Intraoperative bleeding amount prediction enhancing system for transurethral prostatic enucleation

Abstract

The invention relates to the technical field of prostatic enucleation, and discloses a transurethral prostatic enucleation intraoperative bleeding amount prediction enhancing system, which specifically comprises an end-side multi-mode intelligent sensing module, a transurethral prostatic enucleation module and a transurethral prostatic enucleation module, the edge side causal perception and collaborative decision engine is deployed in an embedded computer provided with an image processing unit in an operating room and receives side signals; a cloud federated causal model library and a digital twin module; according to the in-situ, synchronous and real-time monitoring of a microstructure, a mechanical state and a cell metabolism level in an operation is realized through the acoustic sensor, the optical sensor, the mechanical sensor and the electrical impedance sensor which are integrated on the end side; especially, by calculating a metabolic stress index (MSI), the system can capture extremely early physiological premonition such as cell energy metabolism disorder before a blood vessel is physically broken for tens of seconds, a prewarning window is greatly advanced, and a key time advantage is created for active intervention.

Description

Technical Field

[0001] This invention relates to the field of prostate enucleation technology, specifically to an enhanced system for predicting intraoperative bleeding during transurethral prostate enucleation surgery. Background Technology

[0002] Transurethral enucleation of the prostate (TURP) is a mainstream minimally invasive surgical procedure for treating benign prostatic hyperplasia (BPH). Its core challenge lies in the effective management of bleeding risk during the procedure. Current clinical practice heavily relies on the surgeon's visual experience and tactile intuition, lacking objective and quantitative technical support for predicting potential bleeding points. Specifically, current intraoperative monitoring primarily depends on endoscopic imaging and vital signs, which are "outcome" indicators after bleeding has occurred and have significant lag, leading to passive hemostasis and potentially missing the optimal intervention window. Although preoperative imaging can assess glandular morphology and blood supply, it cannot reflect the dynamic tissue mechanics and microcirculatory changes caused by the surgical procedure in real time. In recent years, researchers have attempted to introduce single-modality intraoperative monitoring techniques, such as estimating blood loss based on changes in hemoglobin in the irrigation fluid, or using Doppler ultrasound to monitor local blood flow. While imaging can be performed on bleeding, the former is still post-operative measurement and is susceptible to interference, while the latter is difficult to popularize due to the difficulty in probe integration and the complexity of operation. In terms of intelligent assistance, existing systems are mostly based on traditional machine learning models, learning the statistical correlation between bleeding and certain intraoperative parameters from historical data to provide early warnings. These methods are essentially "black box" correlation predictions, lacking interpretable modeling of the causal mechanism of "why and how tissue tends to bleed," resulting in a high false alarm rate and failing to provide surgeons with causal decision support on "why to issue a warning" and "how to intervene in a targeted manner." In addition, the "islands" formed by the closed data of various medical institutions limit the model's ability to use multi-center data for iterative optimization and generalization improvement. Furthermore, existing human-computer interaction is mostly based on one-way alarms, failing to deeply understand the surgeon's intentions and achieve dynamic and flexible collaborative control. Summary of the Invention

[0003] The purpose of this invention is to provide an enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) to address the problems mentioned in the background section.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate, specifically comprising: End-side multimodal intelligent sensing module: integrated into the end of the surgical instrument, it includes a miniature broadband acoustic sensor assembly, an integrated sweep frequency source optical coherence tomography probe, a three-dimensional torque sensing unit, a multi-band bioelectrical impedance measurement circuit, and a multimodal data synchronization and preprocessing unit, for in-situ, real-time acquisition of physical and physiological signals of tissue interaction; The edge-side causal perception and collaborative decision-making engine is deployed in an embedded computer equipped with an image processing unit in the operating room. It includes a high-speed data stream access and buffer queue, a causal structure learning and update module, a potential outcome model and causal effect estimation unit, a counterfactual reasoning and strategy generation unit, and a surgical instrument kinematic analysis and intent recognition unit. It receives edge-side signals, performs causal discovery, intent recognition, risk calculation, and generates collaborative decision-making suggestions and control instructions. Cloud-based federated causal model library and digital twin module: Deployed on the central server, it maintains and updates the global causal knowledge graph and high-fidelity tissue biophysical model, and provides incremental model download and anonymized data upload services for the edge side; Human-computer two-way interaction and dynamic control module: including augmented reality head-mounted display, force feedback surgical instrument master hand, and voice interaction unit, to realize two-way information flow and dynamic and flexible management of control permissions.

[0005] Preferably, the end-side multimodal intelligent sensing module specifically comprises: Miniature broadband acoustic sensor assembly: It consists of four piezoelectric composite ultrasonic transducers arranged in a cross shape on the end sidewall of the instrument; each transducer operates in the frequency band of 10kHz-1.5MHz and is equipped with a preamplifier and a bandpass filter; the array realizes sound source localization by measuring the time difference of arrival of sound waves, and the localization algorithm is based on the generalized cross-correlation-phase transformation method; Integrated sweep frequency source optical coherence tomography probe: This probe is coaxially integrated with the instrument working channel and uses a sweep frequency laser source with a center wavelength of 1300nm. B-scan is achieved through fiber optic coupler and miniaturized two-dimensional scanning galvanometer. After the interference signal is received by the balanced detector, the depth reflectivity profile is reconstructed through Fourier transform.

