function, Level∈{0.05,0.10,0.15,0.20,0.25}, and Base represents the base adjustment amount.

[0275] Recommendations for adjusting anesthetic drugs include dose adjustment, administration rate adjustment, and drug substitution. Dose adjustment takes into account the patient's weight, age, and liver function status. The formula is:

[0276] Dose=Standard_Dose·Weight_Factor·Age_Factor·Liver_Factor

[0277] Among them, Weight_Factor, Age_Factor and Liver_Factor represent the factors affecting the dosage by weight, age and liver function respectively. The formula for adjusting the dosage rate is:

[0278] Rate new =Rate current (1+Adjustment)

[0279] Drug substitution plans are based on the hepatic metabolic characteristics of different anesthetic drugs, taking into account the patient's liver function and blood flow status.

[0280] Blood flow protection strategies include recommendations for fluid management, vasoactive drug use, and surgical procedure adjustment. Fluid management recommendations are based on the patient's hemodynamic parameters and surgical blood loss, and a reasonable infusion rate and type are calculated. Recommendations for the use of vasoactive drugs are based on mean arterial pressure and cardiac output, and pressor or vasodilator drugs are recommended. Surgical procedure adjustment recommendations are based on blood flow status, and it is recommended to adjust the surgical procedure method or timing. For example, when liver blood flow decreases significantly, it is recommended to temporarily remove the Pringle clamp or reduce the clamping time.

[0281] The decision recommendations of the clinical decision aid are generated in structured text using the SOAP format:

[0282] The S (subjective data) section describes the risk status, including risk level, risk trend and affected parameters; the O (objective data) section describes the objective parameters and change trends, including the current values ​​of each parameter and their change rates; the A (assessment) section includes risk assessment and possible causes, based on the knowledge base and reasoning results analysis; the P (plan) section includes specific intervention recommendations, including adjustment direction, magnitude and expected effects.

[0283] Each recommendation is accompanied by a corresponding level of evidence and strength of recommendation, along with supporting evidence. Decision recommendations are presented to clinicians via a visual interface designed based on the principles of human-computer interaction for medical decision support, ensuring clear and easy-to-read information and highlighting important recommendations.

[0284] like Figure 2 As shown, the present invention provides a liver blood flow and anesthesia depth combined monitoring system during liver surgery, the workflow of which includes the following steps:

[0285] S1. Multi-source data acquisition. The multi-source data collector receives and collects patient-related signals related to liver blood flow (hepatic artery blood flow, portal vein blood flow, total hepatic blood flow, hepatic artery resistance index, portal vein resistance index), signals related to anesthesia depth (bispectral index, entropy index, anesthetic drug concentration, and raw EEG data), and systemic hemodynamic parameters (central venous pressure, mean arterial pressure, and cardiac output). Data are digitized at a sampling rate of 1000 Hz, ensuring a time synchronization accuracy of ±0.1 ms. The collector performs preliminary data processing, including removing significant artifacts, range checking, and unit conversion. Artifact detection uses a sliding window method with a window length of 10 sampling points. The processed data is packaged according to a unified data format, with a fixed header identifier of 0xAA55AA55. The data is then transmitted to the signal preprocessor via a reliable TCP / IP-based transport protocol.

[0286] S2. Signal preprocessing. The signal preprocessor performs filtering, denoising, and data verification on the raw signals from the multi-source data collector. A fourth-order Butterworth bandpass filter with a passband of 0.5–100 Hz and a roll-off of 24 dB / octave is applied to the EEG signal; a fourth-order Butterworth bandpass filter with a passband of 0.01–10 Hz is applied to the liver blood flow signal; and a fourth-order Butterworth low-pass filter with a cutoff frequency of 5 Hz is applied to the hemodynamic signal. Denoising is performed using wavelet transform technology, using the db4 wavelet basis function with a five-level decomposition layer and soft thresholding applied to the wavelet coefficients. The signal quality index (SQI) is calculated for the processed signal based on signal-to-noise ratio, stability, and artifact count. Signal segments with an SQI below 60 are marked or repaired. Short-duration (<1 second) low-quality data are repaired using cubic spline interpolation, while long-duration (>1 second) low-quality data are repaired using autoregressive model prediction. Time domain features, frequency domain features and time-frequency features are extracted from the preprocessed signal. The feature extraction adopts the sliding window method with a window length of 10 seconds and an overlap rate of 50%.

