A patient intraoperative anti-falling bed early warning system and method

Abstract

This invention discloses a patient intraoperative fall-prevention early warning system and method, belonging to the field of intelligent operating room monitoring technology. The system includes: a multimodal sensing module for real-time acquisition of multi-dimensional sensing data of the patient's intraoperative status; the multi-dimensional sensing data includes pressure distribution data, posture motion data, optical positioning data, and environmental risk data; an edge computing terminal for performing fusion processing on the multi-dimensional sensing data, including timestamp alignment, filtering, and correlation analysis; an intelligent early warning and braking module for providing graded early warnings based on the probability of a fall and triggering corresponding braking measures; and a central monitoring platform for visually displaying the patient's status and risk level, and managing historical data. This invention can integrate multi-dimensional sensing information and is suitable for patient safety monitoring in scenarios involving anesthesia, fixed positioning, or intraoperative movement.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology in operating rooms, and more specifically, to a patient intraoperative bed fall prevention early warning system and method. Background Technology

[0002] As the core area of ​​a medical institution, patient safety is always the primary concern in the operating room. During surgery, patients may experience impaired consciousness and muscle relaxation under anesthesia, and factors such as surgical positioning and instrument handling can lead to a potential risk of falling off the operating table. Intraoperative falls can not only cause secondary injuries and disrupt the surgical process, but may also trigger medical disputes and liability issues. With the increasing complexity of surgeries and increasingly stringent medical quality standards, establishing an effective intelligent early warning system to prevent intraoperative falls has become an urgent need for modern operating room safety management.

[0003] Currently, bed rail straps combined with manual rounds are mainly used in clinical practice to prevent patients from falling out of bed. However, the repeated disassembly and reassembly of bed rail straps during complex surgeries that require frequent adjustments to patient position is time-consuming and affects surgical efficiency, and it cannot handle sudden agitation in patients emerging from anesthesia. Regarding manual rounds, medical staff need to focus on assisting with the surgical procedure and cannot continuously monitor changes in patient position. While anesthesiologists are responsible for monitoring vital signs, their main attention is focused on respiratory and circulatory indicators, making it difficult to detect minor shifts in patient position in a timely manner. In recent years, single-sensor monitoring has emerged, employing technologies such as pressure sensors, inertial measurement units, and video surveillance. However, in practice, pressure sensors cannot distinguish between active position adjustments and passive slippage, leading to frequent false alarms. Inertial measurement units are susceptible to vibrations from the operating table and electromagnetic interference from instruments such as electrosurgical units. Video surveillance has a limited field of view after the use of sterile surgical drapes and raises privacy concerns.

[0004] It is evident that existing technologies have significant shortcomings: First, the false alarm rate remains high. A single sensor cannot comprehensively determine the true cause of changes in the patient's condition. For example, when medical staff adjust the patient's position, the pressure sensor may detect a shift in the center of pressure, and the high-frequency electromagnetic interference generated during electrosurgical treatment may cause the inertial sensor to misinterpret it as violent movement by the patient. According to clinical statistics, the false alarm rate of existing single-sensor systems is generally above 20%. Second, the response speed is insufficient, leading to missed intervention opportunities. Existing systems mostly rely on post-event analysis, resulting in processing delays. However, there is often only a few seconds between a patient's abnormal positioning and the actual fall, and existing technologies lack real-time linkage with the operating table control system, making it impossible to automatically activate the bed surface braking. Third, there is a lack of adaptability to the special characteristics of the surgical environment. Complex interference factors such as electrosurgical operation, metal instrument reflections, and operating table lifting and lowering, if not effectively identified and filtered, will lead to distorted monitoring data. At the same time, traditional restraint measures are difficult to use in special surgical positions such as the lateral decubitus position or lithotomy position. Therefore, there is an urgent need to develop an intelligent fall prevention bed system that can integrate multi-dimensional sensor information, effectively distinguish between normal operation and real risks, adapt to the complex environment of the operating room, and achieve real-time early warning and automatic intervention.

[0005] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes a patient intraoperative bed fall prevention early warning system and method, which has the advantages of multimodal data fusion for accurate monitoring, real-time risk assessment, and automatic braking intervention, thereby solving the problems of high false alarm rate, delayed response, and lack of coordinated intervention in existing technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, a patient intraoperative bed fall prevention early warning system is provided, comprising: a multimodal sensing module for real-time acquisition of multi-dimensional sensing data of the patient's intraoperative status, including pressure distribution data, posture motion data, optical positioning data, and environmental risk data; an edge computing terminal for performing fusion processing on the multi-dimensional sensing data, including timestamp alignment, filtering, and correlation analysis, calculating the probability of bed fall, and outputting a bed fall early warning signal; an intelligent early warning and braking module for providing graded early warnings based on the probability of bed fall and triggering corresponding braking measures; and a central monitoring platform for visually displaying the patient's status and risk level, and managing historical data. According to another aspect of the present invention, a method for preventing patients from falling off the bed during surgery is also provided. This method includes: real-time acquisition of multi-dimensional sensor data on the patient's intraoperative status based on a multi-modal sensing module; fusion processing of the multi-dimensional sensor data, including timestamp alignment, filtering, and correlation analysis, to calculate the probability of a fall and output a fall warning signal; graded warnings based on the fall risk probability and triggering corresponding braking measures; and visual display of the patient's status and risk level, and management of historical data through a central monitoring platform.

[0008] The beneficial effects of this invention are as follows: (1) This invention effectively solves the problem of high false alarm rate of existing single sensor system by multimodal sensing fusion. The system collects four types of data at the same time: pressure distribution, posture motion, optical positioning and environmental risk. It analyzes the synergistic relationship between them through pressure-posture time-series correlation model. When medical staff adjust the patient's position, pressure change and posture change are synchronous. The system can identify this synergistic mode and judge it as normal operation. However, when a real fall from the bed occurs, it is characterized by asynchronous features where pressure changes first and posture follows. The pressure center shifts rapidly to the edge of the bed, but the trunk posture responds with a lag. The system can accurately distinguish between the two situations based on this. At the same time, the environmental sensor can identify normal operation features such as the operating table motor start signal and the electrosurgical working time window. When these signals are detected, the risk weight is automatically reduced, thereby avoiding false alarms caused by normal surgical operations. This multi-dimensional comprehensive discrimination mechanism significantly reduces the false alarm rate and reduces interference with the surgical process.

