An intelligent management platform for whole-cycle off-hospital rehabilitation of severe pancreatitis patients

Abstract

The application relates to the technical field of medical rehabilitation monitoring, and discloses a whole-cycle hospital-outside rehabilitation intelligent management platform for a severe pancreatitis patient, which comprises an interactive terminal, food point cloud data of which is acquired to determine an initial total physical volume; a server, from which a single maximum safe intake volume calculated in combination with a patient tolerance baseline is acquired, and on the basis of which a three-dimensional virtual boundary model is generated and superimposed on a real food surface; during eating, food consumption is continuously calculated to generate an intake volume slice sequence. A perception node monitors a physiological state of an abdomen of the patient to output a tension fluctuation sequence, and a volume tension buffer gradient is generated by operating the intake volume slice sequence and the tension fluctuation sequence. When the gradient exceeds a threshold value, the interactive terminal changes a rendering feature of the virtual boundary model to output a visual cut-off warning. The server acquires a historical sequence of corresponding data to complete whole-cycle risk assessment. The application associates physical intake and digestive tract stress state data, and reduces the risk of volume overload.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation monitoring technology, specifically to an intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis. Background Technology

[0002] For patients with severe pancreatitis undergoing outpatient rehabilitation, strict management of food volume is crucial to preventing relapse. Current dietary management protocols typically rely on patients or their families to manually estimate food volume based on experience. This lack of a real-time, objective correlation between external physical intake and the stress levels within the digestive tract is problematic. This static, experience-based estimation method cannot dynamically adjust the single-session food restriction standard according to the patient's current pathological recovery and clinical scores, making it highly susceptible to relapse due to overloading of the food volume in a single meal.

[0003] Meanwhile, when using surface sensors to acquire abdominal physiological monitoring data in routine out-of-hospital environments, the collected signals are easily affected by the patient's physical activity. Changes in trunk posture and diaphragmatic displacement during respiration can superimpose significant physical deformation interference and electrophysiological baseline drift into the sensor acquisition channel. Current technologies lack a joint noise stripping mechanism for different spatial postures and respiratory phases, failing to effectively eliminate mechanical interference. This leads to decreased accuracy of extracted tension fluctuations and gastric electrical signals, making them unreliable underlying data for assessing gastrointestinal tension.

[0004] Furthermore, when a patient's food intake is detected to be approaching or reaching a risk threshold, existing auxiliary management systems mostly rely on conventional audible alarms or text pop-ups for notification. In the absence of mandatory medical supervision outside of hospitals, these routine alarms are easily ignored by patients due to subjective appetite or habits, leading to delayed responses. Because of the lack of an objective and mandatory intervention mechanism, interventions for overeating often cannot be reliably implemented, resulting in ineffective food management. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis. This platform solves the problems of mismatch between the physical intake of external food and the physiological stress state inside the digestive tract during the outpatient rehabilitation period of patients with severe pancreatitis, and the fact that surface monitoring signals are easily affected by trunk posture and respiratory movements, leading to judgment errors.

[0006] To achieve the above objectives, this invention provides an intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis, comprising: A sensing node is used to monitor the patient's abdominal physiological state and output a tension fluctuation sequence; and to calculate and generate a volume tension buffer gradient based on the intake volume slice sequence sent by the interactive terminal and the tension fluctuation sequence. An interactive terminal is used to acquire food point cloud data and determine the initial total physical volume, obtain the maximum safe intake volume for a single intake from the server and generate a three-dimensional virtual boundary model based on the maximum safe intake volume for a single intake, calculate the food consumption based on continuously acquired food point cloud data and the initial total physical volume during the patient's eating process, generate an intake volume slice sequence and send it to the sensing node; and use the volume tension buffer gradient to change the rendering features of the three-dimensional virtual boundary model and output a visual truncation warning. The server is used to calculate the maximum safe intake volume per session based on the patient's comprehensive baseline data and send it to the interactive terminal; and to acquire the historical sequence of the volumetric tension buffer gradient and the record of the visual truncation warning, thereby completing the intelligent management of the patient's full-cycle outpatient rehabilitation.

[0007] Preferably, the server includes a secure access management module, used for: Retrieve the preset standard gastric emptying volume reference constant and obtain the patient's comprehensive baseline data; Extract the comprehensive tolerance assessment baseline from the comprehensive baseline data and multiply it by the baseline volume adjustment weight to obtain the tolerance baseline product. Add the tolerance baseline product to the preset benchmark relaxation constant to obtain the baseline relaxation factor. The clinical score is extracted from the comprehensive baseline data and multiplied by the clinical score penalty coefficient to obtain the clinical penalty product. The negative value of the clinical penalty product is used as the exponent of the natural constant to obtain the penalty decay factor. The maximum safe intake volume for a single dose is calculated by multiplying the standard gastric emptying volume reference constant, the baseline relaxation factor, and the penalty attenuation factor, and then the maximum safe intake volume for a single dose is sent to the interactive terminal.

[0008] Preferably, the interactive terminal includes: a depth sensing component for acquiring time-of-flight point cloud data of target objects within the field of view for food identification and volume calculation; a positioning component for capturing the patient's gesture coordinates and mapping them to the abdominal surface to generate a pain space matrix and outputting it to the sensing node; and a rendering component for generating a three-dimensional virtual boundary model based on the single maximum safe intake volume and performing superimposed rendering, and dynamically adjusting the rendering features of the three-dimensional virtual boundary model based on the volume tension buffer gradient according to the hardware interrupt interception command sent by the sensing node and outputting a visual truncation warning.

[0009] Specifically, the interactive terminal includes: a depth perception component, used to collect time-of-flight point cloud data of target objects within the field of view for food feature recognition, initial total physical volume of food and real-time intake volume calculation; a positioning component, used to run a synchronous positioning and mapping algorithm, capture the patient's gesture coordinates and map them onto the abdominal surface, generate a pain space matrix, and output the pain space matrix to the perception node; a rendering component, used to generate a three-dimensional virtual boundary model based on the single maximum safe intake volume, and perform perspective overlay rendering of the three-dimensional virtual boundary model and real food; and a component for receiving a hardware interruption interception command sent by the perception node, changing the rendering features of the three-dimensional virtual boundary model according to the hardware interruption interception command and outputting a visual truncation warning.

[0010] Preferably, when generating a 3D virtual boundary model and completing perspective overlay rendering, the rendering component is specifically used for: Using the geometric center of the target food as the origin, a three-dimensional virtual boundary model of the corresponding spatial volume is generated based on the single maximum safe intake volume, and the determinant value of the third-order rotation sub-matrix contained in the pose tracking matrix output by the positioning component in real time is calculated. When the absolute difference between the determinant value of the third-order rotation submatrix and the preset unit determinant reference value is greater than the preset rotation distortion threshold, the inverse matrix of the pose tracking matrix of the previous effective visual sampling period is invoked. When the absolute difference is not greater than the rotational distortion threshold, the inverse matrix of the pose tracking matrix in the current visual sampling period is used. Using the inverse matrix that is invoked or adopted, the three-dimensional virtual boundary model is transformed from the objective world coordinate system to the local camera coordinate system, and the three-dimensional virtual boundary model is rendered so that the three-dimensional virtual boundary model is overlaid on the real food surface after perspective alignment.

[0011] Furthermore, the rendering component is also used to: dynamically adjust the rendering features of the three-dimensional virtual boundary model according to the comprehensive pain weight parameter corresponding to the pain space matrix output by the positioning component and the volume tension buffer gradient fed back by the perception node, and output a visual truncation warning signal when the volume tension buffer gradient exceeds a preset threshold.

[0012] Preferably, the sensing node includes: an electromyographic electrode for acquiring electrogastric signal; an inertial measurement unit for acquiring spatial posture matrix data under different standard body positions; a strain sensor for acquiring abdominal wall tension data under different standard body positions; and a smart chip for obtaining a postural respiratory baseline mapping table based on the electrogastric signal, the spatial posture matrix data, and the abdominal wall tension data.

[0013] Preferably, the smart chip is specifically used for: Simultaneously acquire spatial posture matrix data and abdominal wall tension data under different standard body positions; The static stretch offset caused by trunk posture and the periodic respiratory fluctuation caused by diaphragmatic movement are separated from the abdominal wall tension data, and the respiratory timing phase is determined based on the periodic respiratory fluctuation. The pitch and roll angles in the spatial attitude matrix data are set as the primary index dimensions, the breathing timing phase is set as the secondary index dimension, and a multi-dimensional data grid is constructed. The static stretch offset and the periodic breathing fluctuation are added to obtain the tension deformation superposition value. The pitch angle, roll angle, and respiratory timing phase are cross-combined to form grid coordinates, and the corresponding tension deformation superposition values ​​are filled into the nodes of the multidimensional data grid to generate a body position respiratory baseline mapping table.

[0014] Preferably, the smart chip is further configured to perform mechanical noise stripping on the acquired electrogastric signal and abdominal wall tension data based on the body position respiratory baseline mapping table, and output a pure tension fluctuation sequence and a pure electrogastric sequence.

[0015] Preferably, the smart chip outputs a pure tension fluctuation sequence and a pure gastric electrical sequence, specifically including: The pitch angle, roll angle and breathing timing phase extracted in real time are input into the body position breathing baseline mapping table to obtain the mechanical interference compensation value corresponding to the current physical state; The mechanical interference compensation value is subtracted from the real-time acquired abdominal wall tension data to output a pure tension fluctuation sequence as the tension fluctuation sequence; Simultaneously, the mechanical interference compensation value is converted into an electrophysiological baseline drift using a preset electromechanical conversion mapping coefficient, and a pure gastric electrosequence is output by subtracting the electrophysiological baseline drift from the real-time acquired gastric electrogram signal.

[0016] Preferably, the positioning component is further configured to capture the three-dimensional coordinates of the gesture, generate a pain space matrix, and send it to the sensing node.

[0017] Preferably, the depth sensing component is further configured to acquire time-of-flight point cloud data and construct a two-dimensional planar mapping mesh containing multiple discrete mesh nodes in the objective world coordinate system; The positioning component is specifically used to calculate the vertical projection point of the absolute three-dimensional coordinates of the hand feature points in the objective world coordinate system onto the two-dimensional plane mapping grid as the pain center coordinates, and to calculate the square of the Euclidean distance between the discrete grid nodes and the pain center coordinates on the two-dimensional plane. The square of the Euclidean distance is divided by a preset pain radiation space variance parameter to obtain the attenuation variable. The negative value of the attenuation variable is used as the exponent of the natural constant to calculate the pain weight value. The pain weight values ​​corresponding to all the discrete grid nodes are filled into a matrix data structure to generate a pain space matrix, and the pain space matrix is ​​sent to the sensing node.