[0287] S3. Physiological coupling model calculation. The physiological coupling model calculator constructs a correlation model between hepatic hemodynamics and anesthesia depth based on the signal features provided by the preprocessor. The monitored parameters are organized into three state vectors: the hepatic blood flow state vector F, the anesthesia state vector A, and the systemic hemodynamic state vector H. A system of differential equations describes the dynamic relationship between the three state vectors. The state transition function is implemented using a recurrent neural network. The network structure consists of two layers of LSTM units, with 128 and 64 neurons in each layer, respectively. The differential equations are numerically solved using the fourth-order Runge-Kutta method, with a solution step size of 0.1 seconds. The hepatic blood flow-anesthesia depth coupling index (HAACI) is calculated based on the nonlinear mapping between hepatic blood flow parameters and anesthesia depth parameters. The weight coefficients are dynamically adjusted according to the type of anesthetic drug and individual patient characteristics. For propofol anesthesia, the initial weights are set to w1 = 0.4, w2 = 0.4, and w3 = 0.2; for sevoflurane anesthesia, the initial weights are set to w1 = 0.5, w2 = 0.3, and w3 = 0.2. The particle filter algorithm is used to estimate and update the model parameters in real time, and the number of particles is set to 1000. The calculator updates the calculation results once per second and transmits the HAACI index to the risk prediction and early warning device.

[0288] S4. Surgical operation recognition. The surgical operation recognizer identifies the current surgical operation type based on the signal features transmitted by the preprocessor. A two-stream convolutional neural network and long short-term memory network (CNN-LSTM) hybrid architecture is used to identify 13 typical hepatobiliary surgical operations, including liver mobilization, portal vessel exposure, Pringle clamp initiation, and Pringle clamp removal. The convolutional network consists of three convolutional layers: the first layer consists of 64 3×7 convolutional kernels, the second layer consists of 128 3×5 convolutional kernels, and the third layer consists of 256 3×3 convolutional kernels. The sequence analysis layer is configured as a two-layer BiLSTM, with 64 LSTM units in each direction, interposed with a dropout layer with a dropout rate of 0.5. The operation classification layer uses a fully connected network architecture, with 256 neurons in the first fully connected layer and 14 neurons in the second fully connected layer. Real-time recognition is achieved using a sliding window method with a window length of 10 seconds and a window sliding step of 0.5 seconds. When the probability of an operation exceeds the threshold θrec = 0.85, it is initially identified as that operation. The operation state must last for at least 2 seconds before it is confirmed. The recognizer outputs the recognition result every 2 seconds, and the recognition result is transmitted to the risk prediction and early warning system and clinical decision support system as a surgical environment parameter.

[0289] S5. Brain-liver information flow analysis. The brain-liver information flow analyzer synchronously analyzes EEG and liver blood flow data and calculates the information transfer relationship between them. EEG data are subjected to time-frequency analysis using wavelet packet transform (WPT). Using the db4 wavelet basis function, a five-level decomposition yields 32 subbands, which are then reorganized into five physiologically significant frequency bands: δ (0.5-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (30-100 Hz). Liver blood flow data are adaptively decomposed using empirical mode decomposition (EMD), typically selecting the first three IMFs as eigenmodes. Continuous signals are converted to symbolic sequences using the sorting pattern method with an embedding dimension of m = 3 and a time delay of τ = 1. Based on the symbolic sequences, the transfer entropy in both the brain-to-liver and liver-to-brain directions is calculated. Conditional entropy is calculated using the k-nearest neighbor (k-NN) method with k = 4. The final BLITI index was calculated as the weighted sum of the weight coefficient, normalized transfer entropy, and synergy measure. For propofol anesthesia, α = 0.6, β = 0.3, and γ = 0.1; for sevoflurane anesthesia, α = 0.7, β = 0.2, and γ = 0.1. The analysis process used a sliding window method with a 60-second window length and a 50% overlap. The BLITI index was updated every 10 seconds and the results were transmitted to the risk prediction and early warning system and clinical decision support system.