[0009] (2) This invention overcomes the shortcomings of traditional monitoring and early warning technologies with slow response. The system uses an edge computing terminal to complete data fusion and risk assessment locally, avoiding the network delay of traditional cloud processing. It can quickly update the probability of falling from the bed and immediately trigger the electromagnetic brake of the operating table to automatically lock the bed surface when an anomaly is detected. The entire response process is much faster than the manual reaction time. In terms of system deployment, the inertial measurement unit is fixed to the patient's torso skin surface with medical adhesive, the pressure sensor array is embedded inside the operating mattress, and the optical positioning unit is deployed on the top of the operating room. The entire system does not occupy the space of the operating area and does not change the existing surgical procedure. In addition, the system provides a graded early warning mechanism. When the risk is low, only light prompts are given, which does not affect the focus of the operation. Only when the risk is confirmed to be high will the braking and voice alarm be activated. While ensuring real-time response, it improves the acceptance and willingness of medical staff to use the system. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of a patient intraoperative bed fall prevention early warning system according to an embodiment of the present invention; Figure 2 This is a specific implementation diagram of an edge computing terminal in a patient intraoperative bed fall prevention early warning system according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating a method for preventing patients from falling off the bed during surgery, according to an embodiment of the present invention.

[0012] In the picture: 1. Multimodal sensing module; 2. Edge computing terminal; 3. Intelligent early warning and braking module; 4. Central monitoring platform. Detailed Implementation

[0013] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0014] According to embodiments of the present invention, a patient intraoperative bed fall prevention early warning system and method are provided.

[0015] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a patient intraoperative bed fall prevention early warning system is provided. The patient intraoperative bed fall prevention early warning system includes: a multimodal sensing module 1, used to collect multi-dimensional sensing data of the patient's intraoperative status in real time, including pressure distribution data, posture motion data, optical positioning data, and environmental risk data; an edge computing terminal 2, used to perform fusion processing on the multi-dimensional sensing data, including timestamp alignment, filtering, and correlation, and calculate the probability of bed fall risk, and output a bed fall early warning signal; an intelligent early warning and braking module 3, used to provide graded early warnings based on the probability of bed fall risk and trigger corresponding braking measures; and a central monitoring platform 4, used to visually display the patient's status and risk level, and manage historical data.

[0016] In one embodiment, when the edge computing terminal 2 performs fusion processing on multi-dimensional sensor data, including timestamp alignment, filtering, and correlation analysis, calculates the probability of a fall from the operating table, and outputs a fall warning signal, the process includes: aligning the timestamps of the multi-dimensional sensor data using a hardware clock synchronization protocol to uniformly convert sensor data from different coordinate systems to the operating table coordinate system; using a sub-sensor filtering scheme to filter the sensor data converted to the operating table coordinate system, establishing a multi-modal dataset to eliminate instrument interference and multipath effects; and based on the multi-modal dataset, establishing a correlation model between sensors and generating multi-modal feature vectors through dynamic weight fusion.

[0017] In one embodiment, the multimodal sensing module 1 includes: a pressure sensor array for embedding within the surgical mattress to monitor the contact pressure distribution between the patient and the bed surface and acquire pressure distribution data; an inertial measurement unit group for wearing on the patient's torso and limbs to collect posture motion data in real time to monitor changes in the patient's posture; an optical positioning unit for being deployed on the ceiling of the operating room to track the patient's optical positioning data and calculate the distance to the edge of the bed; and an environmental sensor for monitoring environmental risk data and providing environmental context information for fall risk assessment.

[0018] It should be noted that this invention proposes a patient intraoperative bed fall prevention early warning system, comprising: a multimodal sensing module 1, specifically including: a pressure sensor array embedded in the surgical mattress for monitoring patient pressure distribution, including contact area and center of gravity shift; an inertial measurement unit (IMU) sensor group, the sensors being fixed to the patient's torso skin surface with disposable medical adhesive, a fixation method that does not require puncture or implantation and can be worn on the patient's torso or limbs for real-time acquisition of acceleration and angular velocity data; an optical positioning unit (UWB) deployed on the operating room ceiling for tracking the patient's three-dimensional spatial position; and environmental sensors for monitoring environmental risk factors such as ground humidity and equipment layout. An edge computing terminal 2 is used to fuse multi-sensor data in real time and calculate the probability of bed fall using a machine learning model. An intelligent early warning and braking module 3 is used to implement tiered early warning, including low-risk audible and visual alarms, medium-risk vibration alerts, and high-risk automatic braking. A central monitoring platform 4 is used to visualize patient status, risk level, and historical data, supporting collaborative management across multiple operating rooms. The multimodal sensing module 1 is connected to the intelligent early warning and braking module 3 through the edge computing terminal 2, and the intelligent early warning and braking module 3 is connected to the central monitoring platform 4.

[0019] It should also be noted that the goal of the edge computing terminal 2 in this invention to fuse multi-sensor data in real time is to unify the processing of data from heterogeneous sensors such as pressure sensor arrays, IMUs, and UWBs, eliminate noise, and extract effective features, such as... Figure 2 As shown, it includes: Step 1: Data Synchronization and Alignment. Multi-dimensional sensor data is timestamped using a hardware clock synchronization protocol, unifying sensor data from different coordinate systems to the operating table coordinate system. This includes timestamp alignment and spatial coordinate unification. During timestamp alignment, data from each sensor is synchronized using a hardware clock synchronization protocol such as PTP, ensuring precise temporal matching of data from different sensors. The spatial coordinate unification process transforms UWB positioning data from the global coordinate system and IMU data from the patient's local coordinate system to the operating table coordinate system, establishing a coordinate reference for subsequent multimodal data fusion.

[0020] Specifically, the coordinate system definitions in this invention are shown in Table 1, which includes the explicit definition of three key coordinate systems. Among them, the UWB global coordinate system takes the corner of the operating room wall as the origin and is mainly used to provide the absolute position information of the patient; the IMU local coordinate system takes the center of the patient's torso as the origin and is mainly used to measure the relative motion of the patient, including acceleration and angular velocity data; the operating table coordinate system takes the geometric center of the operating table as the origin and serves as a reference coordinate system for unifying all sensor data to facilitate risk calculation.

[0021] Table 1 Coordinate System Definition Specifically, the coordinate system transformation step includes two key transformation processes: First, the transformation from UWB global coordinates to the operating table coordinate system is performed, and the UWB origin is aligned to the center of the operating table through translation transformation; then, the transformation from IMU local coordinates to the operating table coordinate system is performed, and the IMU data is transformed to the operating table coordinate system through rotation transformation using attitude quaternions, and translation compensation is performed according to the patient's initial position calibration offset.