[0006] Preferably, the three-dimensional torque sensing unit has a strain gauge Wheatstone bridge array positioned near the end of the instrument. Multi-band bioelectrical impedance measurement circuit: A four-electrode method is used, with two pairs of metal electrodes placed at the end of the instrument; the excitation electrode applies a multi-frequency sinusoidal voltage signal V with a frequency range of 1kHz-1MHz and an amplitude less than 100mVpp. in (f); Measurement electrode collects response current I out (f), the real part Z′(f) and imaginary part Z′′(f) of the impedance are extracted by a lock-in amplifier; then, the measurement spectrum is fitted to the Cole-Cole model by nonlinear least squares method.

[0007] Preferably, the edge-side causal perception and collaborative decision-making engine specifically comprises: High-speed data stream access and buffer queue: Several first-in-first-out buffer queues are created to receive different modal data streams from the end side; each queue implements a circular buffer structure, and when the amount of data in the buffer reaches the preset window length W, the subsequent processing thread is triggered to ensure real-time performance; Causal Structure Learning and Update Module: This module uses a sliding window approach to process multimodal time series data {X}. t}, X t ∈R d To conduct the analysis; firstly, using PC The algorithm performs conditional independence tests to learn the causal framework; for any pair of variables (X... i X j Given the condition set S, the partial correlation coefficient ρ is calculated. ij |S and test whether it is zero to determine conditional independence; Gaussian process regression is used to perform nonlinear conditional independence test; Potential Outcome Model and Causal Effect Estimation Unit: For the learned causal graph, a dual machine learning approach is used to estimate the conditional average treatment effect for specific intervention variables T and outcome variables Y. First, two nuisance functions are fitted using random forest or deep neural networks: the outcome prediction model and the propensity score model, respectively. Then, cross-fitting is used to obtain the unbiased estimator.

[0008] Preferably, the counterfactual reasoning and strategy generation unit calculates counterfactual results based on a structural causal model for the current observation X=x and the proposed intervention; Surgical Instrument Kinematics Analysis and Intent Recognition Unit: This unit receives the instrument end-effector pose q from the robot controller. t ∈SE(3) and the interaction force F from the force sensing module t First, calculate the motion characteristics, including the terminal linear velocity v. t angular velocity ω t The features, including acceleration and force-motion covariance, together with the tissue type labels segmented from real-time OCT images, constitute the observation vector Ot. A long short-term memory network-conditional random field hybrid model is used to decode the observation sequence and output the basic surgical action label corresponding to each time step. The high-level intent recognizer is based on a hidden semi-Markov model, which maps the basic action sequence to a finite intent state It∈{exploration, dissection, hemostasis}, and its state transition probability and duration distribution are learned through historical surgical data.

[0009] Preferably, the cloud-based federated causal model library and digital twin module specifically include: Differential Privacy Federated Aggregator Server: This server coordinates K clients for federated learning; in each training round, the server allocates the global model parameters θ. gBroadcast to clients; each client k has local dataset D k Calculate the gradient of the loss function The gradient is then clipped to ensure its L2 norm does not exceed a threshold. Subsequently, the client adds Gaussian noise to the gradient: and will Encrypted upload; the server uses a secure multi-party computation protocol to decrypt and aggregate gradients.

[0010] Preferably, the human-computer two-way interaction and dynamic control module specifically includes: Dynamic permission decision-maker: This decision-maker calculates the situation-aware safety margin ξ(t) in real time; its inputs include: causal risk estimate R(t), metabolic stress index MSI(t), and counterfactual prediction variance σ. 2 CF(t), Physician-operated tremor power spectral density P tremor (f) Integration in the 8–12 Hz frequency band, and the directional consistency ρ between system commands and doctor's operations. align (t); these inputs are mapped to scalars ξ(t) through a normalization function and a learnable multilayer perceptron; the weights of the multilayer perceptron are optimized through reinforcement learning to maximize the long-term safety-efficiency trade-off gains.

[0011] Preferably, it also includes a multimodal data synchronization and preprocessing unit: a built-in synchronization controller based on a field-programmable gate array (FPGA) provides a unified hardware trigger clock for all sensors, with a synchronization error of less than 1 microsecond; the raw data first undergoes baseline correction and power frequency notch filtering, then is segmented according to fixed time windows, and a timestamp based on a GPS-disciplined atomic clock is added; the preprocessed data stream is sent to the edge side via a gigabit Ethernet interface in the form of User Datagram Protocol (UDP) packets, specifically: Multi-source signal synchronization and resampling; R from the impedance measurement circuit i (t) sequence, StO2(t) sequence from OCT common-path spectrometer, and envelope energy P from a specific frequency band of acoustic sensor. mito The (t) sequence is resampled to a unified time base by linear interpolation at a sampling frequency of 100Hz.

[0012] Preferably, the causal structure learning and updating module employs the following hybrid approach to process causal discovery in multimodal time series data: Variable definition and time delay processing: Each physical quantity in the multimodal data stream is defined as a time series variable; at the same time, possible time delay versions of each variable are defined to construct an extended variable set; Skeleton learning based on conditional independence test; PC operation on extended variable set. Algorithm; For each pair of variables, condition variables are progressively added to the possible condition set S, and partial correlation and kernel-based independence tests are used; the test statistic is compared with the critical value; this stage outputs an undirected graph skeleton; Direction determination and V-structure identification; identification of all potential triples XZY that satisfy the V-structure on the skeleton.

