A pressure ulcer risk real-time assessment system based on finite element analysis and deep learning

Abstract

The application discloses a kind of real-time assessment system of pressure sore risk based on finite element analysis and deep learning, it is related to medical monitoring and biomechanics simulation field;Solve the problem that existing assessment lacks real-time monitoring of deep tissue stress, dynamic early warning is insufficient.The application acquires body surface pressure, temperature and continuous pressure duration in real time by flexible pressure sensing blanket;Offline finite element simulation constructs parameterized human-mattress model, generates the paired data set of body surface pressure and deep tissue equivalent stress;Physical information deep learning model is based on pre-training real-time inference deep tissue equivalent stress distribution;Risk evolution calculation combines temperature and continuous pressure duration, and according to damage evolution law, calculates cumulative damage risk index and residual safety time;Visual early warning generates risk thermodynamic diagram and triggers hierarchical early warning and intervention instruction push.The application realizes real-time, accurate, individualized assessment and early warning of pressure sore risk.

Description

Technical Field

[0001] This invention belongs to the field of medical monitoring and biomechanical simulation, specifically a real-time pressure ulcer risk assessment system based on finite element analysis and deep learning. Background Technology

[0002] Pressure ulcers are a common complication in clinical nursing, and their occurrence is closely related to the continuous application of pressure to the body surface, leading to ischemia, hypoxia, and damage in deep tissues. Traditional pressure ulcer risk assessment mainly relies on the subjective judgment of nursing staff and tools such as the Braden Rating Scale, which makes it difficult to achieve real-time and quantitative monitoring of the stress state of deep tissues. In recent years, the development of flexible pressure sensing technology has made it possible to continuously collect the distribution of body surface pressure, but body surface pressure information alone cannot directly reflect the equivalent stress at the bone-muscle interface, which is considered one of the key mechanical factors in the development of pressure ulcers. Finite element simulation methods can simulate the distribution of deep tissue stress under different body positions, body types, and mattress conditions by constructing a human-mattress biomechanical model, but the simulation calculation is time-consuming and cannot meet the needs of real-time assessment. Deep learning technology has the ability to quickly map high-dimensional nonlinear relationships, providing a new approach for inferring deep tissue stress from body surface pressure distribution in real time. In addition, factors such as temperature and duration of continuous pressure (i.e., duration of tissue ischemia) have a significant impact on the rate of tissue damage evolution, requiring the establishment of a comprehensive risk evolution model. Therefore, constructing a pressure ulcer risk assessment system that integrates finite element simulation, deep learning, and multi-physical quantity sensing has important clinical significance.

[0003] The following problems exist in the existing technology:

[0004] Traditional pressure ulcer risk assessment relies on subjective judgment or static indicators, which cannot monitor the distribution of pressure on the patient's body surface and the stress state of deep tissues in real time.

[0005] While traditional finite element simulation can calculate the equivalent stress of deep tissues, the calculation is time-consuming and difficult to use for real-time dynamic clinical assessment.

[0006] Existing technologies only focus on surface pressure, ignoring the accelerating effects of temperature and duration of localized continuous pressure on the rate of tissue damage evolution, and lack a damage accumulation model that comprehensively considers multiple physical quantities.

[0007] Current technologies lack dynamic early warning and visual intervention guidance for pressure ulcer risk, and cannot provide the remaining safe time and targeted pressure relief measures in advance. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a real-time pressure ulcer risk assessment system based on finite element analysis and deep learning to solve the above-mentioned technical problem.

[0009] The first aspect of this invention provides a real-time pressure ulcer risk assessment system based on finite element analysis and deep learning, comprising the following modules:

[0010] The flexible pressure sensing blanket module includes a flexible pressure sensing blanket, which integrates a pressure sensing array and a temperature sensing node for real-time acquisition of pressure distribution data, temperature distribution data and duration of continuous pressure on each area of ​​the patient's body surface.

[0011] The offline finite element simulation module is used to construct a parametric human body-mattress finite element model. Through multi-condition simulation, it generates paired data of surface pressure distribution and deep tissue equivalent stress to form a pre-training dataset.