[0018] Preferably, the smart chip is further used to perform buffering operations on the discrete intake volume slice sequence to obtain an effective buffer volume sequence.

[0019] Preferably, the step of performing a buffering operation on the discrete ingestion volume slice sequence to obtain an effective buffer volume sequence specifically includes: Establish a time sliding window, extract all historical intake volume slice values ​​recorded within the time sliding window, and extract the physical time difference experienced from the current visual sampling cycle. Divide the physical time difference by a preset viscoelastic relaxation time to obtain the time decay ratio. Take the negative value of the time decay ratio as the exponent of the natural constant to calculate the time-varying weight. Multiply all the historical intake volume slice values ​​by the corresponding time-varying weights and perform discrete summation to output the effective buffer volume sequence.

[0020] Preferably, the sensing node generates the volumetric tension buffer gradient using the intake volume slice sequence and the tension fluctuation sequence via the smart chip, specifically including: Extract the pure tension value of the pure tension fluctuation sequence at the current unified sampling point and the pure tension value of the historical sampling points after a preset differential time window length. Subtract the pure tension value of the current unified sampling point from the pure tension value of the historical sampling points to obtain the macroscopic tension physical increment. Extract the effective buffer volume sequence at the current unified sampling point and the effective buffer volume sequence at historical sampling points, and subtract the effective buffer volume sequence at the current unified sampling point from the effective buffer volume sequence at historical sampling points to obtain the macroscopic equivalent intake volume increment; When it is determined that the macroscopic equivalent intake volume increment is greater than the preset volume increment threshold, the macroscopic tension physical increment is divided by the macroscopic equivalent intake volume increment to obtain the basic tension response slope. The comprehensive pain weight parameter is calculated based on the pain space matrix. The pain weight value is used as the comprehensive pain weight parameter and multiplied by the preset pain amplification gain coefficient to obtain the gain product. The gain product is added to the preset gain reference constant and then multiplied by the basic tension response slope to output the volume tension buffer gradient.

[0021] Specifically, the comprehensive pain weight parameter is calculated based on the pain space matrix by: traversing the pain space matrix, sorting the pain weight values ​​in descending order, and taking the arithmetic mean of the top preset number of pain weight values ​​as the comprehensive pain weight parameter.

[0022] Preferably, the interactive terminal utilizes the volumetric tension buffer gradient to change the rendering features of the three-dimensional virtual boundary model and outputs a visual truncation warning, specifically including: When the volume tension buffer gradient generated by several consecutive visual sampling cycles that meet the preset confirmation period is greater than the preset tolerance gradient threshold, the smart chip sends a hardware interrupt interception command to the interactive terminal. The interactive terminal suspends the scanning thread of the food point cloud data based on the hardware interrupt interception instruction, and forces the transparency parameter in the material shader of the three-dimensional virtual boundary model to be overwritten to zero through the rendering component, and switches the diffuse color channel of the three-dimensional virtual boundary model to red warning, thereby outputting a warning color rendering signal as the visual truncation warning.

[0023] Preferably, the server acquires the historical sequence of the volumetric tension buffer gradient and the record of the visual truncation warning, as well as the comprehensive pain weight parameter and the pure gastric electrical sequence, to complete the intelligent management of the patient's full-cycle outpatient rehabilitation.

[0024] Preferably, the server further includes a full-cycle rehabilitation management module, specifically used for: Extract the total number of times the visual truncation warning is triggered within a preset evaluation period from the records of the visual truncation warning. Divide the total number of times by the physical time span of the evaluation period and perform normalization to obtain the average forced blocking frequency feature. Simultaneously, the arithmetic mean of all historical sequences of the volumetric tension buffer gradients within the evaluation period is extracted and normalized to obtain the average gastrointestinal stress sensitivity characteristics. The arithmetic mean of all the comprehensive pain weight parameters within the evaluation period is extracted and normalized to obtain the average pain characteristics; The average offset of the pure gastric electroencephalogram sequence relative to the preset steady-state baseline of the gastric electroencephalogram is extracted and normalized to obtain the average offset feature; By using preset cutoff frequency penalty weights, tension gradient influence weights, pain feedback attenuation weights, and gastric electrical deviation penalty weights, the average forced blockage frequency feature, the average gastrointestinal stress sensitivity feature, the average pain feature, and the average deviation feature are weighted and summed to output a full-cycle risk index, thus completing the intelligent management of the patient's full-cycle outpatient rehabilitation.

[0025] This invention calculates the maximum safe intake volume per feeding based on the patient's comprehensive baseline data via a server. An interactive terminal performs volume measurement on food point cloud data, generates an intake volume slice sequence and a three-dimensional virtual boundary model, and combines this with real-time monitoring of the patient's abdominal physiological state by sensing nodes to calculate the volume tension buffer gradient. This allows for the correlation and matching of external physical food intake with the physiological stress state within the digestive tract. Simultaneously, the joint analysis of abdominal physiological state monitoring and intake volume calculation reduces interference from trunk posture and respiratory movements on surface monitoring and judgment, improving the accuracy of dietary management during outpatient rehabilitation. This enables dynamic limitation of single feeding amounts, reducing the risk of relapse due to overeating. Attached Figure Description

[0026] Figure 1 This is an overall architecture diagram of the intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis, as presented in this invention. Figure 2 The flowchart of the intelligent management method for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to the present invention is shown below; Figure 3 This is a schematic diagram of the mechanical noise stripping and clean timing signal extraction logic of the present invention; Figure 4 This is a schematic diagram of gesture coordinate capture and pain space matrix mapping according to the present invention; Figure 5 This is a schematic diagram illustrating the calculation of the viscoelastic delay response decay and effective buffer volume sequence of the present invention; Figure 6 This is a schematic diagram of the volume tension buffer gradient calculation and high-frequency intervention truncation of the rendering boundary in this invention; Figure 7 This is a diagram illustrating the outpatient full-cycle risk assessment and closed-loop feedback interaction of medical staff and patients, as presented in this invention. Figure 8 This is a simulation diagram of the feeding volume monitoring and viscoelastic buffering conversion data of the present invention; Figure 9 This is a simulation diagram comparing the original mechanical signal and the pure tension after stripping noise in this invention. Figure 10 This is a simulation diagram of the volumetric tension buffer gradient and closed-loop cutoff data of the present invention. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings.

[0028] See Figure 1 This invention provides an intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis, comprising: sensing nodes, an interactive terminal, and a server. The sensing nodes are deployed on the patient's abdominal surface. Each sensing node includes a smart chip and electromyographic electrodes for acquiring electrogastrogram (EGG) signals, strain sensors for acquiring abdominal wall tension data, and an inertial measurement unit for acquiring spatial attitude matrix data, all connected to the smart chip via data communication. The interactive terminal can be an augmented reality head-mounted display worn by the user, comprising a depth sensing component, a positioning component, and a rendering component. The depth sensing component acquires time-of-flight point cloud data of target objects within the field of view; the positioning component captures gesture coordinates and generates a pain space matrix; and the rendering component generates and outputs a computer graphics overlay. A local area network (LAN) communication connection is established between the interactive terminal and the sensing nodes. The server establishes WAN communication connections with both the sensing nodes and the interactive terminal, for storing comprehensive baseline data, distributing control parameters, and performing comprehensive risk index calculations.

[0029] like Figure 2 As shown, the present invention provides a method for intelligent management of full-cycle outpatient rehabilitation for patients with severe pancreatitis, comprising: S1. In-hospital baseline anchoring and edge perception initialization. The server acquires the patient's comprehensive baseline data upon discharge, including the comprehensive tolerance assessment baseline, electrogastrogram steady-state baseline, and clinical score, and distributes this comprehensive baseline data to the smart chip at the sensing node. The sensing node simultaneously collects abdominal wall tension data and spatial posture matrix data in three standard body positions: supine, sitting, and upright. The smart chip generates and solidifies a postural respiratory baseline mapping table based on the abdominal wall tension data and spatial posture matrix data.

[0030] S2. Visual Feature Extraction and Asymmetric Sampling Control. The interactive terminal identifies eating features through the depth sensing component and sends spatiotemporal synchronization pulses to the sensing nodes. After receiving the spatiotemporal synchronization pulses, the intelligent chip triggers a heterogeneous asymmetric sampling strategy. The intelligent chip calls the body position and respiratory baseline mapping table and performs mechanical noise stripping processing logic on the gastric electrocardiogram electrical signals collected by the electromyography electrodes and the abdominal wall tension data collected by the strain sensor, outputting a pure tension fluctuation sequence and a pure gastric electrocardiogram sequence.

[0031] S3. Objective Intake Calculation and Virtual Boundary Rendering Steps. The interactive terminal calculates the initial total physical volume based on the food point cloud data obtained by the depth perception component, and obtains the maximum safe intake volume for a single consumption from the server. The rendering component generates a 3D virtual boundary model based on the maximum safe intake volume for a single consumption, and overlays and renders the 3D virtual boundary model onto the surface of the real food. During the eating process, the interactive terminal continuously calculates the consumption of food point cloud data and generates a discrete sequence of intake volume slices.

[0032] S4. Cross-modal voxel fusion and delay buffer truncation steps. The interactive terminal captures the 3D coordinates of gestures through the positioning component, generating a pain space matrix. The smart chip uses a sliding window algorithm to perform convolution operations on the ingested volume slice sequence, and calculates the volumetric tension buffer gradient by combining the first derivative of the pure tension fluctuation sequence. When the volumetric tension buffer gradient reaches a preset tolerance gradient threshold, the interactive terminal controls the rendering component to change the transparency of the 3D virtual boundary model and outputs a warning color rendering signal.