[0022] Specifically, the implementation process and technical effectiveness verification of coordinate system transformation are illustrated using the intraoperative calibration procedure as an example. During the preoperative calibration phase, when the patient lies supine on the operating table, the system records the UWB coordinates as the reference position of the origin of the operating table coordinate system. Simultaneously, the IMU is initialized, recording initial quaternions and offset parameters. During the intraoperative real-time transformation phase, the UWB data is directly transformed to the operating table coordinate system through simple translation operations, while the IMU data undergoes real-time rotation transformation using quaternions and the initial offset is superimposed to achieve coordinate unification. Experimental verification shows that, after accuracy verification, the positional error between the UWB and IMU after transformation is less than 2 cm, and the time for a single transformation is less than 1 millisecond, meeting the real-time processing requirements of an edge computing terminal with a computing power of no less than 100MHz.

[0023] In one embodiment, a multi-sensor filtering scheme is used to filter sensor data transformed to the operating table coordinate system to establish a multi-modal dataset. This includes: based on pressure distribution data transformed to the operating table coordinate system, a combination of hardware-level low-pass filtering and software-level Kalman filtering is used to suppress noise and eliminate high-frequency pulses generated by the electrosurgical unit; based on attitude motion data transformed to the operating table coordinate system, a complementary filtering algorithm is used to fuse low-frequency accelerometer signals and high-frequency gyroscope signals with fixed weight coefficients to eliminate accelerometer jitter and gyroscope drift caused by operating table motor vibration; a random sampling consensus algorithm is used on optical positioning data transformed to the operating table coordinate system to determine interior points within the time window to eliminate abnormal positioning points caused by metal instrument reflection or multipath effects; and a multi-modal dataset is generated based on the filtered pressure distribution data, attitude motion data, and optical positioning data, combined with environmental risk data.

[0024] Step 2: Data preprocessing. A multi-sensor filtering scheme is used to filter the sensor data transformed to the operating table coordinate system to establish a multi-modal dataset.

[0025] Specifically, the edge computing terminal employs a sensor-based filtering scheme to filter sensor data transformed to the operating table coordinate system. This process includes three core steps: pressure sensor data preprocessing, IMU sensor attitude calculation, and UWB positioning data calibration. For pressure sensor data, the system uses a combination of hardware-level low-pass filtering and software-level Kalman filtering to suppress noise based on the pressure distribution data transformed to the operating table coordinate system, effectively eliminating high-frequency pulse interference generated by the electrosurgical unit. After filtering, feature extraction is performed, including pressure center trajectory calculation and contact area change rate analysis. The system assesses patient positional stability by calculating the movement trajectory of the pressure center coordinates, while simultaneously monitoring the dynamic changes in the contact area between the patient and the mattress to identify slippage or abnormal positioning.

[0026] It should be noted that the pressure sensor filtering in this invention, i.e., pressure sensor data preprocessing, employs a combined hardware and software filtering approach. Hardware-level filtering involves deploying a low-pass filter at the sensor signal input to suppress high-frequency pulse noise generated by the electrosurgical unit. Software-level filtering uses Kalman filtering combined with wavelet transform to separate noise frequency bands and eliminate electrosurgical interference signals. In this invention, IMU filtering, i.e., IMU sensor attitude calculation, employs a combined complementary filtering and Kalman filtering scheme. In this invention, complementary filtering includes accelerometer calculation of the initial attitude angle, gyroscope integration calculation of the angle, and complementary filtering fusion.

[0027] Specifically, from the accelerometer data a=[ a x , a y , az Calculate the pitch angle θ acc and roll angle φ acc : ; In the formula, a x , a y , a z The components of the accelerometer on the three axes.

[0028] Specifically, the gyroscope measures angular velocity. ω =[ ω x , ω y , ω z ] ω =[ ω x , ω y , ω z The angle is obtained by integration: ; In the formula, Δ t This represents the sampling time interval.

[0029] Specifically, the angle estimate from the accelerometer is combined with the integral of the angular velocity from the gyroscope: ; In the formula, α as The filter gain determines the level of trust in the gyroscope; α The closer it is to 1, the more it relies on the gyroscope, resulting in a fast response but large drift; α The closer it is to 0, the more it relies on the accelerometer, which is stable but has poor dynamic response.

[0030] Specifically, in this invention, Kalman filtering predicts attitude through state equations, corrects noise through measurement equations, and dynamically adjusts the Kalman gain to adapt to operating table vibration interference. The state vector x of the Kalman filter... k =[ θ , φ , w b ] T In the formula, θ is the pitch angle. φ This refers to the roll angle. w b The gyroscope has zero bias; the state prediction equation for the Kalman filter is: ; In the formula, F k The state transition matrix is ​​based on the gyroscope angular velocity integral; w k For process noise, i.e., the covariance matrix Q, a diagonal matrix is ​​used. This is then used to measure the angular velocity of the gyroscope. ω x , ω y Predict the angle at the next moment and estimate the zero bias. w b .

[0031] Specifically, the measurement equation for Kalman filtering is used to correct noise, and its expression is: ; ; In the formula, the measurement vector z k =[ θ acc , φ acc ] T H is calculated by the accelerometer. k This is the observation matrix, used to map the state to the measurement space.

[0032] Specifically, the gain calculation for Kalman filtering: ; In the formula, The prediction error covariance matrix is ​​used; finally, the state is updated, expressed as: .

[0033] It should be noted that this filtering process mainly removes hardware noise such as operating table vibration and gyroscope drift. A true assessment of the risk of a fall from the bed requires combining this with multimodal fusion, as discussed later. In practice, relying solely on IMU data can easily lead to misjudgments. For example, a patient turning over will change their trunk tilt angle, but turning over and a true fall from the bed are different. When turning over, the center of pressure moves very slowly, generally only about 5 cm / s, and the pressure distribution is relatively uniform. The angular velocity measured by the IMU is only about 10 degrees / s, and the entire process lasts relatively smoothly for three or four seconds. However, a true fall from the bed is completely different. The center of pressure will suddenly shift rapidly by more than 10 cm / s, the contact area will decrease by more than 30%, the IMU can detect a sudden change in trunk tilt angle exceeding 30 degrees, a sudden change in angular velocity, and an acceleration close to free fall. More importantly, during a normal turn, the changes in pressure and posture are synchronous, while in a fall from the bed, it is a passive slip, with pressure changing first and posture following. The correlation coefficients between the two are significantly different. Therefore, this invention integrates data from four sensors: pressure, IMU, UWB, and environment. It analyzes the synergistic relationship between them through a pressure-posture time-series correlation model, and then combines environmental sensors to confirm whether there are medical personnel operating or whether there is interference from instruments. This is the only way to accurately distinguish between normal movement and a real risk of falling off the bed, thus avoiding false alarms.