[0012] The physical information deep learning inference module is used to pre-train the built-in physical information deep learning model based on the pre-trained dataset; and to obtain the deep tissue equivalent stress distribution tensor based on the pressure distribution data and temperature distribution data through the pre-trained physical information deep learning model.

[0013] The risk evolution calculation module is used to calculate the cumulative damage risk index and remaining safe time at the corresponding positions of each sensing node of the flexible pressure sensing blanket based on the equivalent stress distribution tensor of the deep tissue, combined with the temperature distribution data and the duration of continuous pressure on each region, and based on the tissue damage risk evolution law.

[0014] The visualization and early warning module is used to generate and display a risk heat map based on the risk index of each sensing node of the flexible pressure sensing blanket in the cumulative damage risk index, and to trigger graded early warning and intervention instructions based on the cumulative damage risk index and its rate of change.

[0015] Preferably, the flexible pressure sensing blanket module includes:

[0016] The pressure acquisition unit, composed of the pressure sensor array, is used to acquire pressure values ​​at various spatial points on the patient's body surface in real time at a preset sampling frequency, forming pressure distribution data.

[0017] The temperature acquisition unit is composed of the temperature sensing nodes, which are distributed in the flexible pressure sensing blanket at a density lower than that of the pressure sensing array. It is used to collect the temperature values ​​of various areas of the patient's body surface in real time and form temperature distribution data.

[0018] The judgment unit is used to determine whether a pressure unloading event has occurred in the body surface area corresponding to each pressure sensor based on the real-time pressure value of each pressure sensor in the pressure sensor array, and to maintain the duration of continuous pressure on the area corresponding to each pressure sensor; wherein, a pressure unloading event is defined as the time when the real-time pressure value of the pressure sensor drops from above a preset pressure unloading judgment threshold to below a preset pressure unloading judgment threshold and remains below the preset pressure unloading judgment threshold for a period of time exceeding a preset time window; the duration of continuous pressure is defined as the time elapsed from the moment the most recent pressure unloading event occurred to the current moment.

[0019] Preferably, the offline finite element simulation module includes:

[0020] The model building unit is used to construct a parametric human-mattress finite element model. The human geometric model is constructed in layers according to the bone layer, muscle layer, fat layer, and skin layer. The material properties of each layer are defined using a preset biomechanical constitutive model. The physiological morphological parameters of the human geometric model are parametrically adjusted through a variety of adjustable individual parameters. The mattress geometric model is constructed according to preset geometric dimensions and material properties, and the type of the mattress geometric model can be parametrically selected through a variety of preset types.

[0021] The working condition configuration unit is used to generate a variety of simulation working conditions by combining and configuring body shape parameters, mattress geometry model type and body position parameters. The body position parameters include supine and lateral lying, and lateral lying includes a variety of preset tilt angles.

[0022] The calculation and pairing unit is used to perform finite element calculations for each simulation condition, extract the contact pressure distribution on the skin surface of the human geometric model as the body surface pressure distribution, and extract the maximum equivalent stress at the bone-muscle interface as the deep tissue equivalent stress. The body surface pressure distribution and the deep tissue equivalent stress are paired and stored as a sample record in the pre-training dataset.

[0023] Preferably, the physical information deep learning inference module includes:

[0024] The model building unit is used to build a physical information deep learning model using a hybrid architecture of U-Net and gated recurrent units; wherein, U-Net is used to extract spatial features from the pressure distribution data, and the gated recurrent unit is used to extract temporal features from the stored historical pressure distribution data;

[0025] The pre-training unit is used to pre-train the physical information deep learning model based on the body surface pressure distribution in the pre-training dataset as input and the deep tissue equivalent stress as label, so as to obtain the pre-trained physical information deep learning model.