[0033] S5. Risk Closed-Loop Assessment and Log Generation Steps. The intelligent chip exits the heterogeneous asymmetric sampling strategy after the eating action is completed. The server acquires the event records and comprehensive pain weight parameters of the high-frequency intervention cutoff logic triggering the rendering boundary uploaded by the interactive terminal, as well as the pure tension fluctuation sequence, volumetric tension buffer gradient, and pure gastric electrical sequence uploaded by the sensing nodes, and performs synchronous integration according to the timestamps. Based on a preset multimodal feature weighted fusion model, the server calculates the full-cycle risk index by combining the event records of the high-frequency intervention cutoff logic at the rendering boundary, the volumetric tension buffer gradient, the pure gastric electrical sequence, and the comprehensive pain weight parameters. Based on the full-cycle risk index, a visualized vital sign analysis report containing the pure tension fluctuation curve, the pure gastric electrical sequence, and the event records of the high-frequency intervention cutoff logic at the rendering boundary is generated and pushed to the receiving terminal.

[0034] Specifically, in step S1, in-hospital baseline anchoring and edge-aware initialization are performed. The server receives the patient's comprehensive baseline data upon discharge through the medical data interface, which includes the comprehensive tolerance assessment baseline, electrogastrogram steady-state baseline, and clinical score.

[0035] The comprehensive tolerance assessment baseline is a fusion of biochemical indicators measured during the patient's in-hospital intensive care period, specifically including serum albumin and C-reactive protein (CRP) concentration characteristics. For the quantitative calculation of the comprehensive tolerance assessment baseline, the server extracts the measured concentrations of serum albumin and CRP from the patient, respectively. Using preset upper and lower limits for healthy populations, the measured serum albumin concentration is positively linearly normalized to obtain a positive nutrition index, and the measured CRP concentration, which characterizes the degree of negative inflammation, is negatively linearly normalized to obtain a negative inflammation index. The preset upper and lower limits for healthy populations are retrieved by the server from the latest version of the industry standard database of reference intervals for routine clinical biochemical tests published by the National Health Commission through a medical data interface, providing a benchmark scale for physiological indicators. Based on the physical causal relationship of clinical metabolic consumption, the server assigns a first weighting coefficient and a second weighting coefficient to the positive nutrition index and the negative inflammation index, respectively, and then outputs the comprehensive tolerance assessment baseline through weighted summation. As a preferred approach, the first weighting coefficient ranges from 0.6 to 0.7, and the second weighting coefficient ranges from 0.3 to 0.4. This range is determined based on the fact that the weight of malnutrition-related risk factors in patients recovering from severe pancreatitis is statistically significantly greater than that of residual acute inflammation. The steady-state baseline of gastric electrograms is the baseline electrophysiological frequency value of the stomach obtained by clinical instruments when the patient is fasting and resting in bed. The clinical score encompasses quantitative scores used to assess the functional status of multiple organ systems in patients.

[0036] The server sends the integrated baseline data to the sensing node via a wide area network communication link. The smart chip contains a radio frequency transceiver module and a memory. It receives the integrated baseline data through the radio frequency transceiver module and writes the integrated tolerance assessment baseline, electrogastrogram steady-state baseline, and clinical score into the memory in the form of structured data blocks, thus completing the anchoring and solidification of the integrated baseline data at the sensing node.

[0037] The system performs physical calibration of the sensing nodes on the patient's body surface. The patient is instructed to maintain three standard postures: supine, sitting, and upright, and to maintain calm breathing during these postures. The sensing nodes simultaneously and continuously acquire abdominal wall tension data and spatial posture matrix data in each posture until a preset posture maintenance time threshold is reached. This threshold is calculated based on the average duration required for healthy adults to complete at least five full respiratory cycles at rest, as measured by clinical medical statistics, to ensure that the acquired data covers the complete respiratory fluctuation characteristics. Due to the differences in sampling frequencies of the heterogeneous sensors, the intelligent chip uses its internal global clock as a unified time reference, timestamping the two heterogeneous data streams at the microsecond level. Using the signal with the higher sampling frequency as a reference, spline interpolation is performed on the signal with the lower sampling frequency to align the abdominal wall tension data and spatial posture matrix data on the same time profile.

[0038] The strain sensor can be a flexible piezoresistive sensor or a flexible capacitive sensor. The strain sensor is attached to the surface of the patient's abdominal wall skin, converting the mechanical deformation of the abdominal wall skin caused by breathing and torso bending into changes in resistance or capacitance. The intelligent chip contains a high-precision analog-to-digital converter, which converts the changes in resistance or capacitance into discrete digital signals, outputting the corresponding abdominal wall tension data.

[0039] The inertial measurement unit (IMU) includes a triaxial microelectromechanical system (MEMS) accelerometer and a triaxial MEMS gyroscope. The IMU synchronously outputs linear acceleration and angular velocity signals of the patient's torso in three-dimensional space, and after hardware integration and filtering, outputs spatial attitude matrix data containing pitch and roll angle data. For the calculation process of the spatial attitude matrix acquired by the IMU, those skilled in the art can use the quaternion method instead of the Euler angle method for attitude fusion calculation, avoiding the spatial matrix singularity produced by the Euler angle method when the patient's torso pitch angle approaches 90°. The quaternion-based attitude calculation algorithm is a well-known technology in the field and will not be elaborated upon here.

[0040] The intelligent chip generates and stores a postural respiratory baseline mapping table based on abdominal wall tension data and spatial posture matrix data. To eliminate physical interference from body posture and breathing movements on strain sensors in subsequent daily life, a mechanical noise benchmark is established. The intelligent chip performs digital signal processing on the abdominal wall tension data acquired during the calibration phase, separating the static stretching offset caused by specific trunk postures and the periodic respiratory fluctuations caused by diaphragmatic movement. For the method used by the intelligent chip to filter the abdominal wall tension data and separate the periodic respiratory fluctuations, those skilled in the art can use time-domain low-pass filtering algorithms or frequency-domain analysis algorithms based on fast Fourier transform to extract the corresponding frequency components.

[0041] The intelligent chip uses pitch and roll angles from the spatial attitude matrix data as the primary indexing dimensions and respiratory timing phases extracted from abdominal wall tension data as the secondary indexing dimension, constructing a multi-dimensional data grid in memory. The intelligent chip identifies the time interval between two adjacent peaks of periodic respiratory fluctuations in the abdominal wall tension data as a complete respiratory physiological cycle. To map the physical time domain to a unified polar coordinate angle domain and prevent overflow during hardware division operations due to the denominator approaching zero, the intelligent chip performs piecewise calculations based on the following respiratory timing phase calculation formula: ; in, This represents the discretized respiratory timing phase of the calculated output. It represents the time interval between two adjacent identified peaks, and its physical meaning is the duration of a complete respiratory physiological cycle. This represents the preset high-frequency interference threshold, used to determine whether the current cycle is an effective breathing cycle, and is set to 0.5 seconds based on the human physiological limit. A microsecond-level timestamp representing the current sampling point. The timestamp representing the start peak of the current complete respiratory physiological cycle. This represents a null value that is forcibly output due to the judgment that it is a pseudo-signal caused by an instantaneous mechanical collision in the external environment. It represents the full-cycle phase constant mapped to polar coordinates for a complete respiratory physiological cycle. The normalized dimensionless proportionality coefficient representing the current duration as a percentage of the total duration of the respiratory physiological cycle.

[0042] when Less than When this happens, the truncation logic is triggered directly, and the high-frequency mutation data segment is discarded; when Not less than At that time, the time offset of the current sampling point relative to the starting peak of the current complete respiratory physiological cycle and the proportion of the entire cycle are projected into a 360° circular space, completing the standardization and normalization of continuous timestamps to discrete respiratory timing phases. The intelligent chip fills the static stretch offset corresponding to the cross combination of different spatial posture matrix data and respiratory timing phases with the tension deformation superposition value of periodic respiratory fluctuations into the corresponding nodes of the multidimensional data grid, generating a body position respiratory baseline mapping table.

[0043] The intelligent chip stores the postural-respiratory baseline mapping table in its memory and loads it into its internal memory when performing out-of-hospital monitoring tasks. During independent operation out of the hospital, the intelligent chip directly inputs the real-time acquired spatial attitude matrix data and respiratory temporal phase into a multi-dimensional data grid.

[0044] When the input continuous floating-point value fails to match the discrete node coordinates of the multidimensional data grid, the intelligent chip locates the three-dimensional grid cell where the input coordinate point is located, retrieves the data of the eight vertices of the three-dimensional grid cell, and uses a trilinear interpolation algorithm to calculate the smoothed mechanical interference compensation value based on the relative position of the input coordinate point in each coordinate dimension. Then, in subsequent monitoring, based on the mechanical interference compensation value, mechanical noise stripping logic is performed on the real-time acquired noisy abdominal wall tension data and electrogastrogram signals to output a pure tension fluctuation sequence and a pure electrogastrogram sequence.

[0045] Specifically, the visual feature extraction and asymmetric sampling control steps are performed in S2, which include the following steps: S21. Steps for recognizing out-of-hospital eating scenarios and acquiring time-of-flight point cloud data. In out-of-hospital rehabilitation scenarios, the sensing nodes default to running in a low-power steady-state monitoring mode. When the patient prepares to eat and wears the interactive terminal, the interactive terminal activates the depth sensing component to continuously scan the field of view, outputting time-of-flight point cloud data containing three-dimensional spatial coordinates, and extracting geometric contour features and motion trajectory features based on this.

[0046] S22. Point cloud preprocessing, feature fusion classification, and comprehensive feeding confidence determination steps. The interactive terminal uses a three-dimensional convolutional neural network, which sequentially includes a voxel partitioning layer, a three-dimensional convolutional feature extraction layer, and a fully connected classification layer, to perform feature fusion and intent classification. In the data preprocessing stage, the interactive terminal downsamples the extracted time-of-flight point cloud data into a matrix containing a fixed number of points, which is set to 1024 or 2048. The tensor dimension of the time-of-flight point cloud matrix is ​​N×3, where N is the number of sampling points and 3 represents the Cartesian coordinate components of each sampling point. The voxel partitioning layer maps the Cartesian coordinate components to a three-dimensional voxel grid with a fixed resolution. The three-dimensional convolutional feature extraction layer uses a three-dimensional convolutional kernel to extract local spatial and temporal dynamic features from the voxel grid. The fully connected classification layer flattens the extracted three-dimensional spatial dynamic features and outputs the geometric contour matching degree and trajectory intent matching degree through an activation function. For the downsampling operation of the time-of-flight point cloud data, those skilled in the art can use a voxel grid filtering algorithm. Based on the geometric contour matching degree and trajectory intention matching degree output by the 3D convolutional neural network, the interactive terminal calculates the current comprehensive feeding confidence degree through weighted summation logic. Specifically, the interactive terminal assigns a first constant weight ranging from 0.3 to 0.4 to the geometric contour matching degree and a second constant weight ranging from 0.6 to 0.7 to the trajectory intention matching degree. Since the periodic movement of the hand approaching the face has a stronger physical causal determinant in determining feeding intention, the trajectory intention matching degree is given a higher weight. When the calculated comprehensive feeding confidence degree crosses a preset confidence trigger threshold, the interactive terminal determines that the patient has entered the feeding preparation stage. The confidence trigger threshold is set to 0.85 and is obtained from the optimal cutoff point of the receiver operation characteristic curve plotted by the server based on historical sample data.