[0034] Specifically, in this invention, UWB filtering employs the RANSAC algorithm to eliminate multipath effects, including: assuming that the UWB positioning data conforms to the linear motion model pt = pt-1 + v × Δt within a 1-second window; randomly selecting two points within the time window to fit the linear model, iterating 100 times; calculating the Euclidean distance d from other points to the model, classifying d ≤ 10cm as inliers and d > 20cm as outliers and removing them; and constraining UWB position updates based on IMU data, with a maximum permissible velocity V. max =2m / s, if the limit is exceeded, IMU data interpolation will be used instead.

[0035] Specifically, the method for calculating the pressure center trajectory is based on real-time calculation of the patient's pressure center coordinates and analysis of their movement trajectory using pressure sensor array data. The pressure center calculation employs a weighted average algorithm, obtaining the pressure center coordinates by multiplying the position coordinates of each sensor unit by its corresponding pressure value, summing the results, and then dividing by the total pressure value. The sensor unit's position coordinates represent its geometric center position in the operating table coordinate system, the pressure value is the real-time pressure data detected by that unit, and the total pressure is the sum of the pressure values ​​from all sensor units. Pressure center trajectory analysis calculates the velocity by dividing the difference in pressure center coordinates between adjacent moments by the time interval. A risk of falling from the bed is identified when the pressure center movement speed exceeds 15 mm / s or the movement distance exceeds the bedside safety zone threshold of 50 mm. The contact area change rate is calculated by multiplying the number of sensor units with pressure values ​​exceeding a preset threshold by the unit area. A contact area change rate exceeding 20% ​​per second is considered an abnormal posture.

[0036] Specifically, the IMU sensor attitude calculation employs a Madgwick filtering algorithm based on quaternion fusion to calculate the patient's torso tilt angle. During the input data preprocessing stage, the system performs unit unification and calibration on the sensor data from the accelerometer, gyroscope, and magnetometer. Static calibration removes zero bias and ensures that the IMU's coordinate axes are aligned with the patient's torso in the forward, backward, left, right, up, and down directions. Quaternion updating, the core process of attitude calculation, first updates the quaternions through gyroscope data integration. A differential equation is used to describe the influence of angular velocity on the quaternions, and this is followed by discretization and updating. Then, accelerometer data is used to correct gyroscope drift, and the quaternion parameters are adjusted by calculating the gravity direction error and using a gradient descent algorithm. The final output is a weighted quaternion fusing gyroscope integration and acceleration correction. The dynamic weighting coefficients are typically set between 0.98 and 0.995, determined based on the sensor signal-to-noise ratio.

[0037] It should be noted that the expression for the pressure center offset velocity in this invention is: ; ; In the formula, p i For the first i Readings from a pressure sensor; x i and y i The first i The horizontal and vertical coordinates of each sensor in the mattress coordinate system; N The total number of pressure sensors (e.g., a 10×10 array is used in this embodiment). N =100); CoP x ( t ) and CoP y ( t ) are respectively t The coordinates of the pressure center at that moment; Δ t The sampling time interval; v CoP The pressure center offset velocity; the expression for the contact area in this invention is: ; In the formula, II( p i > p th ) is an indicator function, if p i > p th If so, it is counted as 1; otherwise, it is counted as 0. p thThe pressure threshold is used; the expression for the rate of change in this invention is: ; If Δ A / Δ t <-0.2cm 2 A value of / ms indicates a sudden decrease in contact area, which may indicate a fall from the bed.

[0038] Specifically, the sensor data is the angular velocity of the gyroscope. w =[ w x , w y , w z [The unit is rad / s; the acceleration of the accelerometer] a =[ a x , a y , a z ], unit m / s 2 .

[0039] It should also be noted that in this invention, quaternions are the core of attitude calculation, and are defined as attitude quaternions. q =[ q 0, q 1, q 2, q 3], of which, q 0 is the real part. q 1, q 2 ,q 3 represents the imaginary part; this quaternion includes gyroscope data integration, accelerometer calibration, and fusion output.

[0040] Specifically, the gyroscope data integration is the differential equation that updates the quaternion through angular velocity, expressed as: ; In the formula, This is quaternion multiplication; w Convert the angular velocity measured by the gyroscope to a pure quaternion [0, w x , w y , w z Discretize and update the sampling period Δ t The expression is: .

[0041] Specifically, this invention employs the Madgwick algorithm to correct gyroscope drift using accelerometer data. First, the gravity direction error is calculated. Assuming the gravity vector should be [0,0,1] under the current attitude, the gravity direction is predicted using quaternions: ; The error vector is then the sum of the accelerometer data (a) and the predicted gravity. cross product: ; In the formula, a For accelerometer data, To predict gravity, gradient descent correction is then performed, adjusting the quaternion based on the error: ; In the formula, β This represents the filter gain.

[0042] Specifically, the final quaternion is a weighted fusion of the gyroscope integral and the acceleration correction, expressed as: ; In the formula, γ The weights are dynamic and depend on the sensor signal-to-noise ratio.

[0043] It should also be noted that, in calculating the torso tilt angle, this invention uses quaternions. q =[ q 0, q 1, q 2, q 3] Solving for Euler angles, the expression for the roll angle is: ; The expression for the pitch angle is: ; The yaw angle does not need to be calculated because the operating table is fixed.

[0044] Specifically, the present invention ensures the reliability of multimodal data through the implementation method of Euler angle calculation using quaternions and UWB positioning data processing. The torso tilt angle is calculated by using quaternion components to calculate the roll and pitch angles. The roll angle is calculated using the arctangent function to determine the ratio of the imaginary components of the quaternions, and the pitch angle is calculated using the arcsine function to determine a specific combination of quaternion components. Since the operating table position is fixed, there is no need to calculate the yaw angle. The anomaly detection mechanism removes acceleration abrupt changes caused by operating table movement by setting thresholds. For UWB optical positioning data, the system uses a random sampling consensus algorithm to determine the interior points of data points within the time window. Multipath effect errors are corrected through dynamic calibration combined with operating table position information, effectively removing abnormal positioning points caused by metal instrument reflections. The filtered pressure distribution data, attitude motion data, optical positioning data, and environmental risk data are combined to finally generate a multimodal dataset for subsequent risk assessment.