[0290] S6. Risk prediction and early warning. The risk prediction and early warning device integrates the outputs of the physiological coupling model calculator, surgical operation identifier, and brain-liver information flow analyzer to predict potential risks. A dynamic Bayesian network (DBN) was used to construct the risk assessment model. This DBN model consists of a three-layer network structure with a total of 34 nodes. A particle filter algorithm was used for state prediction, with N = 1000 particles, to achieve multi-step prediction. The prediction time windows included three scales: 30 seconds, 45 seconds, and 60 seconds, with a prediction granularity of 5 seconds per step. Risk states were assessed based on an individualized safety domain, represented by a multidimensional Gaussian distribution. The individualization coefficient α in the mean vector calculation formula was 0.7. The Mahalanobis distance was used to quantify the deviation between the current state and the safety domain. Distances across different prediction time windows were integrated, with weighting coefficients β1 = 0.5, β2 = 0.3, and β3 = 0.2. Risk is categorized into four levels based on Mahalanobis distance: Green (Safe): DM∈[0,2.0); Yellow (Caution): DM∈[2.0,4.0); Orange (Warning): DM∈[4.0,6.0); and Red (Danger): DM≥6.0. A trend alert function is implemented, triggering a trend alert when the risk increase rate exceeds a threshold of θrate=0.4 / s. The alerter updates the risk assessment every 5 seconds. When the risk level reaches Warning or Danger, an alert signal is triggered and transmitted to the clinical decision support system.

[0291] S7. Clinical Decision Support. The clinical decision support system generates clinical intervention recommendations based on the output of the risk prediction and early warning system, combined with individual patient characteristics and surgical stage information. Decision support functions are implemented based on a knowledge base system and a reasoning engine. The knowledge base consists of a hepatobiliary surgery knowledge base, an anesthesia management rule base, and a case library. A hybrid reasoning architecture is employed, encompassing rule-based, case-based, and model-based reasoning. Rule-based reasoning employs a forward-chaining reasoning strategy, with rule matching implemented using the Rete algorithm. Case-based reasoning utilizes a four-step "retrieve-reuse-modify-retain" model, with case retrieval based on the k-nearest neighbor (k-NN) algorithm. Model-based reasoning uses a physiological-pharmacological model to predict intervention effects, and an Emax model for pharmacodynamics. Fuzzy logic is used to handle uncertain knowledge, with the Mamdani method employed for fuzzy reasoning. Multi-objective optimization is supported, with the objective function defined as the weighted sum of the utility functions of each objective. Reinforcement learning, based on the Q-learning algorithm, is employed to continuously optimize the decision strategy. Generated intervention recommendations include three categories: anesthesia depth adjustment recommendations, anesthetic drug adjustment recommendations, and blood flow protection strategies. These recommendations are generated using the SOAP format to generate structured text. General recommendations are updated every minute, while emergency recommendations (risk level ≥ warning) are pushed immediately.

[0292] In summary, the present invention provides a joint monitoring system for liver blood flow and anesthesia depth during liver surgery, which realizes the joint monitoring and management of liver blood flow and anesthesia depth through seven key modules (multi-source data collector, signal preprocessor, physiological coupling model calculator, surgical operation identifier, brain-liver information flow analyzer, risk prediction and early warning device, and clinical decision assistant). The system overcomes the problems of mutual independence, lack of real-time performance and correlation analysis in the existing technology, and improves the accuracy of risk prediction and the scientific nature of clinical decision-making. The main advantages include improving surgical safety, optimizing anesthetic drug management, reducing false alarm interference, supporting individualized monitoring and optimizing clinical workflow, which is of great significance to improving the quality of hepatobiliary surgery and patient prognosis.

[0293] For those skilled in the art, modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all be included in the scope of protection of the present invention.

Claims

1. A combined monitoring system for liver blood flow and anesthesia depth during liver surgery, characterized in that: include: Multi-source data collector, used to collect patient liver blood flow-related signals, anesthesia depth-related signals, and systemic hemodynamic parameters; A signal preprocessor, configured to perform filtering, denoising and feature extraction on the original signal transmitted from the multi-source data collector; Physiological coupling model calculator, used to construct a correlation model between liver hemodynamics and anesthesia depth based on preprocessed signal characteristics and calculate the liver blood flow-anesthesia depth coupling index; a surgical operation identifier for identifying the type of surgical operation based on preprocessed signal features; Brain-liver information flow analyzer, used to synchronously analyze EEG and liver blood flow data and calculate the information transmission relationship between the two; A risk prediction and early warning device is used to integrate the output results of the physiological coupling model calculator, surgical operation identifier, and brain-liver information flow analyzer to predict potential risks; A clinical decision aid is used to generate clinical intervention recommendations based on the output of the risk prediction and early warning device and individual patient information.