[0047] The three-dimensional convolutional neural network is trained before leaving the management platform. It collects a large number of point cloud sequences of eating actions from healthy subjects and patients in the recovery period under different lighting conditions and different types of tableware as sample data. Medical experts manually label the actual occurrence intervals of eating actions to form hard labels. During the training process, the cross-entropy loss function is used to calculate the error between the predicted output of the three-dimensional convolutional neural network and the hard labels. The backpropagation algorithm combined with the adaptive moment estimation optimizer is used to iteratively update the weight matrix inside the three-dimensional convolutional neural network until the value of the cross-entropy loss function converges to a stable interval.

[0048] S23. Spatiotemporal Synchronization Pulse Generation and Distribution Steps. At the moment of entering the feeding preparation stage, the interactive terminal distributes a spatiotemporal synchronization pulse containing a compensation timestamp and a hardware interrupt command to the sensing node via the local area network communication link. Specifically, before distributing the command, the interactive terminal pre-sends a probe data packet to the sensing node and receives an acknowledgment receipt to calculate the round-trip time difference under the current network environment; the interactive terminal extracts the initial timestamp determining the moment of entering the feeding preparation stage, adds half of the round-trip time difference to the initial timestamp as unidirectional delay compensation, and generates the compensation timestamp.

[0049] S24. Heterogeneous Asymmetric Sampling Strategy Triggering and High-Frequency Sampling Switching Steps. The internal interrupt controller of the intelligent chip receives the spatiotemporal synchronization pulse and responds to the hardware interrupt command, immediately triggering the heterogeneous asymmetric sampling strategy. In low-power steady-state monitoring mode, the intelligent chip alternately reads data collected by the electromyography electrodes and strain sensors at a symmetrical, extremely low sampling frequency. After triggering the heterogeneous asymmetric sampling strategy, the intelligent chip actively suspends the conventional sampling channels for low-priority physiological features such as heart rate monitoring, and redirects the internal bus bandwidth and the computing resources of the analog-to-digital converter to the electromyography electrodes and strain sensors. The sampling frequency of the strain sensor acquiring abdominal wall tension data is increased to the 100 Hz level to fully capture the transient response of gastric smooth muscle mechanical expansion caused by swallowing, without causing hardware overload of the analog-to-digital converter.

[0050] Refer to the following steps S25-S210 Figure 3 As shown.

[0051] S25. Multi-source physiological data synchronous acquisition step. Under high-frequency sampling, the intelligent chip synchronously acquires the spatial attitude matrix data output by the inertial measurement unit, the abdominal wall tension data acquired by the strain sensor, and the electrogastrography signal acquired by the electromyography electrode.

[0052] S26. Respiratory Timing Phase Calculation Step. Given the mechanical deformation interference caused to the strain sensor by trunk movements during eating and deep breathing, the intelligent chip executes mechanical noise stripping logic to extract the pitch and roll angles from the spatial attitude matrix data, and calculates the current respiratory timing phase from the low-frequency baseline drift of the electrogastrogram signal.

[0053] S27. Mechanical Interference Compensation Value Acquisition Step. The intelligent chip uses the pitch angle, roll angle, and breathing timing phase as addressing input parameters, calls the pre-stored body position breathing baseline mapping table, and combines interpolation calculations to obtain the mechanical interference compensation value corresponding to the current physical state.

[0054] S28. Pure Tension Fluctuation Sequence Generation Steps. Based on the principle of linear additivity of multi-source signals in the same time dimension, the intelligent chip uses the real-time acquired abdominal wall tension data to subtract the corresponding mechanical interference compensation value, outputting a pure tension fluctuation sequence. Quantization is performed using the following mechanical noise stripping formula: ;in, Representative moment The pure tension value in the pure tension fluctuation sequence. Representative moment The abdominal wall tension data is obtained by collecting data from strain sensors and then converting it from analog to digital. This represents a mechanical disturbance compensation function mapping model based on specific spatial and temporal dimensions. Representative moment Pitch angle variable in spatial attitude matrix data. Representative moment The roll angle variable in the spatial attitude matrix data. Representative moment Discretized respiratory timing phase variables extracted and calculated from electrogastrogram signals. Input parameters for querying the three-dimensional joint index of the body position respiratory baseline mapping table, used for positioning time. The mechanical interference compensation value corresponding to the patient's working conditions. Through differential stripping calculation using the aforementioned mechanical noise stripping formula, the intelligent chip eliminates external mechanical stretching artifacts on the abdominal wall surface caused by the patient bending over to eat and breathing fluctuations, extracting the true tension mutation data.

[0055] S29. Electrophysiological baseline drift compensation and pure gastric electroencephalogram (GEG) sequence generation steps. The intelligent chip uses a preset electromechanical conversion mapping coefficient to convert the mechanical interference compensation value into an electrophysiological baseline drift, and subtracts the electrophysiological baseline drift from the real-time acquired GEG signal to output a pure GEG sequence.

[0056] S210, Pure physiological sequence caching step. The intelligent chip synchronously and sequentially pushes the pure tension fluctuation sequence and the pure gastric electrical sequence into its local high-speed cache queue.

[0057] In step S3, the objective intake measurement and virtual boundary rendering steps are performed. While the sensing node outputs the pure tension fluctuation sequence and the pure gastric electrical sequence, the interactive terminal simultaneously performs quantitative measurement and physical constraint on the patient's eating behavior at the visual level.

[0058] The interactive terminal calculates the initial total physical volume based on the food point cloud data acquired by the depth sensing component. Due to the extremely high spatial density of the original time-of-flight point cloud data, to avoid memory overflow in the interactive terminal's graphics processor, the terminal first uses a voxel grid downsampling algorithm to reduce the dimensionality of the food point cloud data. After dimensionality reduction, a statistical filtering algorithm is used to remove outliers caused by sensor thermal noise. For the data cleaning process based on voxel grid downsampling and statistical filtering, those skilled in the art can use standard point cloud library functions. Subsequently, based on the physical prior assumption that the food is placed on a table or plate in the eating scenario, the interactive terminal uses a random sampling consensus algorithm to perform planar segmentation on the dimensionality-reduced and denoised food point cloud data, removing point cloud sets belonging to the table background and tableware tray, and extracting the pure point cloud set of the target food. For the process of 3D planar extraction using the random sampling consensus algorithm, those skilled in the art can use an iterative hypothesis and verification mathematical model; the planar segmentation algorithm is a well-known technique in the field and will not be elaborated here. The interactive terminal uses a Poisson surface reconstruction algorithm to fit the pure point cloud set of the target food into a closed 3D mesh model. The logic for generating a closed 3D mesh model and calculating the volume of the closed mesh using the Poisson surface reconstruction algorithm can be implemented by those skilled in the art using the discrete divergence theorem combined with Gaussian integrals. The process of calculating the 3D mesh volume integration is a well-known technique in the field and will not be elaborated upon here. Through volume integration calculations, the interactive terminal calculates the initial total physical volume of the target food.

[0059] The interactive terminal obtains the maximum safe intake volume for a single dose from the server based on the comprehensive tolerance assessment baseline and clinical score from step S1. The calculation process uses the following baseline volume mapping formula: ; in, This represents the maximum safe intake volume calculated and output by the server in a single instance. The standard gastric emptying volume for a single healthy adult is a preset reference constant, which in this embodiment is set to 300 ml based on the statistical mean of the digestive tract anatomy of Chinese people. This represents the baseline of the comprehensive tolerance assessment measured and written to the server when the patient is discharged. This represents the baseline volume adjustment weight. Represents the baseline relaxation factor. Tolerance baseline product, This is a baseline relaxation constant. This represents the clinical score assessed when the patient is discharged. This represents the penalty coefficient for clinical scoring. Represented by the natural constant ( The penalty decay factor is obtained by using an exponential function with the base of the negative value of the clinical penalty product. As a preferred method, and The values ​​range from 0.1 to 0.15 and from 0.05 to 0.08, respectively. Since the risk of local tissue edema corresponding to the clinical score of pancreatitis lesions has a decisive physical determinant effect on the single maximum safe intake volume limit, a nonlinear, strong penalty constraint is imposed using a negative exponential mapping. However, the baseline for systemic comprehensive tolerance assessment only reflects the basal metabolic surplus, so a smaller linear weighting coefficient is used for a gentle relaxation. The server distributes the calculated single maximum safe intake volume to the interactive terminals via a wide area network.

[0060] The interactive terminal receives the maximum safe intake volume for a single operation and controls the rendering component to calculate a 3D virtual boundary model with a spatial volume equal to the maximum safe intake volume for a single operation, using the geometric center of the target food as the origin. Before calculating the inverse matrix of the pose tracking matrix output in real time by the positioning component, the rendering component extracts the third-order rotation submatrix from the pose tracking matrix and calculates its determinant value. When the absolute difference between this determinant value and 1 (i.e., the unit determinant reference value) is greater than a preset rotation distortion threshold, it is determined that the current pose tracking matrix has degraded. In this embodiment, the rotation distortion threshold is... The value is set based on the truncation error limit of the single-precision floating-point operation of the graphics processor inside the interactive terminal, so as to avoid the division-by-zero overflow crash caused by matrix invertibility without excessively triggering the rejection condition.