[0026] Preferably, the physical information deep learning inference module further includes:

[0027] The physical constraint unit is used to add force balance constraint terms and constitutive relation constraint terms to the loss function during the pre-training process of the physical information deep learning model. The force balance constraint term is the difference between the integral value of the body surface pressure distribution obtained by converting the deep tissue equivalent stress distribution tensor output by the untrained physical information deep learning model through the preset mechanical constitutive model and the patient's weight. The constitutive relation constraint term is the deviation between the directional difference between the gradient of the deep tissue equivalent stress distribution tensor output by the untrained physical information deep learning model and the gradient of the input body surface pressure distribution and the value calculated by the preset mechanical constitutive model.

[0028] The real-time inference unit is used to output the deep tissue equivalent stress distribution tensor by fusing spatial and temporal features with the pressure distribution data and temperature distribution data as inputs, based on the pre-trained physical information deep learning model.

[0029] Preferably, the risk evolution calculation module includes:

[0030] The calculation unit is used to calculate the equivalent stress distribution tensor of the deep tissue, the temperature distribution data, and the duration of continuous pressure in each region. The cumulative damage risk index and risk evolution rate of each sensing node of the flexible pressure sensing blanket are calculated based on the tissue damage risk evolution law.

[0031] The evolutionary pattern of tissue damage risk is as follows:

[0032]

[0033] in, Where D is the risk evolution rate, and D is the cumulative damage risk index. For deep tissue equivalent stress, The tissue's stress tolerance threshold, The stress sensitivity index is T, where T is temperature. Baseline body temperature As a temperature acceleration factor, Based on the damage rate constant, This is the reference time constant.

[0034] Preferably, the risk evolution calculation module further includes:

[0035] An integration unit is used to numerically integrate the risk evolution rate and update the cumulative damage risk index.

[0036] Boundary processing unit, used to limit the updated cumulative damage risk index to the range of 0 to 1;

[0037] The safety time calculation unit is used to calculate the remaining safety time at the corresponding position of each sensing node of the flexible pressure sensing blanket based on the current cumulative damage risk index and the risk evolution rate, wherein the remaining safety time is the time required for the current cumulative damage risk index to rise to 1.

[0038] Preferably, the visual early warning module includes:

[0039] The visualization unit is used to map the risk index of each sensing node of the flexible pressure sensing blanket in the cumulative damage risk index to a color gradient, generate a risk heat map, and overlay the risk heat map onto the human body model for display.

[0040] The early warning unit is used to determine the early warning level of each sensor node based on the risk index and its rate of change of each sensor node corresponding to the location of the flexible pressure sensing blanket in the cumulative damage risk index.

[0041] The instruction unit is used to generate intervention instructions based on the spatial distribution of the warning level and the cumulative injury risk index. The intervention instructions include the turning direction, turning angle, and decompression measures.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This invention uses a flexible pressure sensing blanket to collect surface pressure, temperature and duration of continuous pressure in real time. Combined with a physical information deep learning model, it can infer the equivalent stress distribution tensor of deep tissues from the surface pressure distribution in real time, thus making up for the deficiency that deep tissue stress cannot be directly monitored in clinical practice.

[0044] This invention utilizes offline finite element simulation to generate paired datasets for pre-training deep learning models, achieving inference speeds that meet the needs of real-time clinical monitoring while ensuring mechanical accuracy (real-time calculations can be performed on edge computing devices based on the optimized model structure and sampling frequency), thus solving the problems of slow finite element simulation calculations and inability to be applied in real time.

[0045] This invention constructs a differential equation for the evolution of tissue damage risk, which includes stress, temperature, and duration of continuous pressure, and dynamically calculates the cumulative damage risk index and remaining safe time corresponding to each sensing node, thereby achieving quantitative risk assessment with multiple factors coupled.

[0046] This invention generates a risk heat map based on the risk index and its rate of change, and triggers graded early warning and intervention command push, which significantly improves the timeliness and accuracy of pressure ulcer prevention. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0048] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0049] Please see Figure 1 This invention is a real-time pressure ulcer risk assessment system based on finite element analysis and deep learning, comprising the following modules:

[0050] The flexible pressure sensing blanket module includes a flexible pressure sensing blanket, which integrates a pressure sensing array and a temperature sensing node for real-time acquisition of pressure distribution data, temperature distribution data and duration of continuous pressure on each area of ​​the patient's body surface.