[0045] In one embodiment, establishing a sensor inter-modal correlation model based on a multimodal dataset and generating multimodal feature vectors through dynamic weight fusion includes: mapping each modal feature in the multimodal dataset to a shared feature space through a trainable weight matrix to construct a ternary representation of query vector, key vector, and value vector; calculating the correlation score between each modality based on the dot product operation of the query vector and the key vector, converting the correlation score into dynamic weight coefficients through a soft maximization function to form an intermodal attention mapping matrix; establishing a sensor inter-modal correlation model, and using the attention mapping matrix to perform weighted fusion of the value vectors of each modality to generate a multimodal feature vector containing cross-modal correlation information.

[0046] In one embodiment, establishing a sensor-interrelationship model includes: calculating the pressure center offset velocity based on pressure distribution data, calculating the trunk tilt angle change rate based on posture motion data, and constructing a pressure-posture feature pair using the pressure center offset velocity and the trunk tilt angle change rate; performing a time-series window sliding calculation on the pressure-posture feature pair to obtain a pressure-posture cross-correlation coefficient, and establishing a pressure-posture time-series correlation model based on the pressure-posture cross-correlation coefficient; calculating the patient's distance from the bed edge based on optical positioning data, calculating the pressure contact area based on pressure distribution data, verifying the geometric consistency between the patient's distance from the bed edge and the pressure contact area, and outputting a position-pressure spatial consistency coefficient; identifying device interference events and slippery ground events based on environmental risk data, and weighting the pressure-posture time-series correlation model and the position-pressure spatial consistency coefficient based on the confidence level of the device interference events and slippery ground events to generate a sensor-interrelationship model.

[0047] In one embodiment, the pressure center offset velocity is obtained by calculating the displacement of the pressure center positions of adjacent sampling points and dividing by the corresponding time interval; the torso tilt angle change rate is obtained by fusing gyroscope angular velocity data and accelerometer data to calculate attitude quaternions, solving the attitude quaternions to obtain the roll angle and pitch angle, and calculating the ratio of the difference between the roll angle and pitch angle of adjacent sampling points to the time interval.

[0048] Specifically, the expression for the rate of change of the trunk tilt angle is: ω t ( t )=(| φ ( t )- φ ( t -Δ t )|+| θ ( t )- θ ( t -Δ t )|) / Δ t ; In the formula, φ ( t )and θ ( t ) are respectively t The roll and pitch angles at any given moment.

[0049] In one embodiment, establishing a time-series correlation model for posture control based on posture cross-correlation coefficients includes: dividing posture feature pairs into multiple time-series segments according to a preset time step; calculating the Pearson correlation coefficient between the pressure center offset velocity and the rate of change of trunk tilt angle for each time-series segment to obtain a time-series correlation coefficient sequence; the expression for the Pearson correlation coefficient is: ρ(P,I)=Cov(P,I) / (σP×σI); where P represents the pressure center offset velocity sequence, I represents the rate of change of trunk tilt angle sequence, Cov(P,I) represents the covariance of the two sequences, and σP and σI represent the standard deviations of the two sequences, respectively; calculating the moving average correlation coefficient based on the time-series correlation coefficient sequence; marking a synchronous change mode when the moving average correlation coefficient is greater than a preset threshold, and marking an asynchronous change mode when the moving average correlation coefficient is less than a preset threshold; marking time-series segments according to synchronous and asynchronous change modes; and constructing a posture-series correlation model by combining the marking results of the time-series segments with the corresponding posture feature pairs.

[0050] Step 3: The implementation process of establishing a sensor correlation model based on a multimodal dataset first requires constructing attitude feature pairs and a time-series correlation coefficient sequence.

[0051] Specifically, the pressure-posture feature pair refers to a binary data combination formed by combining the pressure center offset velocity and the trunk tilt angle change rate, used to describe the coordination characteristics of patient positional changes. The temporal correlation coefficient sequence is a numerical sequence formed by dividing the pressure-posture feature pair into multiple temporal segments with a 50-millisecond time step, and calculating the Pearson correlation coefficient between the pressure center offset velocity and the trunk tilt angle change rate for each segment. This sequence reflects the evolutionary trend of the correlation strength between pressure changes and posture changes over time. The pressure-posture temporal correlation model is a mathematical model built based on the temporal correlation coefficient sequence. By calculating the moving average correlation coefficient and setting thresholds of 0.7 and 0.3, synchronous and asynchronous change modes are labeled, forming a predictive model describing the temporal relationship between pressure and posture changes.

[0052] Specifically, the attitude timing correlation model in this invention is constructed using an autoregressive moving average model combined with correlation analysis, and its expression is: R(τ) = E[P(t) × I(t + τ)]; In the formula, R(τ) represents the pressure-attitude cross-correlation function with a time delay of τ, and P(t) represents... t The velocity of the pressure center offset at a given moment, I( t +τ) represents t The rate of change of the trunk tilt angle at time +τ, E[·] represents the mathematical expectation operation; The sensor correlation model in this invention adopts a multi-level fusion architecture, and its expression is: Ms = α × Mp + β × Cp + γ × We; In the formula, Mp represents the pressure attitude temporal correlation model, Cp represents the pressure spatial consistency coefficient, We represents the environmental risk weighting coefficient, and α, β, and γ are fusion weight parameters that satisfy α+β+γ=1.

[0053] Specifically, the generation process of the attention mapping matrix and multimodal feature vectors embodies the core mechanism of cross-modal information fusion. The attention mapping matrix is ​​constructed by mapping four modal features (pressure distribution, posture motion, optical positioning, and environmental risk) to a 128-dimensional shared feature space to build query vectors, key vectors, and value vectors. A 4×4 weight matrix is ​​formed by calculating the correlation score using the dot product of the query vector and the key vector, and then converting it into dynamic weight coefficients through a soft maximization function. This matrix quantifies the correlation strength and importance distribution among different sensor modalities. The multimodal feature vector is a 256-dimensional comprehensive feature vector generated by weighted fusion of the value vectors of each modality using the attention mapping matrix. It contains cross-modal correlation information and temporal dynamic features. The first 64 dimensions encode the temporal correlation features of pressure and posture, the middle 64 dimensions encode the spatial consistency features of pressure and posture, and the last 128 dimensions encode the environmental risk weighted features, forming a unified feature representation for subsequent bed fall risk assessment.