[0061] If the current pose tracking matrix is ​​determined to be degraded, the rendering component directly calls the inverse matrix of the previous valid visual sampling period to perform subsequent calculations. The rendering component uses the calculated inverse matrix to transform the 3D virtual boundary model from the objective world coordinate system to the local camera coordinate system of the interactive terminal. Then, it uses the graphics application programming interface to render the 3D virtual boundary model as a 3D virtual boundary model with semi-transparent materials. Finally, it displays the 3D virtual boundary model on the interactive terminal's screen by perspective alignment and overlay with the food in the real scene.

[0062] During the eating process, the interactive terminal continuously calculates the consumption of food point cloud and generates a discrete sequence of ingested volume slices. The depth sensing component continuously scans the food at a fixed visual sampling frequency (e.g., 30 Hz), reconstructing a closed 3D mesh model of the remaining food in real time and calculating the current remaining physical volume to match the Nyquist frequency of human hand movements during eating, while also considering the lower limit of the global shutter exposure time of the depth sensing component. To eliminate time-of-flight point cloud gaps and abrupt changes in volume calculation caused by hands or utensils obscuring the food during eating, the interactive terminal introduces median filtering logic in the temporal dimension. The interactive terminal performs filtering smoothing on the remaining physical volume for a consecutive preset number of filtering cycles (e.g., 5). The number of filtering cycles corresponds to approximately 0.16 seconds of physical time; this time window can cover the duration of most instantaneous occlusion by utensils in front of the viewer's line of sight without causing excessive smoothing distortion of the volume consumption curve.

[0063] The interactive terminal extracts the remaining physical volume of food from the previous visual sampling period after median filtering and smoothing, and subtracts it from the remaining physical volume of food in the current visual sampling period to obtain the initial difference in remaining physical volume. A monotonic truncation function is used to extract the maximum value between the initial difference in remaining physical volume and zero, which is then used as the intake volume slice value extracted in the current visual sampling period. As the visual sampling period increases, the interactive terminal pushes the intake volume slice values ​​obtained in each period into a one-dimensional array in chronological order, and uses the terminal's internal high-precision hardware clock to attach a microsecond-level timestamp to each intake volume slice value, thereby generating a discrete intake volume slice sequence containing time-series labels. The interactive terminal synchronously sends the intake volume slice sequence to the sensing node via a local area network communication link, serving as the objective physical input parameter for subsequent buffer analysis combined with the pure tension fluctuation sequence.

[0064] See Figure 4 In step S4, cross-modal voxel fusion and delayed buffer truncation are performed. During the patient's eating process, in addition to the objective calculation of the intake volume in step S3, the local discomfort caused by the patient's gastric distension is monitored and quantified simultaneously.

[0065] When a patient touches or presses their abdomen due to discomfort, the interactive terminal continuously tracks the gesture within its field of view using a positioning component. In this embodiment, the positioning component runs a simultaneous localization and mapping (SLAM) algorithm. For the process of using SLAM to construct a sparse point cloud of the surrounding environment in real time and calculate the six-degree-of-freedom pose of the interactive terminal in the objective world coordinate system, those skilled in the art can employ a visual odometry method based on feature points combined with a nonlinear optimization backend, which is a well-known technique in the field and will not be elaborated upon here.

[0066] In step S41, the interactive terminal obtains the three-dimensional coordinates of the hand feature points at the current moment from the positioning component. As a preferred method, the hand feature points are defined as the center point of the patient's palm or the tip of the index finger. The center point of the palm and the tip of the index finger are selected as specific input parameters because they have the most concentrated force characteristics and the highest pointing accuracy when humans perform instinctive avoidance actions such as pain relief.

[0067] In step S42, since the 3D coordinates directly output by the positioning component are located in the local camera coordinate system of the interactive terminal, the interactive terminal extracts the pose tracking matrix at the current moment. Before performing spatial coordinate transformation, the interactive terminal performs an orthogonality check on the third-order rotation submatrix in the pose tracking matrix and calculates the determinant value of the third-order rotation submatrix. When the absolute difference between the determinant value and 1 is greater than the preset orthogonality tolerance error (set to 1), the orthogonality tolerance error is determined. When the pose tracking matrix at the current moment is determined to be distorted or singular, the interactive terminal forcibly discards the current frame data and calls the pose tracking matrix from the previous valid moment for smooth replacement. The orthogonal tolerance error is extracted and calculated through multiple extreme viewpoint tracking tests combined with the hardware calibration manual of the graphics processor to ensure compatibility with the lower limit of basic accuracy.

[0068] In step S43, after confirming that the pose tracking matrix is ​​valid, the interactive terminal performs matrix multiplication on the three-dimensional coordinate vector of the hand feature points in the local camera coordinate system and the pose tracking matrix to calculate the absolute three-dimensional coordinates of the hand feature points in the objective world coordinate system.

[0069] In step S44, in order to map the spatial position of the gesture to the patient's abdominal surface, the interactive terminal uses a depth sensing component to acquire time-of-flight point cloud data of the patient's torso surface and constructs a two-dimensional planar mapping mesh that fits the contour of the patient's abdomen in the objective world coordinate system. The two-dimensional planar mapping mesh consists of multiple discrete mesh nodes arranged at a fixed physical interval (set to 1.0 cm).

[0070] In step S45, the interactive terminal calculates the vertical projection point of the absolute three-dimensional coordinates of the hand feature points in the objective world coordinate system onto the two-dimensional plane mapping grid, and uses the coordinates of the vertical projection point as the coordinates of the pain center.

[0071] The radial distribution of pain sensation from internal organs on the body surface exhibits spatial attenuation characteristics. The interactive terminal uses the pain center coordinates as a reference and employs a two-dimensional Gaussian kernel function to mathematically model the diffusion effect of the pain range, calculating its influence on surrounding discrete grid nodes and generating a pain space matrix. The calculation process uses the following pain weight distribution formula: ; in, The coordinates in the pain space matrix are The pain weight values ​​of discrete grid nodes. This represents an exponential function with the natural constant as its base. and Representing the first in the two-dimensional planar mapped mesh Line 1 The horizontal and vertical physical coordinates of the discrete grid nodes in the local coordinate system of the abdomen. and These represent the horizontal and vertical physical coordinates of the pain center, respectively. This represents the spatial variance parameter of pain radiation. The preset pain radiation attenuation constant is set, for example, to 3.0 to 5.0 cm. This represents the square of the Euclidean distance between the current discrete grid node and the coordinates of the pain center in the two-dimensional plane. This represents a decaying variable.

[0072] In the above calculations, the interactive terminal performs mandatory verification. The value of is greater than zero to avoid anomalies. The pain weight distribution formula transforms the single-dimensional point coordinates into a continuous weight distribution on a two-dimensional plane, so that the farther the discrete grid node is from the coordinates of the pain center, the closer the calculated pain weight value is to zero.

[0073] In step S46, the interactive terminal traverses all discrete grid nodes, filling the calculated pain weight values ​​into the corresponding grid node positions to generate a pain space matrix. To ensure data time alignment, the interactive terminal attaches a microsecond-level timestamp to the pain space matrix and sends it to the sensing node in real time via a local area network communication link. After receiving the pain space matrix, the sensing node extracts its attached microsecond-level timestamp and performs validity verification and delay compensation on the timestamp using a global clock reference pre-synchronized via a network time protocol. After verification, the smart chip stores the pain space matrix in a local high-speed cache queue.

[0074] See Figure 5 After performing gesture coordinate capture and pain space matrix mapping based on synchronous localization and mapping, the intelligent chip of the sensing node combines the ingested volume slice sequence to perform convolution calculation of viscoelastic delayed response decay and effective buffer volume sequence. In this embodiment, based on the biophysical characteristics of human gastric smooth muscle, the gastric wall tissue, as a typical viscoelastic medium, does not complete its mechanical expansion response to the increase in internal volume instantaneously. When food enters the stomach, the gastric wall smooth muscle undergoes stress relaxation and creep, resulting in a significant nonlinear hysteresis effect in the tension change of the abdominal wall surface. Therefore, the actual physical expansion load of the stomach is determined by the cumulative effect of all ingested volume slices over a historical time period after time-delay decay.

[0075] In step S51, the smart chip extracts the ingestion volume slice sequence with microsecond-level timestamps stored in the local cache queue. The ingestion volume slice sequence serves as the basic input data for subsequent viscoelastic delay response decay and effective buffer volume sequence convolution calculations.

[0076] In step S52, the smart chip extracts the clinical score obtained in step S1 and the viscoelastic relaxation time of the gastric smooth muscle. Dynamic calculations are performed. Specifically, based on the degree of severe local tissue edema reflected in the clinical score, the intelligent chip calculates the viscoelastic relaxation time of the gastric smooth muscle by multiplying a preset baseline relaxation time constant by the sum of the product of the clinical score and a preset relaxation penalty coefficient. As a preferred approach, the baseline relaxation time constant ranges from 100 to 150 seconds, specifically 120 seconds in this embodiment; the relaxation penalty coefficient ranges from 0.10 to 0.20, specifically 0.15 in this embodiment, to reflect the objective characteristic of slower elastic recovery of damaged gastric wall tissue. To avoid division-by-zero overflow anomalies during exponential decay operations, the intelligent chip applies a hardware lower limit clamp to the viscoelastic relaxation time of the gastric smooth muscle in its underlying hardware logic to ensure the output... It is greater than or equal to the reference relaxation time constant.

[0077] In step S53, the smart chip is based on the viscoelastic relaxation time. Given a fixed visual sampling period, the maximum delay length W of the time sliding window is calculated. Specifically, the intelligent chip divides three times the viscoelastic relaxation time of the gastric smooth muscle by the fixed visual sampling period and rounds down the quotient to obtain the maximum delay length. Before performing the division truncation calculation, the intelligent chip forcibly verifies whether the fixed visual sampling period is greater than a preset effective duration threshold, which is set as follows: To completely avoid the risk of division-to-zero overflow, the effective duration threshold matches the lower limit of hardware precision. The truncation calculation logic is based on the following: when the physical time difference spans three viscoelastic relaxation times, the weight coefficient of the exponential function has decayed to less than 5%, therefore, earlier historical data is forcibly truncated to reduce the number of multiply-accumulate operations of the intelligent chip and prevent stack overflow.