[0051] The offline finite element simulation module is used to construct a parametric human body-mattress finite element model. Through multi-condition simulation, it generates paired data of surface pressure distribution and deep tissue equivalent stress to form a pre-training dataset.

[0052] The physical information deep learning inference module is used to pre-train the built-in physical information deep learning model based on the pre-trained dataset; and to obtain the deep tissue equivalent stress distribution tensor based on the pressure distribution data and temperature distribution data through the pre-trained physical information deep learning model.

[0053] The risk evolution calculation module is used to calculate the cumulative damage risk index and remaining safe time at the corresponding positions of each sensing node of the flexible pressure sensing blanket based on the equivalent stress distribution tensor of the deep tissue, combined with the temperature distribution data and the duration of continuous pressure on each region, and based on the tissue damage risk evolution law.

[0054] The visualization and early warning module is used to generate and display a risk heat map based on the risk index of each sensing node of the flexible pressure sensing blanket in the cumulative damage risk index, and to trigger graded early warning and intervention instructions based on the cumulative damage risk index and its rate of change.

[0055] Specifically, the flexible pressure sensing blanket module collects real-time pressure and temperature distribution data on the patient's body surface through a pressure sensing array and temperature sensing nodes integrated on the flexible pressure sensing blanket. Simultaneously, the module's internal judgment unit continuously monitors pressure changes in each sensing area, identifies the existence of "pressure unloading events" (i.e., whether the pressure drops below a safe threshold), and calculates and maintains the duration of continuous pressure on each area accordingly. The offline finite element simulation module is responsible for building the physical model. This module constructs a layered human-mattress finite element model including bones, muscles, fat, and skin. By configuring various body types, mattress types, and positions (such as supine, lateral, and different tilt angles) for simulation, the system performs finite element calculations. The calculation results are paired with the surface pressure distribution and the equivalent stress in the deeper layers (bone-muscle interface) to generate a pre-training dataset for training the AI ​​model. After acquiring the pre-training dataset, the physical information deep learning inference module pre-trains the built-in physical information deep learning model using the dataset. During actual operation, real-time collected pressure and temperature distribution data are used as input, and the pre-trained physical information deep learning inference module performs inference, outputting the deep tissue equivalent stress distribution tensor. The risk evolution calculation module combines the deep tissue equivalent stress distribution tensor, real-time temperature data, and the duration of continuous pressure on each region. Based on the preset tissue damage risk evolution law (including a coupled formula of stress, temperature, and time), it calculates the cumulative damage risk index corresponding to each node of the flexible pressure sensing blanket. It also further calculates the remaining safe time for that region, i.e., how long the current state must continue before tissue damage reaches a critical value. Finally, the visualization and early warning module maps the risk index of each sensing node to a color gradient, generates a risk heatmap, and overlays it onto the human body model. Simultaneously, based on the value of the risk index and its rate of change, the system triggers tiered early warnings and generates specific intervention instructions (such as turning direction, angle, or pressure relief measures), thereby achieving closed-loop management of pressure ulcer risk.

[0056] In one embodiment of the present invention, the flexible pressure sensing blanket module includes:

[0057] The pressure acquisition unit, composed of the pressure sensor array, is used to acquire pressure values ​​at various spatial points on the patient's body surface in real time at a preset sampling frequency, forming pressure distribution data.

[0058] The temperature acquisition unit is composed of the temperature sensing nodes, which are distributed in the flexible pressure sensing blanket at a density lower than that of the pressure sensing array. It is used to collect the temperature values ​​of various areas of the patient's body surface in real time and form temperature distribution data.