[0013] Preferably, the security aggregation protocol executed by the differential privacy federated aggregation server is as follows: Initialization and key distribution: The server generates a pair of homomorphically encrypted public key pk and private key sk for each round of training, and broadcasts pk to all participating clients; at the same time, a secret sharing scheme is adopted to distribute a master key MK to all clients, so that more than t clients are needed to jointly recover the key.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. Through end-side integrated acoustic, optical, mechanical, and electrical impedance sensors, in-situ, synchronous, and real-time monitoring of tissue microstructure, mechanical state, and cellular metabolic level is achieved during surgery. In particular, by calculating the "metabolic stress index (MSI)," the system can capture very early physiological precursors such as cellular energy metabolism disorders tens of seconds before physical rupture of blood vessels, significantly advancing the warning window and creating a key time advantage for proactive intervention, effectively avoiding the lag of traditional methods that rely on visible bleeding. 2. Decision-making basis shifts from "statistical association" to "causal explanation": The system utilizes the causal discovery and counterfactual reasoning engine on the edge side to automatically construct and update the "structural causal model" of surgical operations in real time from high-dimensional time-series data. This not only enables high-precision prediction of bleeding risk but also clearly reveals the causal path of risk generation, providing doctors with interpretable decision support on "why to issue a warning" and "how to intervene." This significantly reduces the high false alarm rate caused by spurious correlations in traditional machine learning models and enhances clinical credibility. 3. Model capabilities have evolved from "static isolation" to "dynamic collaboration": Relying on the cloud-based federated causal model library and digital twin module, the system, while strictly protecting the privacy of patient data from various medical institutions, achieves continuous fusion and sharing of multi-center knowledge through secure aggregation algorithms. This enables the system to utilize a wider range of diverse case data to continuously optimize and generalize its early warning and causal models, forming a continuously evolving collective intelligence. This effectively solves the problem of poor model generalization ability and difficulty in adapting to individual differences caused by medical "data silos." 4. Human-computer interaction has been upgraded from "one-way alarm" to "two-way collaboration": The system understands the doctor's surgical steps and goals through the intent recognition module, and seamlessly and flexibly switches between multiple modes such as "fully automatic", "shared control" and "enhanced guidance" based on real-time risk and doctor status through a dynamic permission manager. This collaborative mode, in which the doctor is the leader and the system is the intelligent assistant, provides intuitive guidance through multiple channels such as augmented reality, spatial audio and force feedback. It not only ensures the bottom line of safety, but also respects and enhances the doctor's operational sovereignty and sense of control, greatly reducing cognitive load and improving the smoothness and overall efficiency of the surgery. Attached Figure Description

[0015] Figure 1 This is a structural block diagram of an enhanced system for predicting intraoperative blood loss during transurethral resection of the prostate. Figure 2 This is a structural block diagram of an end-side multimodal intelligent sensing module for an enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate. Figure 3 This is a structural block diagram of a marginal-side causal perception and collaborative decision-making engine for an enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate. Figure 4 This is a structural block diagram of a cloud-based federated causal model library and digital twin module for an enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate. Figure 5 The diagram shows the structural block of a human-computer two-way interaction and dynamic control module for an enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This invention provides a technical solution: an enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP), specifically comprising: End-side multimodal intelligent sensing module: integrated into the end of the surgical instrument, it includes a miniature broadband acoustic sensor assembly, an integrated sweep frequency source optical coherence tomography probe, a three-dimensional torque sensing unit, a multi-band bioelectrical impedance measurement circuit, and a multimodal data synchronization and preprocessing unit, for in-situ, real-time acquisition of physical and physiological signals of tissue interaction; The edge-side causal perception and collaborative decision-making engine is deployed in an embedded computer equipped with an image processing unit in the operating room. It includes a high-speed data stream access and buffer queue, a causal structure learning and update module, a potential outcome model and causal effect estimation unit, a counterfactual reasoning and strategy generation unit, and a surgical instrument kinematic analysis and intent recognition unit. It receives edge-side signals, performs causal discovery, intent recognition, risk calculation, and generates collaborative decision-making suggestions and control instructions. Cloud-based federated causal model library and digital twin module: Deployed on the central server, it maintains and updates the global causal knowledge graph and high-fidelity tissue biophysical model, and provides incremental model download and anonymized data upload services for the edge side; Human-computer two-way interaction and dynamic control module: including augmented reality head-mounted display, force feedback surgical instrument master hand, and voice interaction unit, to realize two-way information flow and dynamic and flexible management of control permissions.