[0078] In step S54, the smart chip uses a one-dimensional discrete convolution operation to perform mathematical convolution calculations on the ingested volume slice sequence with time-series labels and a preset viscoelastic delay response kernel function, using the following effective buffer volume sequence convolution formula: ; in, The discrete time index representing the current visual sampling period is a positive integer that monotonically increases with objective physical time. Representing the The effective buffer volume sequence calculated from each visual sampling cycle represents the equivalent physical expansion load actually borne by the stomach at the current moment. This represents the maximum delay length of the time sliding window, which is the number of historical data points involved in the convolution calculation. This parameter represents the historical delay index within the sliding window, and its value ranges from... to . This represents a discrete summation operation on all decayed historical intake volume slice values ​​within a time sliding window. Representing the The ingress volume slice value extracted from each visual sampling cycle. Represented by the natural constant ( The time-varying weights are obtained by using an exponential function with the base of the time decay ratio (i.e., the negative of the time decay ratio). This represents the time interval between two adjacent ingested volume slices, i.e., the fixed visual sampling period duration. This represents the viscoelastic relaxation time of the gastric smooth muscle. Representing the The visual sampling period is the distance from the current visual sampling period. The absolute physical time difference experienced by each visual sampling cycle. Through the effective buffer volume sequence convolution formula, the intelligent chip transforms the discrete, objective intake volume slice sequence into a continuous effective buffer volume sequence that includes tissue physiological hysteresis effects.

[0079] In step S55, due to the asynchronous nature of the heterogeneous sensor data streams arriving at the local buffer queue and the difference in sampling frequency, the smart chip uses the microsecond-level timestamp of the internally maintained global clock to perform zero-order-preserved interpolation operations on the effective buffer volume sequence and the pain space matrix with the highest sampling frequency as a unified reference, thereby achieving point-to-point anchoring of cross-modal data on a unified time profile.

[0080] See Figure 6 Further, in step S4, high-frequency intervention truncation of volumetric tension buffer gradient calculation and rendering boundary is performed. After the intelligent chip at the sensing node completes point-to-point anchoring of cross-modal data such as pure tension fluctuation sequence, effective buffer volume sequence, and pain space matrix on a unified time profile, the intelligent chip continuously assesses changes in the physical expansion compliance of the patient's stomach.

[0081] In step S61, the smart chip extracts the pure tension value from the pure tension fluctuation sequence corresponding to the same timestamp in the local cache queue and the effective buffer volume value from the effective buffer volume sequence.

[0082] In step S62, the comprehensive pain weight parameter is calculated. To avoid misjudgment caused by the extraction of a single extreme value due to local muscle spasms of the patient or accidental noise from the sensor, this step traverses all discrete grid nodes in the pain space matrix, sorts the pain weight values ​​of all discrete grid nodes in descending order, extracts the pain weight values ​​of the top preset number of extreme value nodes (e.g., fixed at 5), calculates the arithmetic mean, and defines it as the comprehensive pain weight parameter at the current time. Comprehensive pain weighting parameter It represents the strongest risk-avoidance pain signal feedback exhibited by the patient's abdominal surface during the current eating phase.

[0083] In step S63, the volumetric tension buffer gradient is calculated. This step aims to accurately quantify the nonlinear stress rate of gastric smooth muscle under external food intake stimulation. The smart chip divides the difference in pure tension values ​​within a fixed time window by the difference in the effective buffer volume sequence, and combines this with a comprehensive pain weighting parameter to perform nonlinear sensitivity amplification. The following volumetric tension buffer gradient formula is used for quantification: ; in, Representing the The volumetric tension buffer gradient is calculated from a uniform sampling point. This represents the preset differential time window length, used to calculate the slope of macroscopic changes across multiple sampling points to smooth high-frequency interference. Representing the The pure tension value in the pure tension fluctuation sequence corresponding to each uniform sampling point. Representing the The pure tension value in the pure tension fluctuation sequence corresponding to a unified sampling point. It represents the physical increment of macroscopic tension. Representing the The effective buffer volume sequence corresponding to a unified sampling point. Representing the The effective buffer volume sequence corresponding to a unified sampling point. This represents the increase in the macroscopically equivalent intake volume. This represents the gain baseline constant, ensuring that the management platform retains the basic volumetric tension response slope as the primary criterion even when the overall pain weight parameter is zero. This represents the slope of the basic tension response. This represents the preset pain amplification gain coefficient, used to adjust the multiplicative amplification weight of the patient's subjective pain perception on the objective tension gradient. Representing the The comprehensive pain weight parameter is extracted from a unified sampling point. This represents the gain product.

[0084] In this embodiment, the differential time window length The time window is set to 15, corresponding to a physical time window of 0.5 seconds. This differential time window length is sufficient to span the high-frequency tension fluctuation cycle caused by respiration and heartbeat, thereby extracting the macroscopic volumetric tension response slope. Specifically, the differential time window length is determined by performing Fourier transform spectral analysis on the combined monitoring data of abdominal surface electromyography and tension from one hundred healthy subjects during standard eating and swallowing movements. The lower limit period of the dominant frequency interference band is extracted and then calculated using a 2x oversampling method. A preset volume increment threshold is also included. for This value matches the truncation error limit of microprocessor single-precision floating-point arithmetic during division to prevent overflow. Pain amplification gain coefficient. The value ranges from 1.5 to 2.0, and is specifically 1.8 in this embodiment. The pain amplification gain coefficient is an empirical multiplier determined by correlation analysis between subjective pain feedback and objective gastric wall electromyographic activity enhancement in multiple clinical patients. It quantifies the degree of nonlinear positive feedback amplification of neuropathic pain on gastrointestinal stress.

[0085] In this calculation step, the intelligent chip introduces a conditional branch truncation mechanism at the underlying hardware logic level. This mechanism works when the calculated difference in the effective buffer volume sequence is less than or equal to the volume increment threshold. When this occurs, it indicates that no effective intake has occurred, and the volumetric tension buffer gradient at the current unified sampling point is directly applied. The value is assigned to zero; conversely, only if the difference is greater than zero. Only when it is used as the denominator will the division instruction be executed.

[0086] In step S64, the intelligent chip compares the calculated volumetric tension buffer gradient with a preset tolerance gradient threshold in real time. The tolerance gradient threshold is a physical critical constant extracted by researchers based on historical vital sign data of acute pancreatitis patients during their in-hospital recovery period before gastric spasm and acute dilatation, using logistic regression classification fitting. In this embodiment, the tolerance gradient threshold is set to 45.0 Pa / mL. It is obtained by the server from a clinical vital sign dataset containing historical pure tension fluctuation sequences and historical effective buffer volume sequences of multiple acute pancreatitis patients during their in-hospital recovery period. The clinical vital sign dataset is labeled to mark abnormal time points where clinical gastric spasm and acute tissue dilatation occur. Subsequently, the server extracts historical volumetric tension buffer gradient features within a preset extraction window before the abnormal time points and inputs them into the logistic regression classifier for binary classification supervised training. During training, the gradient descent algorithm is used to iteratively optimize the weight parameters of the logistic regression classifier, using maximizing the area under the receiver operating characteristic curve as the objective function. The optimal decision boundary value for distinguishing between a safe compensatory state and an overload spasm state is calculated, and this optimal decision boundary value is set as the tolerance gradient threshold.

[0087] In step S65, to avoid false triggering caused by a single high-frequency noise, the volume tension buffer gradient calculated by visual sampling cycles for a number of consecutive preset confirmation cycles (set to 10, corresponding to a continuous confirmation time of about 0.33 seconds) is determined to be greater than the tolerance gradient threshold. This means that the current physical expansion load of the stomach has exceeded or reached the safe compensation limit of the patient's digestive system.

[0088] In step S66, the smart chip sends a hardware interrupt interception command with the highest execution priority to the interactive terminal through the local area network communication link.

[0089] In step S67, after receiving the hardware interrupt interception command, the interactive terminal immediately suspends the continuous scanning and calculation thread of the target food's three-dimensional point cloud data, controls the rendering component to forcibly overwrite the transparency parameter in the material shader of the three-dimensional virtual boundary model to zero, making it suddenly become a completely opaque visual occlusion surface, and directly switches the diffuse color channel of the three-dimensional virtual boundary model to a high-contrast red warning.

[0090] See Figure 7 In step S5, after performing the high-frequency intervention truncation steps of volume tension buffer gradient calculation and rendering boundary, the server further performs out-of-hospital full-cycle risk assessment and closed-loop feedback logic of medical information to achieve long-term dynamic monitoring of the recovery status of patients with acute pancreatitis after discharge.

[0091] S71. After a single feeding process, the interactive terminal timestamps the event record that triggers the high-frequency intervention cutoff logic of the rendering boundary, the volumetric tension buffer gradient within the corresponding time period, and the comprehensive pain weight parameter. This data, along with the pure tension fluctuation sequence and pure gastric electrical sequence uploaded by the sensing nodes, is then encrypted and uploaded to the server via a wide area network communication link. The server archives the received multimodal monitoring data into its internal relational database.

[0092] S72. Based on structured archived historical data, the server periodically performs a full-cycle risk assessment of the patient's outpatient gastrointestinal recovery status. The server uses a multimodal feature weighted fusion model to calculate the full-cycle risk index, quantifying the probability of pathological recurrence in patients under unstructured outpatient conditions. The calculation formula is as follows: ; in, The full-cycle risk index, calculated and output by the server, measures the overall probability that a patient's digestive system is on the verge of overload. This represents the preset truncation frequency penalty weight. This represents the total number of times the high-frequency intervention truncation logic for triggering the rendering boundary is within the set evaluation period. This represents the physical time span of the evaluation period. This represents the average forced blocking frequency characteristic per unit time after normalization. This represents the weights that influence the preset tension gradient. The constant is 1, which helps in calculating the arithmetic mean of the total sample. This represents the total number of samples of volumetric tension buffer gradients recorded during the evaluation period. This represents a discrete summation operation performed on all samples within the evaluation period. The discrete sample number representing the volumetric tension buffer gradient recorded within the evaluation period, with values ​​ranging from... continuously monotonically increasing to . Representing the The volumetric tension buffer gradient corresponding to each sample. This represents the average gastrointestinal stress sensitivity characteristics within the assessment period after normalization. This represents the preset pain feedback attenuation weight. This represents the average pain characteristic after normalization, which is the arithmetic mean of the comprehensive pain weight parameters recorded during the assessment period. This represents the preset penalty weight for gastric electrical deviation. This represents the normalized value of the average offset characteristic of the pure gastric electrograph sequence relative to the steady-state baseline of the electrograph during the evaluation period.