[0054] Specifically, the cross-modal relationship modeling in this invention includes: first, dynamically allocating the weights of each modality through an attention mechanism to generate weighted fusion features: ; In the formula, f i These are the original feature vectors for each modality. αi The attention weights are then assigned; next, the features of each modality are mapped to a shared space to construct a ternary representation of the query, key, and value vectors; finally, the weights are calculated, expressed as: ; In the formula, d k is the feature dimension, used to scale gradient stability.

[0055] In one embodiment, when calculating the probability of a bed fall, the edge computing terminal 2 includes: quantifying the feature importance of multimodal feature vectors using a gradient boosting decision tree, and determining the weight coefficients of each feature vector through iterative residual fitting; integrating the weight coefficients of each feature vector into the random forest training process to construct a weighted random forest classifier; during the construction of the weighted random forest classifier, using an improved weighted Gini impurity calculation criterion to split the decision tree; during the training of the weighted random forest classifier, introducing a focus loss function to replace the traditional cross-entropy loss to handle the imbalance problem of bed fall samples, and outputting the probability of a bed fall through a weighted voting mechanism of each decision tree after training.

[0056] In one embodiment, decision tree splitting using an improved weighted Gini impurity calculation criterion includes: calculating the traditional Gini impurity of the training dataset based on a training dataset composed of multimodal feature vectors, and combining this with gradient boosting of the feature weights assigned to the decision tree to form a weighted Gini impurity formula; evaluating the information gain of each candidate split point according to the weighted Gini impurity formula, and selecting the features and split points that minimize the weighted Gini impurity for node splitting; and constructing several decision trees by randomly sampling from the training dataset using a bootstrap aggregation algorithm to complete the decision tree splitting.

[0057] Step 4: When calculating the probability of falling from a bed, edge computing terminal 2 adopts a hybrid model architecture that combines gradient boosting decision tree with weighted random forest. The feature weights generated by GBDT are used to optimize the splitting process of the random forest decision tree.

[0058] It should be noted that the feature extraction and model training process in this invention includes: defining key features such as pressure center offset velocity, IMU attitude change angle, and UWB distance from the bed; normalizing to the [0,1] interval, calculating statistics within a 1-second time window, and generating cross-feature enhancement nonlinear expression; training through gradient boosting trees, statistically analyzing the number of feature splits and the reduction in MSE in all trees, and outputting feature importance scores.

[0059] Specifically, the intraoperative real-time monitoring and feedback optimization in this invention includes: real-time processing at the edge terminal, i.e., updating the probability of falling off the bed every 100ms; triggering a first-level warning with light prompts when the probability is ≥30%, and triggering braking measures when the probability is ≥70%; analyzing false alarms and missed alarms, and updating model parameters through incremental training; adjusting the warning threshold according to FPR and FNR, and using FocalLoss to handle the sample imbalance problem.

[0060] Specifically, the system first uses the LightGBM algorithm as a variant of GBDT to calculate feature importance. A LightGBM classifier is constructed and trained on the training dataset. The weight coefficients of each feature are obtained through the feature contribution score automatically calculated by the model; for example, the weight of the patient's distance from the bed edge is 0.3, and the weight of the trunk tilt angle is 0.4. The gradient boosting process uses an iterative residual fitting mechanism. In each iteration, the prediction residuals of the current model are calculated, and a new decision tree is built to fit these residuals, gradually optimizing the importance scores of features in the model. Finally, normalized feature weight coefficients are output for subsequent use in the weighted random forest.

[0061] Specifically, the weight coefficients of each feature vector are integrated into the random forest training process to construct a weighted random forest classifier. First, bootstrap sampling is performed, resulting in T decision trees. For each decision tree t, n samples are randomly drawn with replacement from the training dataset D composed of multimodal feature vectors, forming a sub-training set Dt, expressed as: Dt={(x t1 ,y t1 ),(x t2 ,y t2 ),...,(x tn ,y tn )};In the formula, (x ti ,y ti The subset D and n are of the same size as the original training set (duplicate samples are allowed); unselected samples are used to validate model performance. A weighted decision tree is then constructed. When training the decision tree on the subset Dt, the splitting criterion introduces feature weights, i.e., the weighted Gini impurity, expressed as: ; In the formula, Gini weighted ( D t ) represents the weighted Gini impurity, used to measure the disorder of the dataset Dt; w j Features assigned to GBDT j The weight; ΔInfoGain j To be based on features jThe information gain during splitting is considered; then, features and split points that minimize the weighted Gini impurity are selected. Finally, the prediction results are aggregated, that is, for the test sample x, the predictions of all trees are combined. The classification task expression for fall detection is: ; In the formula, ht(x) is the prediction result (0 or 1) of the t-th tree; This is an indicator function that returns 1 when the condition is true. The final decision is made using Bagging, which averages the votes from multiple trees to reduce the sensitivity of a single tree to noisy data. The expression is: ; In the formula, Var(h) is the variance of a single tree; P This represents the average correlation coefficient between the trees.

[0062] Specifically, the system utilizes out-of-bag error estimation to directly assess the model's generalization ability without requiring an additional validation set. It ultimately outputs a probability value for bed fall risk between 0 and 1, and implements graded warnings based on the probability range: a probability less than 0.3 indicates a safe state, 0.3 to 0.6 indicates low risk triggering a light warning, 0.6 to 0.8 indicates medium risk triggering a vibration warning, and a probability greater than or equal to 0.8 indicates high risk triggering operating table braking.

[0063] It should be noted that when the intelligent early warning and braking module 3 issues graded warnings based on the probability of falling from the bed and triggers corresponding braking measures, it adopts a dynamic risk threshold adjustment mechanism to distinguish between three different event types: body position adjustment, instrument interference, and actual bed falls. This effectively avoids false alarms and improves the accuracy of the early warning system. The system achieves intelligent identification of event types by comprehensively analyzing the characteristic performance patterns of multimodal sensors. For patient position changes actively performed by medical staff, such as turning over or adjusting to the lithotomy position, the pressure sensor shows a slow movement of the pressure center at a speed of less than 5 cm / s and a uniform change in the contact area. The IMU sensor shows a continuous change in trunk tilt angle with an angular velocity of less than 10 degrees / s, no drastic acceleration abrupt changes, and an acceleration change of less than 1g. The UWB positioning sensor shows that the patient's position change is synchronized with the movement of the operating table, such as the coordinated change of the patient's coordinates when the bed tilts. The environmental sensor can capture warning information such as the operating table motor start signal or the voice commands of medical staff, such as "prepare to adjust body position".