[0018] Furthermore, the edge-side multimodal intelligent sensing module specifically includes: Miniature broadband acoustic sensor assembly: Composed of four piezoelectric composite ultrasonic transducers arranged in a cross shape on the end wall of the instrument; each transducer operates in the 10kHz-1.5MHz frequency band and is equipped with a preamplifier and bandpass filter; the array achieves sound source localization by measuring the time difference of arrival of sound waves. The localization algorithm is based on the generalized cross-correlation-phase transform method, calculating sensor pairs ( , Delay estimation between received signals: ; ; in, For sensor pairs ( , Estimation of the time delay (time difference) between received signals. The generalized cross-correlation-phase transform function is a function of the time delay τ. For frequency, The Fourier transform of the signal received by the i-th sensor. The complex conjugate of the Fourier transform of the signal received by the j-th sensor; based on the time delay set { Given the known spatial coordinates of the sensor, the three-dimensional coordinates of the sound source are solved using the least squares method; Integrated sweep-frequency source optical coherence tomography probe: This probe is coaxially integrated with the instrument's working channel, employing a sweep-frequency laser source with a center wavelength of 1300nm. B-scan is achieved through a fiber optic coupler and a miniaturized two-dimensional scanning mirror. After the interference signal is received by a balanced detector, the depth reflectivity profile is reconstructed through Fourier transform. ; in, Let z be the reflectivity intensity at depth z. The signal is an interference signal in the wavenumber domain; the system generates a 500×500 pixel cross-sectional image in real time and spatially registers it with the endoscope white light image through feature point matching.

[0019] Furthermore, the three-dimensional torque sensing unit: a strain gauge Wheatstone bridge array is set at the proximal end of the instrument; the output voltage V∈R of each bridge is decoupled using a decoupling algorithm. n Converted to three-dimensional force F∈R 3 and three-dimensional torque M∈R 3 : [F T M T ] T =C·V, where C∈R 6×n The decoupling matrix is ​​obtained through calibration experiments; the sampling frequency is not less than 2kHz, and the measurement resolution is better than 0.01N and 0.1N·mm. Multi-band bioelectrical impedance measurement circuit: A four-electrode method is used, with two pairs of metal electrodes placed at the end of the instrument; the excitation electrode applies a multi-frequency sinusoidal voltage signal V with a frequency range of 1kHz-1MHz and an amplitude less than 100mVpp. in (f); Measurement electrode collects response current I out (f) The real part Z′(f) and imaginary part Z′′(f) of the impedance are extracted by a lock-in amplifier; then, the measurement spectrum is fitted to the Cole-Cole model using a nonlinear least squares method: ; The parameters of intracellular resistance Ri, extracellular resistance Re, and characteristic frequency fc = 1 / (2πτ) are extracted, where, Let f be the complex impedance at frequency f, in ohms (Ω). This is the high-frequency limiting resistance (resistance as ω→∞). It is the DC resistance (resistance when ω→0). The imaginary unit, Angular frequency, The relaxation time constant is expressed in seconds (s). Let fc be the distribution parameter and fc be the characteristic frequency.

[0020] Furthermore, the edge-side causal perception and collaborative decision-making engine specifically includes: High-speed data stream access and buffer queue: Several first-in-first-out buffer queues are created to receive different modal data streams from the end side; each queue implements a circular buffer structure. When the amount of data in the buffer reaches the preset window length W (e.g., corresponding to 1 second of data), the subsequent processing thread is triggered to ensure real-time performance. Causal Structure Learning and Update Module: This module uses a sliding window approach to process multimodal time series data {X}. t}, X t ∈R d To conduct the analysis; firstly, using PC The algorithm performs conditional independence tests to learn the causal framework; for any pair of variables (X... i X j Given the condition set S, the partial correlation coefficient ρ is calculated. ij |S and test whether it is zero to determine conditional independence; Gaussian process regression is used to perform nonlinear conditional independence test; Potential Outcome Model and Causal Effect Estimation Unit: For the learned causal graph, a dual machine learning approach is used to estimate the conditional average treatment effect for specific intervention variables T and outcome variables Y. First, two nuisance functions are fitted using random forest or deep neural networks: the outcome prediction model and the propensity score model, respectively. Then, cross-fitting is used to obtain the unbiased estimator.

[0021] Furthermore, the counterfactual reasoning and strategy generation unit, based on a structural causal model, calculates counterfactual outcomes for the current observation X=x and the proposed intervention. This process is completed through three steps: abduction, action, and prediction. Abduction: Inferring the approximate posterior distribution P(U|x) of the exogenous variable U based on x; Action: Replacing the equation for variable T in the structural equation with T=t′; Prediction: Calculating the distribution of Y based on P(U|x) under the modified model. The system calculates the counterfactual risks under multiple candidate interventions {t1′, t2′} in parallel. And select the intervention that minimizes the risk below the threshold and the energy parameter as the recommended strategy; Surgical Instrument Kinematics Analysis and Intent Recognition Unit: This unit receives the instrument end-effector pose q from the robot controller. t ∈SE(3) and the interaction force F from the force sensing module t First, calculate the motion characteristics, including the terminal linear velocity v. t angular velocity ω t The features, including acceleration and force-motion covariance, together with the tissue type labels (gland, capsule, blood vessel) segmented from the real-time OCT image, constitute the observation vector Ot. A long short-term memory network-conditional random field hybrid model is used to decode the observation sequence and output the basic surgical action label (such as "contact", "push", "cut", "coagulate") corresponding to each time step. The high-level intent recognizer is based on a hidden semi-Markov model, which maps the basic action sequence to a finite intent state It∈{exploration, dissection, hemostasis}, and its state transition probability and duration distribution are learned through historical surgical data.