[0093] Before calculating the full-cycle risk index, to prevent division by zero anomalies, due to... The server determines the discrete integer variable representing the total number of samples as a constant. Is it greater than zero? If the total number of samples... A value of zero indicates that the system is not under effective monitoring, and the server will suspend the full-cycle risk index for the current period or output zero.

[0094] In this embodiment, Set to 7 days. Punishment weight for truncation frequency. Tension gradient affects weights Pain feedback attenuation weight Weight of penalty for gastric electrical offset The sum of these values ​​is forcibly constrained to 1. These weight parameters are all derived by the server through offline training and fitting in the cloud using a support vector regression model based on real follow-up recurrence data of historically discharged patients. The server constructs an offline training sample set, using the average forced blockage frequency feature, average gastrointestinal stress sensitivity feature, mean feature of comprehensive pain weight parameters, and average gastric electrical offset feature as input feature vectors; the corresponding supervision labels are extracted from the hospital's electronic medical record system and defined as the probability score of acute pancreatitis recurrence diagnosed within the corresponding time period after discharge. Subsequently, a radial basis function kernel is used to map the low-dimensional heterogeneous physical characteristics to a high-dimensional space to handle the nonlinear coupling relationship between multi-source data. During the model training phase, the server uses a pre-defined tolerance boundary insensitive loss function with relaxation variables as the optimization objective and iteratively updates the hyperplane parameters using a sequential minimum optimization algorithm.

[0095] After training convergence, the normalized hyperplane normal vector parameters are extracted and solidified as corresponding weights. In this embodiment, the values ​​of each weight parameter are as follows: It ranges from 0.3 to 0.4. The value ranges from 0.3 to 0.35. The value ranges from 0.1 to 0.15. It ranges from 0.1 to 0.2.

[0096] S73. To avoid false positives caused by occasional fluctuations in vital signs or sudden dietary abnormalities in a single period, the server extracts three consecutive assessment periods to calculate the full-cycle risk index, and uses an exponentially weighted moving average algorithm to calculate a smoothed comprehensive risk assessment value. In this embodiment, the preset time decay smoothing coefficient of the exponentially weighted moving average algorithm is 0.6, ensuring that recent data dominates while smoothing out transient changes.

[0097] S74. The server performs a full-cycle risk assessment and classification of the patient's current status based on the calculated full-cycle risk index or the smoothed comprehensive risk assessment value. Specifically, multi-level risk status assessment labels are preset to qualitatively describe the patient's full-cycle risk assessment level. When the comprehensive risk assessment value is less than or equal to the risk warning threshold, the server classifies the current full-cycle risk assessment as the compensatory period (under monitoring). This status classification indicates that the patient's gastrointestinal tract's comprehensive physiological burden under volume expansion and tension fluctuations is still within the compensatory limit of its own organ function, and no obvious pathological recurrence trend is shown. At this time, the management platform maintains a background dynamic monitoring state without triggering real-time warning pushes, realizing continuous quantitative tracking of the patient from health compensation to risk overload.

[0098] The server compares the calculated comprehensive risk assessment value with a preset risk warning threshold. The risk warning threshold is set at 0.75. If the comprehensive risk assessment value exceeds the threshold, the server classifies the current full-cycle risk assessment as an overload warning period and determines that the patient has an excessive risk of acute pancreatitis recurrence or severe loss of dietary control, immediately triggering the closed-loop feedback logic for doctor-patient information. Through an internal report generation engine, the server integrates the pure tension fluctuation curve during the abnormal time period, the electrophysiological baseline drift map of the pure gastric electrical sequence, and the event records of the high-frequency intervention cutoff logic at the rendering boundary to generate a multimodal visualized vital sign analysis report. Finally, the server uses a message queue telemetry transmission protocol to push this vital sign analysis report to the attending physician's reception terminal in real time.

[0099] S75. After reviewing the vital signs analysis report through the patient reception terminal, the attending physician can input medical intervention correction instructions for the patient on the terminal's interactive interface. These instructions include downward adjustment parameters for the comprehensive tolerance assessment baseline and clinical score, proactively tightening the system's tolerance for the patient's food intake volume. The server receives the medical intervention correction instructions from the patient reception terminal, overwrites the patient's baseline data in the relational database, and before the patient next opens the interactive terminal for eating, simultaneously sends the dynamically recalculated maximum safe intake volume per meal, combined with the downward adjustment parameters, and the viscoelastic relaxation time of the gastric smooth muscle to the interactive terminal and sensing nodes.

[0100] This design establishes a complete pathway at the underlying data flow level, from objective physiological feedback from patients to cloud-based risk assessment and then to manual intervention by medical experts. The management platform transforms unsupervised eating behaviors outside the hospital into quantifiable and monitorable clinical indicators, and establishes a baseline through a dynamic correction system, constructing a closed-loop risk management link for patients with acute and severe pancreatitis.

[0101] The following is a detailed description of a specific application example using a patient with severe acute pancreatitis who is in the outpatient recovery period. Figures 8 to 10 In the process, vertical dotted lines are uniformly used to mark the blocking points where the management platform determines that the boundary has been crossed and issues truncation commands. Meanwhile, if... Figure 8 As shown in the top status bar, the management platform of the intelligent monitoring terminal for the full-cycle outpatient rehabilitation of patients with severe pancreatitis is in a real-time data closed-loop activation state, and the operating mode is heterogeneous asymmetric high-frequency sampling (100Hz).

[0102] In-hospital baseline anchoring and multi-source sensor hardware initialization were performed. Before patient discharge, the server retrieved their biochemical indicators via the medical data interface. The measured serum albumin concentration was 35 g / L, and the measured C-reactive protein concentration was 12 mg / L. The server performed linear normalization using the reference interval for healthy individuals, and combined it with a first weighting coefficient of 0.65 and a second weighting coefficient of 0.35 to derive the comprehensive tolerance assessment baseline. The score was 0.72. (Patient's clinical score) The assessment score is 3. The sensing node is deployed on the patient's abdominal wall. The intelligent chip executes and generates a respiratory baseline mapping table, writing the baseline parameters into non-volatile flash memory. The corresponding static parameters are synchronized and displayed in real time. Figure 8 The patient recovery boundary parameter overview panel on the right side of the monitoring interface. Meanwhile, Figure 8 The terminal coordination and intervention command log panel on the right records the following logs: [10:05:12.00] Received the server's comprehensive tolerance assessment baseline and [10:12:45.50] Completed three-position calibration and generated a respiratory baseline mapping table. Figure 8 In the central sensing node's high-frequency data processing and intervention / interception monitoring area, medical staff can perform spectral observations by clicking different tabs.

[0103] Perform 3D reconstruction and virtual volume boundary generation of the eating feature target. The patient wears an interactive terminal to prepare for eating. The interactive terminal acquires point cloud data of the plate and food within the field of view, downsamples and fits it into a closed 3D mesh model, obtaining an initial total physical volume of 450 ml for the target food, triggering a log recording [10:20:01.12] visual acquisition of the initial total physical volume of the food 450.0 mL. The server calls the baseline volume mapping formula: Among them, the standard gastric emptying volume reference constant. 300 ml. Baseline volume adjustment weight. The clinical score penalty coefficient is 0.12. The value is 0.06. The maximum safe intake volume for a single incident is calculated. Approximately 272.2 ml was recorded, and the maximum safe intake volume for a single transaction was issued in the log [10:20:01.15]. The interactive terminal generates a three-dimensional virtual boundary model with the geometric center of the target food as the origin, and overlays and renders it onto the real food surface.

[0104] Asymmetric sampling is triggered and mechanical noise stripping is performed. The depth sensing component confirms that the overall feeding confidence level exceeds the confidence trigger threshold of 0.85. The sensing node increases the sampling frequency of the strain sensor to 100Hz, triggering the heterogeneous asymmetric sampling strategy (100Hz) in the log [10:20:02.00]. Figure 8 The interface status bar displays a synchronization precision of 10ms.

[0105] The patient's trunk bending and chewing respiration during eating caused significant abdominal wall deformation. The intelligent chip used a mechanical noise stripping formula to calculate a pure tension fluctuation sequence, generating [10:20:45.32] which successfully stripped away mechanical deformation noise and output a pure tension fluctuation sequence; in Figure 8 In the monitoring interface, click the "Original Mechanical Signal and Stripping Noise Pure Tension" button to observe the spectrum. Combined with... Figure 9 As can be seen, the gray solid line represents abdominal wall tension data, which exhibits extremely chaotic fluctuations due to interference from the patient's eating and breathing, as well as trunk displacement pulses reaching hundreds of Pascals. However, after applying the stripping logic, the pure tension values ​​(represented by the dark black solid line) in the output pure tension fluctuation sequence clearly weave through the noise, showing a smooth and monotonically increasing trend with eating, successfully restoring the true changes in gastric wall expansion stress. Furthermore, after... Figure 9 After the vertical dotted line (i.e. the blocking point), physical feeding is forcibly cut off by the management platform.

[0106] Ingestion volume extraction and viscoelastic buffer calculation. The depth sensing component continuously calculates the remaining food volume at a frequency of 30 Hz. After 5 cycles of median filtering smoothing and monotonic truncation, a discrete ingestion volume slice sequence containing microsecond-level timestamps is generated. Synchronize to the sensing node.

[0107] The intelligent chip simulates the stress relaxation characteristics of gastric wall smooth muscle to perform efficient buffered volume sequence convolutional transport: ;in, For the first The effective buffer volume sequence output by each visual sampling period. Represents the maximum delay length of the time sliding window (rounded down to three times the relaxation constant). Visual sampling period duration. It takes 0.033 seconds. Based on clinical scores Dynamic calculation is set to 120 seconds multiplied by That is, 174 seconds. Extract the effective buffer volume sequence. Click the "Food Volume Monitoring and Viscoelastic Buffer Conversion" buttons in the function area to view the corresponding charts for verification. Figure 8 As shown, the remaining physical volume of the target food decreases in a step-like manner; the equivalent magnified effective buffer volume sequence curve proves that the convolution algorithm can effectively smooth the historical eating speed and no longer accumulate after passing the blocking point, maintaining dynamic stability.