[0064] Specifically, for sensor noise interference events caused by surgical instrument operation, the system avoids false alarms by identifying the unique patterns of instrument interference. Instrument interference is mainly generated by the operation of surgical equipment such as electrosurgical units and suction devices. Its characteristics include: pressure sensors detecting local pressure changes lasting less than 0.5 seconds without overall pressure center displacement; IMU sensors detecting high-frequency vibration signals with frequencies greater than 50Hz that closely match the electrosurgical unit's operating time but without continuous changes in patient posture; UWB positioning sensors experiencing brief signal loss mainly due to metal instrument reflections, but the patient's actual position remains unchanged; and environmental sensors showing a strong correlation between changes in electrosurgical unit power parameters and the timing of sensor noise occurrence. By establishing a correlation model between the instrument operation time window and abnormal sensor signals, the system can accurately identify and filter false alarms caused by instrument interference.

[0065] Specifically, for real-life fall-from-bed risks caused by unexpected patient movement, the system uses multi-dimensional feature fusion analysis to accurately identify and immediately trigger corresponding braking and warning measures. Real-life fall-from-bed risks, such as limb slippage after anesthesia, exhibit clear danger indicators: pressure sensors detect rapid shifts in the center of pressure with a speed greater than 10 cm / s, a sudden decrease in contact area, or even a unilateral pressure drop to zero; IMU sensors detect sudden changes in trunk tilt angle greater than 30 degrees for more than 1 second, accompanied by free-fall acceleration characteristics approaching 1g; UWB positioning sensors indicate the patient's position exceeds the operating table boundary and is more than 15 cm from the edge, a safety threshold; environmental sensors confirm the absence of instrument operation signals, and a floor humidity risk score greater than 0.7 indicates a slippery floor and additional fall risk. Through this intelligent discrimination mechanism based on multi-sensor feature fusion, the system can immediately trigger high-level warnings and activate the operating table braking device when a real-life fall-from-bed risk occurs, effectively protecting patient safety.

[0066] Specifically, the intelligent early warning and braking module 3 implements graded early warning standards based on the combination of the probability of falling from the bed and risk characteristics. The early warning level is dynamically divided according to the combination of the probability of falling from the bed (P) and risk characteristics. The specific standards are shown in Table 2.

[0067] Table 2 Warning Level Standards Among them, dynamic adjustment means that if the same patient triggers two medium-risk warnings in a row within 10 seconds, the system will automatically upgrade to high-risk; manual intervention is prioritized, meaning that medical staff can manually downgrade the warning, which will be linked with the operating table control system to trigger emergency braking or adjust the tilt angle of the bed.

[0068] Specifically, in this embodiment, the triggering condition for a low-risk warning is that the probability of falling from the bed, P, is between 0.3 and 0.6, and a single risk characteristic, such as the center of pressure moving at a speed greater than 5 cm / s, is detected. The system response is for the operating room LED light strip to flash yellow light at a frequency of 1Hz and for a pop-up message on the central monitoring platform to indicate "Low Risk: Patient Position Change". The triggering condition for a medium-risk warning is that the probability of falling from the bed, P, is between 0.6 and 0.8, and for a combination of two risk characteristics, such as the center of pressure moving at a speed greater than 8 cm / s and the trunk tilt angle greater than 20 degrees. The system response is for the smart bracelet worn by medical staff to generate a short vibration pattern of 3 times as a reminder and to lock the manual adjustment function of the operating table, requiring a password to unlock. The high-risk warning is triggered when the probability of falling from bed, P, is greater than or equal to 0.8 and three risk characteristics are combined, such as the pressure center moving at a speed greater than 10 cm / s, the IMU detecting free fall acceleration, and the UWB showing the patient's position exceeding 15 cm from the edge of the bed. The system response measures are to immediately trigger the electromagnetic brake of the operating table to lock the bed surface to prevent further tilting, broadcast a voice alarm "High risk of falling from bed! Bed number A3" on the central monitoring platform, and automatically call the emergency team for emergency handling through the hospital's Internet of Things system.

[0069] like Figure 3 As shown, according to another embodiment of the present invention, a method for preventing patients from falling off the bed during surgery is also provided. The method includes: real-time acquisition of multi-dimensional sensor data of the patient's intraoperative status based on a multi-modal sensor module 1; fusion processing of the multi-dimensional sensor data, including timestamp alignment, filtering, and correlation analysis, and calculation of the probability of falling off the bed, and outputting a fall warning signal; graded warning based on the probability of falling off the bed and triggering corresponding braking measures; and visual display of the patient's status and risk level, and management of historical data through a central monitoring platform 4.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A patient intraoperative bed fall prevention early warning system, characterized in that, include: A multimodal sensing module is used to collect multi-dimensional sensing data of the patient's intraoperative status in real time. The multi-dimensional sensing data includes pressure distribution data, posture motion data, optical positioning data, and environmental risk data. Edge computing terminals are used to perform fusion processing on multi-dimensional sensor data, including timestamp alignment, filtering, and correlation analysis, and to calculate the probability of a bed fall and output a bed fall warning signal. The intelligent early warning and braking module is used to provide graded early warnings based on the probability of falling from the bed and to trigger corresponding braking measures. The central monitoring platform is used to visualize patient status and risk levels, and to manage historical data.

2. The patient intraoperative bed fall prevention early warning system according to claim 1, characterized in that, When the edge computing terminal performs fusion processing on multi-dimensional sensor data, including timestamp alignment, filtering, and correlation analysis, it includes: The hardware clock synchronization protocol is used to timestamp and align the multi-dimensional sensor data, and the sensor data from different coordinate systems are uniformly converted to the operating table coordinate system. A multi-sensor filtering scheme is adopted to filter the sensor data transformed to the operating table coordinate system and establish a multi-modal dataset to eliminate the effects of instrument interference and multipath effects. Based on a multimodal dataset, a correlation model between sensors is established, and multimodal feature vectors are generated through dynamic weight fusion.