[0022] Furthermore, the cloud-based federated causal model library and digital twin module specifically include: Differential Privacy Federated Aggregator Server: This server coordinates K clients (edge ​​nodes) for federated learning; in each training round, the server allocates the global model parameters θ. g Broadcast to clients; each client k has local dataset D k Calculate the gradient of the loss function The gradient is then clipped to ensure its L2 norm does not exceed a threshold. Subsequently, the client adds Gaussian noise to the gradient: and will Encrypted upload; the server uses a secure multi-party computation protocol to decrypt and aggregate gradients: ; Then update the global model: ; in, The average gradient after aggregation. The total number of clients participating in federated learning. These are global model parameters. For learning rate, This is an assignment operator, meaning that the variable on the left is updated with the value on the right. A cross-modal digital twin generative network: This network takes preoperative T2-weighted magnetic resonance imaging (MRI) and diffusion-weighted imaging (DWI) sequences of images as input. The encoder adopts a 3D U-Net architecture to extract multi-scale features. The decoder consists of two branches: one branch outputs a high-resolution 3D anatomical structure mesh model through voxel rendering; the other branch predicts the tissue material properties of each voxel through a physical information neural network, such as the anisotropic Young's modulus tensor E, Poisson's ratio ν, thermal conductivity k, and blood perfusion rate ωb. The network training jointly optimizes the anatomical reconstruction loss and the physical property prediction loss. Physical simulation and model update service: This service receives real-time intervention parameters (such as laser power and application location) and partial state feedback (such as local tissue displacement) from the edge sensing end; based on the digital twin model, it uses the finite element method to solve the coupled biothermodynamic equations. ; in, For divergence operators, Here, denoted as Cauchy stress tensor, is expressed in Pascals (Pa). Let be the displacement field vector, in meters (m). For the temperature field, Volume force vector ; in For gradient operators, For tissue density, To organize specific heat capacity, The rate of change of temperature over time. To improve the thermal conductivity of the tissue, Metabolic heat production rate, For the heat production rate of the external heat source, For blood perfusion rate, Blood density, For the specific heat capacity of blood, The simulation results (such as predicted stress concentration areas and high temperature areas) are converted into risk indicators and compared with the observed risks uploaded from the edge. Uncertain parameters (such as local blood flow changes) in the twin model are updated through Kalman filters and variational inference methods to achieve data assimilation between the model and the real world state.

[0023] Furthermore, the human-machine two-way interaction and dynamic control module specifically includes: Dynamic permission decision-maker: This decision-maker calculates the situation-aware safety margin ξ(t) in real time; its inputs include: causal risk estimate R(t), metabolic stress index MSI(t), and counterfactual prediction variance σ. 2 CF(t), Physician-operated tremor power spectral density P tremor (f) Integration in the 8–12 Hz frequency band, and the directional consistency ρ between system commands and doctor's operations. align (t); these inputs are mapped to scalars ξ(t) through a normalization function and a learnable multilayer perceptron; the weights of the multilayer perceptron are optimized through reinforcement learning to maximize the long-term safety-efficiency trade-off gains; Variable admittance controller: In shared control mode, this controller dynamically adjusts the admittance model parameters from the handheld robot based on ξ(t) and the current task stage; the admittance model equation is: ; in The virtual mass matrix (inertia matrix) is usually a diagonal matrix. For virtual damping matrix, For virtual stiffness matrix, The acceleration is the end position error. The rate of end position error For the end position error, The primary manual force applied to the doctor, Environmental forces; virtual mass Damping and stiffness The constructed diagonal matrix is ​​adjusted online according to ξ(t), where ξ(t) is the situation-aware safety margin, a dimensionless scalar in the range [0, 1]. When ξ(t) is low (high risk), it increases. and To enhance guidance and stability; when ξ(t) is high and the intention is matched, reduce To allow doctors more freedom of action; Multi-channel feedback rendering engine: Visual rendering: Perspective projection transformation is used to overlay virtual content such as 3D risk heatmaps, causal graph edges, and instrument prediction trajectories onto the endoscopic video stream; depth buffering testing is used to solve occlusion problems and ensure the correct front-to-back relationship between virtual information and real tissues; the strength of causal relationships is encoded by the width and opacity of Bézier curves. Haptic rendering: based on the direction F of the virtual force used for navigation. guide Based on the position of the instrument end, calculate the target displacement or vibration frequency of each actuator of the tactile device; use the God-Object and surrogate point algorithm to calculate the penetration depth of the virtual constraint boundary and generate a feedback force proportional to it; Auditory rendering: Spatial filtering of warning sounds is performed using the head-related transfer function library; based on the two-dimensional coordinates of the risk source on the screen, its azimuth angle relative to the center of the screen is calculated, and the corresponding left and right ear filter coefficients are interpolated from the HRTF database. The mono prompt sound is then convolved to generate stereo audio with spatial orientation. An independent security monitoring coprocessor, as a standalone hardware circuit based on field-programmable gate arrays and application-specific integrated circuits, specifically includes: Deterministic physics model computation unit: This unit continuously monitors the umain command from the main system energy control module; Hardware rule checker: This checker has a set of non-programmable safety thresholds embedded in it, such as maximum permissible instantaneous power Pmax, maximum permissible single-point energy Emax_spot, and minimum safe movement speed of the instrument end (vmin) (to prevent static burning); these thresholds are fixed during chip manufacturing; the rule checker compares the output of the deterministic model calculation unit (such as Edep(t)) with the fixed thresholds in parallel. Hardware interrupt generation logic: This logic receives the output signal of the rule checker; once any comparator outputs a trigger signal (e.g., Edep(t)>Emax_spot), the interrupt generation logic will send a low-level active hard-wired reset signal to the main power switch of the energy control unit within nanoseconds, and at the same time send a non-maskable interrupt to the main system CPU; this process does not go through any software layer, ensuring that a safe disconnection can be performed even if the main system software crashes.