[0108] Real-time gradient assessment and hardware interruption interception. The patient experiences a pulling sensation in their abdomen and presses on it with their hand. The interactive terminal captures hand feature points, generates a pain space matrix, and extracts the comprehensive pain weight parameter. The intelligent chip invokes the volumetric tension buffer gradient formula: Among them, to adapt to high-frequency heterogeneous sampling, the differential time window length is... The high-frequency sampling period is set to 300 cycles (3 seconds). The gain reference constant is 1. Pain amplification gain coefficient. Set to 1.8. The tolerance gradient threshold is set to 45.0 Pa / mL.

[0109] Click the "Cross-modal Fusion Tension Buffer Gradient and Closed-Loop Truncation" button, and combine... Figure 10 It can be seen that in the early and middle stages of feeding, due to the smoothing difference mechanism, the volumetric tension buffer gradient curve remains within a low, gradual upward range; when feeding has progressed to approximately 395 seconds (corresponding to the log triggering an early warning [10:26:05.18]: the accumulated physical expansion load is approaching the organ's compensatory limit), the cumulative effect of the effective buffer volume sequence causes the physical expansion load to approach the organ's compensatory limit. At this time, combined with the multiplicative amplification gain from the patient pressing on the abdomen (the combined effect of pain weighting), the volumetric tension buffer gradient shows a significant sharp increase in the graph, instantly exceeding the tolerance gradient threshold of 45.0 Pa / mL (e.g., Figure 10 (As shown by the dashed line in the middle), the trigger log [10:26:36.42] indicates that the volumetric tension buffer gradient continues to exceed the threshold limit.

[0110] When the calculated value exceeds the tolerance gradient threshold, i.e., when it reaches... Figure 10 At the central blocking point, in the intervention determination, the status is displayed as: the blocking point has been triggered and is effective. The smart chip issues a hardware interrupt interception command, executes the intervention action and records the log [10:26:36.45]. The blocking point is triggered, and the high-frequency intervention truncation of the rendering boundary is executed. The interactive terminal immediately activates the interruption logic, changes the virtual boundary material to an impermeable gray-black occluding surface, forcibly cuts off the visual perception link, forces the patient's physical eating behavior to stop, and outputs [10:26:36.48]: the hardware interruption is effective, and the visual perception link is forcibly cut off. Figure 10 As shown, after crossing the blocking point, due to the truncation taking effect and the cessation of eating, the gradient calculation quickly drops and stabilizes at a low level (zero), effectively curbing the risk of stomach wall tearing caused by continued eating.

[0111] The system performs full-cycle risk index calculation and closed-loop feedback. After feeding ends, the interactive terminal and sensing nodes encrypt and upload timestamped multimodal data to the server. The server archives the data using a relational database and assesses the physical time span over 7 days. A multimodal feature weighted fusion model is used to calculate the full-cycle risk index. : Each weight is derived offline through support vector regression model training. If the smoothed comprehensive risk assessment value exceeds the 0.75 threshold, the server generates a multimodal visualization chart through the report generation engine and pushes it to the attending physician's terminal. The doctor inputs a value to adjust the parameters, and the server overwrites the database baseline, completing the closed-loop management. The log panel finally records: [10:26:40.12] The full-cycle risk index is uploaded to the server, and the closed loop ends.

[0112] Depend on Figure 9 It is known that abdominal wall tension data has an extremely low signal-to-noise ratio due to the superposition of respiratory rhythm and large trunk movements; the multidimensional data grid and mechanical noise stripping logic filter the chaotic signal into a monotonically increasing pure stress curve. Most importantly, Figure 9 It clearly demonstrates that when the blocking command is issued, the pure tension curve completely stops its dangerous rise, proving the physical effectiveness of the cutoff control.

[0113] Depend on Figure 8 It can be seen that the effective buffer volume sequence extracted by exponential decay convolution can effectively smooth the step-like historical eating behavior, proving that the management platform can adapt to the viscoelastic hysteresis property of the digestive system smooth muscle and overcome the shortsightedness of external static visual measurements.

[0114] Figure 10 The steep increase triggering mechanism of cross-modal fusion was reproduced. When feeding had been going on for approximately 395 seconds, the accumulation of the effective buffer volume sequence caused the local physical load to approach its limit. Combined with the multiplicative amplification gain from the pain pressure action synchronously captured by the interactive terminal, the originally gentle volumetric tension buffer gradient suddenly increased, exceeding the tolerance gradient threshold of 45.0 Pa / mL. The truncation logic triggered the interruption of feeding. Figure 10 The subsequent gradient curve quickly drops back to zero, preventing overfeeding.

[0115] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application.

Claims

1. A smart management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis, characterized in that, include: Sensing nodes are used to monitor the patient's abdominal physiological state and output tension fluctuation sequences; And calculate and generate a volumetric tension buffer gradient based on the intake volume slice sequence sent by the interactive terminal and the tension fluctuation sequence; An interactive terminal is used to acquire food point cloud data and determine the initial total physical volume, acquire the maximum safe intake volume for a single intake from the server and generate a three-dimensional virtual boundary model based on the maximum safe intake volume for a single intake, and calculate the food consumption based on the continuously acquired food point cloud data and the initial total physical volume during the patient's eating process, generate an intake volume slice sequence and send it to the sensing node. And by using the volumetric tension buffer gradient to change the rendering features of the three-dimensional virtual boundary model and outputting a visual truncation warning; The server is used to calculate the maximum safe intake volume per session based on the patient's comprehensive baseline data and send it to the interactive terminal; and to acquire the historical sequence of the volumetric tension buffer gradient and the record of the visual truncation warning, thereby completing the intelligent management of the patient's full-cycle outpatient rehabilitation.

2. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 1, characterized in that, The interactive terminal includes: The depth sensing component is used to collect time-of-flight point cloud data of target objects within the field of view for food identification and volume calculation. A positioning component is used to capture the patient's gesture coordinates and map them onto the abdominal surface, generate a pain space matrix, and output it to the sensing node; The rendering component is used to generate a three-dimensional virtual boundary model based on the single maximum safe intake volume and perform superimposed rendering. Based on the hardware interrupt interception command sent by the sensing node, it dynamically adjusts the rendering features of the three-dimensional virtual boundary model based on the volume tension buffer gradient and outputs a visual truncation warning.

3. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 2, characterized in that, The rendering component, when generating a 3D virtual boundary model and completing perspective overlay rendering, is specifically used for: Using the geometric center of the target food as the origin, a three-dimensional virtual boundary model of the corresponding spatial volume is generated based on the single maximum safe intake volume, and the determinant value of the third-order rotation sub-matrix contained in the pose tracking matrix output by the positioning component in real time is calculated. When the absolute difference between the determinant value of the third-order rotation submatrix and the preset unit determinant reference value is greater than the preset rotation distortion threshold, the inverse matrix of the pose tracking matrix of the previous effective visual sampling period is invoked. When the absolute difference is not greater than the rotational distortion threshold, the inverse matrix of the pose tracking matrix in the current visual sampling period is used. Using the inverse matrix that is invoked or adopted, the three-dimensional virtual boundary model is transformed from the objective world coordinate system to the local camera coordinate system, and the three-dimensional virtual boundary model is rendered so that the three-dimensional virtual boundary model is overlaid on the real food surface after perspective alignment.

4. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 3, characterized in that, The rendering component is also used to: dynamically adjust the rendering features of the three-dimensional virtual boundary model according to the comprehensive pain weight parameter corresponding to the pain space matrix output by the positioning component and the volume tension buffer gradient fed back by the perception node, and output a visual truncation warning signal when the volume tension buffer gradient exceeds a preset threshold.

5. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 4, characterized in that, The sensing nodes include: Electromyographic electrodes are used to acquire electrogastric signals. Inertial measurement unit, used to acquire spatial attitude matrix data under different standard body positions; Strain sensors are used to collect abdominal wall tension data under different standard body positions; The intelligent chip is used to obtain a postural respiratory baseline mapping table based on the electrogastrogram signal, the spatial posture matrix data, and the abdominal wall tension data.

6. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 5, characterized in that, The smart chip is specifically used for: Simultaneously acquire spatial posture matrix data and abdominal wall tension data under different standard body positions; The static stretch offset caused by trunk posture and the periodic respiratory fluctuation caused by diaphragmatic movement are separated from the abdominal wall tension data, and the respiratory timing phase is determined based on the periodic respiratory fluctuation. The pitch and roll angles in the spatial attitude matrix data are set as the primary index dimensions, the breathing timing phase is set as the secondary index dimension, and a multi-dimensional data grid is constructed. The static stretch offset and the periodic breathing fluctuation are added to obtain the tension deformation superposition value. The pitch angle, roll angle, and respiratory timing phase are cross-combined to form grid coordinates, and the corresponding tension deformation superposition values ​​are filled into the nodes of the multidimensional data grid to generate a body position respiratory baseline mapping table.

7. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 6, characterized in that, The intelligent chip is also used to perform mechanical noise stripping on the acquired electrogastric signal and abdominal wall tension data based on the body position and respiratory baseline mapping table, and output a pure tension fluctuation sequence and a pure electrogastric sequence.

8. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 5, characterized in that, The smart chip is also used to perform buffering operations on the discrete intake volume slice sequence to obtain an effective buffer volume sequence.

9. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 1, characterized in that, The server includes a secure access management module for: Retrieve the preset standard gastric emptying volume reference constant and obtain the patient's comprehensive baseline data; Extract the comprehensive tolerance assessment baseline from the comprehensive baseline data and multiply it by the baseline volume adjustment weight to obtain the tolerance baseline product. Add the tolerance baseline product to the preset benchmark relaxation constant to obtain the baseline relaxation factor. The clinical score is extracted from the comprehensive baseline data and multiplied by the clinical score penalty coefficient to obtain the clinical penalty product. The negative value of the clinical penalty product is used as the exponent of the natural constant to obtain the penalty decay factor. The maximum safe intake volume for a single dose is calculated by multiplying the standard gastric emptying volume reference constant, the baseline relaxation factor, and the penalty attenuation factor, and then the maximum safe intake volume for a single dose is sent to the interactive terminal.

10. The intelligent management platform for the full-cycle outpatient rehabilitation of patients with severe pancreatitis according to claim 2, characterized in that, The positioning component is also used to capture the three-dimensional coordinates of the gesture, generate a pain space matrix, and send it to the sensing node.

Images