3. The patient intraoperative bed fall prevention early warning system according to claim 2, characterized in that, The process of using a multi-sensor filtering scheme to filter sensor data transformed to the operating table coordinate system and establishing a multimodal dataset includes: Based on the pressure distribution data transformed to the operating table coordinate system, a combination of hardware-level low-pass filtering and software-level Kalman filtering is used to suppress noise and eliminate high-frequency pulses generated by the electrosurgical unit. Based on the attitude motion data transformed to the operating table coordinate system, a complementary filtering algorithm is used to fuse the low-frequency signal from the accelerometer and the high-frequency signal from the gyroscope by using fixed weight coefficients, so as to eliminate accelerometer jitter and gyroscope drift caused by the vibration of the operating table motor. The optical positioning data converted to the operating table coordinate system is subjected to a random sampling consensus algorithm to determine the interior points of the data points within the time window, so as to eliminate abnormal positioning points caused by metal instrument reflection or multipath effect. Based on the filtered pressure distribution data, attitude motion data, and optical positioning data, combined with environmental risk data, a multimodal dataset is generated.

4. The patient intraoperative bed fall prevention early warning system according to claim 1, characterized in that, The process of establishing a sensor correlation model based on a multimodal dataset and generating multimodal feature vectors through dynamic weight fusion includes: The modal features in the multimodal dataset are mapped to a shared feature space through a trainable weight matrix to construct a ternary representation of query vector, key vector, and value vector; The relevance scores between each modality are calculated based on the dot product operation of the query vector and the key vector. The relevance scores are then converted into dynamic weight coefficients through a soft maximization function to form an attention mapping matrix between modalities. A model of the correlation between sensors is established, and the value vectors of each modality are weighted and fused using an attention mapping matrix to generate a multimodal feature vector containing cross-modal correlation information.

5. The patient intraoperative bed fall prevention early warning system according to claim 4, characterized in that, The establishment of the sensor inter-sensor correlation model includes: The pressure center offset velocity is calculated based on pressure distribution data, and the trunk tilt angle change rate is calculated based on posture motion data. The pressure center offset velocity and the trunk tilt angle change rate are then combined to form a pressure-posture feature pair. A temporal window sliding calculation was performed on the posture feature pairs to obtain the posture cross-correlation coefficient, and a posture temporal correlation model was established based on the posture cross-correlation coefficient. The distance between the patient and the edge of the bed is calculated based on optical positioning data, and the pressure contact area is calculated based on pressure distribution data. The geometric consistency between the distance between the patient and the edge of the bed and the pressure contact area is verified, and the spatial consistency coefficient of the pressure-positioning space is output. Based on environmental risk data, instrument interference events and ground slippery events are identified. The pressure-attitude time-series correlation model and the spatial consistency coefficient of pressure are weighted according to the reliability of the instrument interference events and ground slippery events to generate a correlation model between sensors.

6. The patient intraoperative bed fall prevention early warning system according to claim 5, characterized in that, The pressure center offset velocity is obtained by calculating the displacement of the pressure center positions of adjacent sampling points and dividing by the corresponding time interval; The expression for the pressure center offset velocity is: ； ； In the formula, p i For the first i Readings from a pressure sensor; x i and y i The first i The horizontal and vertical coordinates of each sensor in the mattress coordinate system; N CoP represents the total number of pressure sensors. x ( t ) and CoP y ( t ) are respectively t The coordinates of the pressure center at that moment; Δ t The sampling time interval; v CoP This refers to the velocity of the pressure center offset. The rate of change of the torso tilt angle is calculated by fusing gyroscope angular velocity data and accelerometer data to obtain the attitude quaternion. The roll angle and pitch angle are obtained from the attitude quaternion, and the ratio of the difference between the roll angle and pitch angle of adjacent sampling points to the time interval is calculated.

7. A patient intraoperative bed fall prevention early warning system according to claim 5, characterized in that, The establishment of the posture time-series correlation model based on posture cross-correlation coefficients includes: The pressure posture feature pair is divided into multiple time segments according to a preset time step. The Pearson correlation coefficient between the pressure center offset velocity and the rate of change of the torso tilt angle is calculated for each time segment to obtain a time correlation coefficient sequence. The moving average correlation coefficient is calculated based on the time-series correlation coefficient sequence. When the moving average correlation coefficient is greater than a preset threshold, it is marked as a synchronous change mode, and when the moving average correlation coefficient is less than the preset threshold, it is marked as an asynchronous change mode. The timing segments are labeled according to synchronous and asynchronous change modes. By combining the labeling results of the timing segments with the corresponding posture feature pairs, a posture timing correlation model is constructed.

8. The patient intraoperative bed fall prevention early warning system according to claim 1, characterized in that, When calculating the probability of a bed fall, the edge computing terminal includes: Gradient boosting decision trees are used to quantify the feature importance of multimodal feature vectors, and the weight coefficients of each feature vector are determined by iterative residual fitting. The weight coefficients of each feature vector are integrated into the random forest training process to construct a weighted random forest classifier; In the construction of the weighted random forest classifier, an improved weighted Gini impurity calculation criterion is used for decision tree splitting; During the training process of the weighted random forest classifier, a focus loss function is introduced to handle the imbalance problem of bed fall samples, and the bed fall risk probability is output through the weighted voting mechanism of each decision tree after training is completed.

9. A patient intraoperative bed fall prevention early warning system according to claim 8, characterized in that, The decision tree splitting using the improved weighted Gini impurity calculation criterion includes: Based on the training dataset composed of multimodal feature vectors, the traditional Gini impurity of the training dataset is calculated, and combined with the feature weights assigned by the gradient boosting decision tree, a weighted Gini impurity formula is formed. The information gain of each candidate split point is evaluated according to the weighted Gini impurity formula, and the feature and split point that minimizes the weighted Gini impurity are selected for node splitting. The decision tree split is completed by randomly sampling from the training dataset using a bootstrap aggregation algorithm to construct several decision trees.

10. A method for preventing patients from falling off the operating table during surgery, employing the patient intraoperative bed-fall prevention early warning system according to any one of claims 1-9, characterized in that, The method includes: Based on the multimodal sensing module, multi-dimensional sensing data of the patient's intraoperative status are collected in real time; Multi-dimensional sensor data is fused, including timestamp alignment, filtering, and correlation analysis, and the probability of a bed fall is calculated, and a bed fall warning signal is output. The system provides tiered warnings based on the probability of a fall from the bed and triggers corresponding braking measures. The central monitoring platform visualizes patient status and risk levels and manages historical data.

Images