[0024] Furthermore, it also includes a multimodal data synchronization and preprocessing unit: a built-in synchronization controller based on a field-programmable gate array (FPGA) provides a unified hardware trigger clock for all sensors, with a synchronization error of less than 1 microsecond; the raw data first undergoes baseline correction and power frequency notch filtering, then is segmented according to a fixed time window (typically 50ms), and a timestamp based on a GPS-disciplined atomic clock is added; the preprocessed data stream is sent to the edge side via a gigabit Ethernet interface in the form of User Datagram Protocol (UDP) packets, specifically: Multi-source signal synchronization and resampling; R from the impedance measurement circuit i The StO2(t) sequence, the StO2(t) sequence from the OCT common-path spectrometer, and the envelope energy P from a specific frequency band (80-120kHz) of the acoustic sensor. mito The (t) sequence is resampled to a unified time base by linear interpolation at a sampling frequency of 100Hz; Feature extraction and sliding window calculation; for each signal, calculate the dynamic features within a sliding window of length Tw = 2s: right For the sequence, the rate of change is approximated using the first-order difference: ; in, Intracellular resistance In time rate of change, For time The intracellular resistance value, Δt = 10 ms; then calculate the intracellular resistance value within the window. The mean; For the StO2(t) sequence, calculate the average value within the current window; For P mito (t) sequence, calculate its relationship with the total acoustic energy P total (t) (1kHz-1MHz) average ratio within the window: ; in, In the time window The average percentage of mitochondrial-related acoustic energy within the mitochondria. In the time window Take the average of the inner values. The acoustic energy in the mitochondrial resonance frequency band (80-120kHz) at time t. The total acoustic energy at time t is calculated using exponential fusion; the resulting three eigenvalues ​​are... StO2 Normalization is performed to make the mean 0 and the variance 1; then, the metabolic stress index (MSI) at the center of the current window is obtained by weighted summation using the following formula: ; in, The metabolic stress index is a dimensionless scalar. , , These are the fixed weight coefficients obtained by training a linear regression model using historical data. It is a linear rectified function. After normalization Window mean, This represents the normalized mean of tissue oxygen saturation within the specified window. This represents the normalized mean of the mitochondrial energy percentage window. The sliding time window length ensures that the ReLU function only contributes positively to the exponent when intracellular resistance increases (which may be related to cell edema); Trend analysis and early warning triggering; continuously calculate and cache the most recent 10 MSI values ​​(i.e., data from the past 20 seconds), use linear regression to fit their trend over time, and obtain the slope kMSI; generate a primary metabolic early warning signal when the following two conditions are met simultaneously: MSI>Θ level (Threshold, for example, 1.5); kMSI>Θ slope (Slope threshold, e.g., 0.05 / second).

[0025] Furthermore, the causal structure learning and update module employs the following hybrid approach to handle causal discovery in multimodal time series data: Variable definition and time delay processing: Each physical quantity in the multimodal data stream (such as laser power P, tissue surface temperature Ts, acoustic high-frequency energy Eac, metabolic stress index MSI, and hemorrhage marker B) is defined as a time series variable; at the same time, possible time delay versions of each variable are defined to construct an extended variable set. Skeleton learning based on conditional independence test; PC operation on extended variable set. Algorithm; For each pair of variables, condition variables are progressively added to the possible condition set S, and partial correlation and kernel-based independence tests are used; the test statistic is compared with the critical value; this stage outputs an undirected graph skeleton; Orientation determination and V-structure identification; Identify all potential triples XZY that satisfy the V-structure on the skeleton; If X and Y are not adjacent and Z is not included in any set S that makes X and Y conditionally independent, then orient the structure as X→Z←Y; Apply the direction propagation rule to determine the orientation of the remaining edges; Fine-grained optimization based on scores; local optimization using score-based search; BIC scoring function used. ; in Cause-and-effect diagram Bayesian information criterion score, To estimate parameters in maximum likelihood Below, data Log-likelihood, For the observation dataset, For the image The maximum likelihood estimate of the corresponding model parameters, The number of model parameters, For the sample size, Given the natural logarithm; use a greedy search (e.g., adding, deleting, reversing edges) to find a graph with a higher BIC score until convergence; Online incremental updates: When a new data window arrives, the module does not learn from scratch, but adopts an incremental update strategy. First, it calculates the change in the independence test p-value of each edge under the new window data. For edges whose p-values ​​change significantly (e.g., from <0.01 to >0.05, or vice versa), it re-runs the skeleton learning, direction determination, V-structure recognition, and score-based fine optimization algorithms based on conditional independence tests on the local subgraph to update the local causal structure, thereby achieving dynamic evolution of the causal graph with lower computational cost.