2. The system according to claim 1, wherein: The multi-source data collector digitizes data at a sampling rate of 1000 Hz, ensuring a time synchronization accuracy of ±0.1 ms; the liver blood flow-related signals include hepatic artery blood flow, portal vein blood flow, total liver blood flow, hepatic artery resistance index, and portal vein resistance index; the anesthesia depth-related signals include bispectral index, entropy index, anesthetic drug concentration, and EEG raw data; the systemic hemodynamic parameters include central venous pressure, mean arterial pressure, and cardiac output.

3. The system according to claim 1, wherein: The signal preprocessor applies a 0.5-100 Hz band-pass filtering algorithm to the EEG signal and a 0.01-10 Hz band-pass filtering algorithm to the liver blood flow signal, and uses a wavelet transform algorithm to eliminate environmental and physiological artifacts; the signal preprocessor calculates a signal quality index and marks or repairs signal segments with a signal quality index lower than 60; the signal preprocessor extracts time domain features, frequency domain features, and time-frequency features from the preprocessed signal.

4. The system according to claim 1, wherein: The physiological coupling model calculator organizes the monitoring parameters into a liver blood flow state vector, an anesthesia state vector, and a systemic hemodynamic state vector, and describes the dynamic correlation between the three vectors through a set of differential equations; the physiological coupling model calculator uses the fourth-order Runge-Kutta method to numerically solve the set of differential equations, and calculates the liver blood flow-anesthesia depth coupling index through a weighted nonlinear function, and the weight coefficient is dynamically adjusted according to the individual characteristics of the patient, the surgical stage, and the type of anesthetic drug.

5. The system according to claim 1, wherein: The surgical operation identifier uses a convolutional neural network to extract the spatial features of the signal, combines it with a long short-term memory network to analyze the temporal change pattern, and automatically identifies hepatobiliary surgical operations; the surgical operation identifier uses a 10-second signal analysis window and outputs a recognition result every 2 seconds. The recognition result is transmitted to the risk prediction and warning device and the clinical decision assistant as a surgical environment parameter.

6. The system according to claim 1, wherein: The brain-liver information flow analyzer decomposes the EEG signal into five frequency bands: δ (0.5-4Hz), θ (4-8Hz), α (8-13Hz), β (13-30Hz) and γ (30-100Hz), and extracts the energy characteristics of each frequency band; based on the transfer entropy theory, the brain-liver information flow analyzer calculates the information flow intensity in two directions: from EEG activity to liver blood flow and from liver blood flow to EEG activity, and synthesizes the brain-liver information transfer index; the brain-liver information flow analyzer updates the calculation results every 10 seconds.

7. The system according to claim 1, wherein: The risk prediction and early warning device predicts system state changes within the next 30-60 seconds based on a dynamic Bayesian network model; the risk prediction and early warning device calculates the Mahalanobis distance between the predicted state and the safety domain to quantify the degree of deviation; the risk prediction and early warning device divides risks into four levels: green (safe), yellow (caution), orange (warning) and red (danger) according to the Mahalanobis distance value; the risk prediction and early warning device updates the risk assessment results every 5 seconds.

8. The system according to claim 1, wherein: The clinical decision assistant has a built-in hepatobiliary surgery knowledge base and an anesthesia management rule base; the clinical decision assistant adopts a hybrid reasoning architecture, including rule-based reasoning, case-based reasoning, and model-based reasoning; the clinical decision assistant generates anesthesia depth adjustment recommendations, anesthetic drug adjustment recommendations, and blood flow protection measures based on risk type and level.

9. The system according to claim 1, wherein: The clinical decision assistant uses a reinforcement learning algorithm to optimize the decision-making strategy; the intervention recommendations generated by the clinical decision assistant use SOAP format structured text, including a description of the intervention direction, adjustment range, and expected effect; the clinical decision assistant updates general recommendations once a minute and pushes emergency recommendations immediately.

Images