[0026] Furthermore, the security aggregation protocol executed by the differential privacy federated aggregation server is as follows: Initialization and key distribution: The server generates a pair of homomorphically encrypted public key pk and private key sk for each round of training, and broadcasts pk to all participating clients; at the same time, a secret sharing scheme is adopted to distribute a master key MK to all clients, so that more than t clients are needed to jointly recover the key; Client-side local computation and masking; client-side gradient gk is calculated locally and clipped to obtain the gradient. Subsequently, the client generates two random masks: one is a locally generated random vector. The other is a public mask generated using the master key MK and the client identity k. (For example, generated via a pseudo-random function); the vector encrypted and uploaded by the client is: ; Secure aggregation; the server collects the encryption vectors from all clients. Then, the sum under the ciphertext is calculated using the addition homomorphic property of homomorphic encryption: ; Since {sk} is generated using a pseudo-random function based on a public seed, the server can calculate it given that all participating clients are aware of it. However, the server is unaware of the individual client's... ; Mask removal and decryption; the server requires all clients to submit their share of a local random mask rk; due to It's a client-side secret; the client won't send it directly. Instead, send a message using As a promise to a certain agreed message, the key pair can only be jointly computed through a secure multi-party computation protocol when the server has collected at least t valid promises. without revealing any individual Subsequently, the server calculates: ; Finally, the server uses the private key Decrypting cfinal yields the aggregated gradient: ; Noise addition and model update; server's aggregated gradient obtained from decryption Add Gaussian noise that meets differential privacy requirements: ; Then use Update global model parameters; this protocol ensures that gradient information of a single client is not inferred by the server or other clients during the aggregation process, while satisfying the definition of centralized differential privacy; in, The encrypted gradient vector uploaded by the kth client. Homomorphic encryption using public key pk For the clipped gradient of the k-th client, A locally generated random mask vector for the k-th client. This is a public mask vector generated based on the master key MK and the client's identity. The sum of encrypted vectors for all clients, The sum of all public masks, The sum of all local random masks, To aggregate the encrypted gradient after removing the mask, The aggregated gradient after decryption. This is the final gradient after adding differential privacy noise. The mean is 0 and the covariance is Gaussian distribution, The standard deviation scaling factor for differential privacy noise. This is the gradient clipping threshold. It is an identity matrix.

[0027] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 process, method, article, or apparatus.

[0028] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An enhanced system for predicting intraoperative blood loss during transurethral resection of the prostate (TURP), characterized in that: Specifically, it includes: End-side multimodal intelligent sensing module: integrated into the end of the surgical instrument, it includes a miniature broadband acoustic sensor assembly, an integrated sweep frequency source optical coherence tomography probe, a three-dimensional torque sensing unit, a multi-band bioelectrical impedance measurement circuit, and a multimodal data synchronization and preprocessing unit, for in situ, real-time acquisition of physical and physiological signals of tissue interaction; The edge-side causal perception and collaborative decision-making engine is deployed in an embedded computer equipped with an image processing unit in the operating room. It includes a high-speed data stream access and buffer queue, a causal structure learning and update module, a potential outcome model and causal effect estimation unit, a counterfactual reasoning and strategy generation unit, and a surgical instrument kinematic analysis and intent recognition unit. It receives edge-side signals, performs causal discovery, intent recognition, risk calculation, and generates collaborative decision-making suggestions and control instructions. Cloud-based federated causal model library and digital twin module: Deployed on the central server, it maintains and updates the global causal knowledge graph and high-fidelity tissue biophysical model, and provides incremental model download and anonymized data upload services for the edge side; Human-computer two-way interaction and dynamic control module: including augmented reality head-mounted display, force feedback surgical instrument master hand, and voice interaction unit, to realize two-way information flow and dynamic and flexible management of control permissions.

2. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 1, characterized in that: The edge-side multimodal intelligent sensing module is specifically: Miniature broadband acoustic sensor assembly: It consists of four piezoelectric composite ultrasonic transducers arranged in a cross shape on the end sidewall of the instrument; each transducer operates in the frequency band of 10kHz-1.5MHz and is equipped with a preamplifier and a bandpass filter; the array realizes sound source localization by measuring the time difference of arrival of sound waves, and the localization algorithm is based on the generalized cross-correlation-phase transformation method; Integrated sweep frequency source optical coherence tomography probe: This probe is coaxially integrated with the instrument working channel and uses a sweep frequency laser source with a center wavelength of 1300nm. B-scan is achieved through fiber optic coupler and miniaturized two-dimensional scanning galvanometer. After the interference signal is received by the balanced detector, the depth reflectivity profile is reconstructed through Fourier transform.

3. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 2, characterized in that: Three-dimensional torque sensing unit: A strain gauge Wheatstone bridge array is set at the proximal end of the instrument. Multi-band bioelectrical impedance measurement circuit: A four-electrode method is used, with two pairs of metal electrodes placed at the end of the instrument; the excitation electrode applies a multi-frequency sinusoidal voltage signal V with a frequency range of 1kHz-1MHz and an amplitude less than 100mVpp. in (f); Measurement electrode collects response current I out (f), the real part Z′(f) and imaginary part Z′′(f) of the impedance are extracted by a lock-in amplifier; then, the measurement spectrum is fitted to the Cole-Cole model by nonlinear least squares method.

4. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 1, characterized in that: The edge-side causal perception and collaborative decision-making engine is specifically as follows: High-speed data stream access and buffer queue: Several first-in-first-out buffer queues are created to receive different modal data streams from the end side; each queue implements a circular buffer structure, and when the amount of data in the buffer reaches the preset window length W, the subsequent processing thread is triggered to ensure real-time performance; Causal Structure Learning and Update Module: This module uses a sliding window approach to process multimodal time series data {X}. t }, X t ∈R d To conduct the analysis; firstly, using PC The algorithm performs conditional independence tests to learn the causal framework; for any pair of variables (X... i X j Given the condition set S, the partial correlation coefficient ρ is calculated. ij |S and check whether it is zero to determine conditional independence; Gaussian process regression was used to test nonlinear conditional independence. Potential Outcome Model and Causal Effect Estimation Unit: For the learned causal graph, a dual machine learning approach is used to estimate the conditional average treatment effect for specific intervention variables T and outcome variables Y. First, two nuisance functions are fitted using random forest or deep neural networks: the outcome prediction model and the propensity score model, respectively. Then, cross-fitting is used to obtain the unbiased estimator.

5. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 4, characterized in that: Counterfactual reasoning and strategy generation unit: Based on the structural causal model, it calculates counterfactual results for the current observation X=x and the proposed intervention; Surgical Instrument Kinematics Analysis and Intent Recognition Unit: This unit receives the instrument end-effector pose q from the robot controller. t ∈SE(3) and the interaction force F from the force sensing module t First, calculate the motion characteristics, including the terminal linear velocity v. t angular velocity ω t Acceleration and force-motion covariance coefficients are used to construct the observation vector Ot, which, together with the tissue type labels segmented from the real-time OCT images, forms the observation vector Ot. A long short-term memory network-conditional random field hybrid model is used to decode the observation sequence and output the basic surgical action label corresponding to each time step. The high-level intent recognizer is based on a hidden semi-Markov model, which maps the basic action sequence to a finite intent state It∈{exploration, dissection, hemostasis}, and its state transition probability and duration distribution are learned through historical surgical data.

6. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 1, characterized in that: The cloud-based federated causal model library and digital twin module are specifically as follows: Differential Privacy Federated Aggregator Server: This server coordinates K clients for federated learning; in each training round, the server allocates the global model parameters θ. g Broadcast to clients; each client k has local dataset D k Calculate the gradient of the loss function The gradient is then clipped to ensure its L2 norm does not exceed a threshold. Subsequently, the client adds Gaussian noise to the gradient: and will Encrypted upload; the server uses a secure multi-party computation protocol to decrypt and aggregate gradients.

7. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 1, characterized in that: The human-computer two-way interaction and dynamic control module is specifically as follows: Dynamic permission decision-maker: This decision-maker calculates the situation-aware safety margin ξ(t) in real time; its inputs include: causal risk estimate R(t), metabolic stress index MSI(t), and counterfactual prediction variance σ. 2 CF(t), Physician-operated tremor power spectral density P tremor (f) Integration in the 8-12Hz frequency band, and the directional consistency ρ between system commands and doctor's operations. align (t); these inputs are mapped to scalars ξ(t) through a normalization function and a learnable multilayer perceptron; the weights of the multilayer perceptron are optimized through reinforcement learning to maximize the long-term safety-efficiency trade-off gains.

8. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 3, characterized in that: It also includes a multimodal data synchronization and preprocessing unit: a built-in field-programmable gate array-based synchronization controller provides a unified hardware trigger clock for all sensors, with a synchronization error of less than 1 microsecond; the raw data first undergoes baseline correction and power frequency notch filtering, then is segmented according to fixed time windows, and timestamped with a time stamp based on a GPS-disciplined atomic clock is added; the preprocessed data stream is sent to the edge side via a gigabit Ethernet interface in the form of User Datagram Protocol (UDP) packets, specifically: Multi-source signal synchronization and resampling; R from the impedance measurement circuit i (t) sequence, StO2(t) sequence from OCT common-path spectrometer, and envelope energy P from a specific frequency band of acoustic sensor. mito The (t) sequence is resampled to a unified time base by linear interpolation at a sampling frequency of 100Hz.

9. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 4, characterized in that: The causal structure learning and updating module employs the following hybrid approach to process causal discovery in multimodal time series data: Variable definition and time delay processing: Each physical quantity in the multimodal data stream is defined as a time series variable; at the same time, possible time delay versions of each variable are defined to construct an extended variable set; Skeleton learning based on conditional independence test; Run PC on the extended variable set Algorithm; For each pair of variables, condition variables are progressively added to the possible condition set S, and partial correlation and kernel-based independence tests are used; the test statistic is compared with the critical value; this stage outputs an undirected graph skeleton; Direction determination and V-structure identification; identification of all potential triples XZY that satisfy the V-structure on the skeleton.

10. The enhanced system for predicting intraoperative bleeding during transurethral resection of the prostate (TURP) according to claim 6, characterized in that: The specific security aggregation protocol executed by the differential privacy federated aggregation server is as follows: Initialization and key distribution: The server generates a pair of homomorphically encrypted public key pk and private key sk for each round of training, and broadcasts pk to all participating clients; at the same time, a secret sharing scheme is adopted to distribute a master key MK to all clients, so that more than t clients are needed to jointly recover the key.

Images