[0059] The judgment unit is used to determine whether a pressure unloading event has occurred in the body surface area corresponding to each pressure sensor based on the real-time pressure value of each pressure sensor in the pressure sensor array, and to maintain the duration of continuous pressure on the area corresponding to each pressure sensor; wherein, a pressure unloading event is defined as the time when the real-time pressure value of the pressure sensor drops from above a preset pressure unloading judgment threshold to below a preset pressure unloading judgment threshold and remains below the preset pressure unloading judgment threshold for a period of time exceeding a preset time window; the duration of continuous pressure is defined as the time elapsed from the moment the most recent pressure unloading event occurred to the current moment.

[0060] Specifically, the flexible pressure sensing blanket module includes a flexible pressure sensing blanket and data acquisition circuitry within the edge computing box. The flexible pressure sensing blanket uses a flexible polyimide substrate material, approximately 3-5 mm thick, and measures 80 cm x 80 cm. It can be laid on the bed surface to cover the patient's back, sacrum, coccyx, and heel areas. The flexible pressure sensing blanket integrates a pressure sensing array and temperature sensing nodes. The pressure sensing array consists of 4096 pressure sensors arranged in 64 rows by 64 columns, used to collect real-time pressure values ​​at various spatial points on the patient's body surface (e.g., the center point of the sacrum and coccyx). Each pressure sensor outputs a pressure value, with a sampling frequency of 5 Hz and units in kilopascals. The temperature sensing nodes consist of 256 surface-mount thermistors arranged in 16 rows by 16 columns, used to collect real-time temperature values ​​in various areas of the patient's body surface (e.g., the left hip area). Each surface-mount thermistor outputs a temperature value in degrees Celsius. The flexible pressure sensing blanket is led out to the edge computing box via flexible cabling.

[0061] The edge computing box internally houses a signal acquisition and processing circuit, which includes a capacitance-to-digital converter (CDC), an analog-to-digital converter (ADC), and a microcontroller. The CDC, connected to the pressure sensor array, converts the capacitance values ​​of each pressure sensor into pressure values; it uses an AD7147 chip. The ADC, connected to the temperature sensor node, converts the resistance values ​​of the thermistors into temperature values. The microcontroller, employing an ARM Cortex-M4 core, connects to both the CDC and ADC, controlling scan timing, performing data calibration, executing decision-making processes, and packaging data.

[0062] A judgment program runs in the microcontroller to determine whether a pressure unloading event has occurred in the body surface area corresponding to each pressure sensor based on the real-time pressure value of each pressure sensor, and to maintain the duration of continuous pressure on the area corresponding to each pressure sensor.

[0063] The pressure unloading threshold ranges from 0.2 kPa to 2.0 kPa, with a typical value of 0.5 kPa. The lower limit of this range is set by rounding up the pressure sensor noise level of the flexible pressure sensing blanket (approximately 0.05 kPa); the upper limit is set based on conventional understanding of surface pressure distribution in human biomechanics, ensuring that the threshold is significantly lower than the normal contact pressure in high-incidence areas of pressure ulcers such as the sacrum, coccyx, and heels.

[0064] A typical example of a pressure unloading event is to set the pressure unloading judgment threshold to 0.5 kPa and the preset time window to 5 seconds. If the patient's body surface pressure drops briefly for 2 seconds due to breathing and then recovers, it will not be recorded as a pressure unloading event. However, if the pressure sensor reading remains below 0.5 kPa for more than 5 seconds after the patient has finished turning over, it will be determined that the area has completed a pressure unloading event.

[0065] The duration of continuous pressure is defined as follows: for each pressure sensor, the duration of continuous pressure is counted from 0. When the real-time pressure value of the pressure sensor exceeds the pressure unloading threshold, the timer continues to accumulate; when a pressure unloading event occurs, the timer resets to zero and restarts. The duration of continuous pressure reflects the duration of tissue ischemia and is the cumulative effect of continuous pressure application over time. For example, if the sacrococcygeal region corresponding to a certain pressure sensor undergoes a turning over at 14:30:00 (a pressure unloading event occurs), the duration of continuous pressure on that region will be 930 seconds by 14:45:30.

[0066] In one embodiment of the present invention, the offline finite element simulation module includes:

[0067] The model building unit is used to construct a parametric human-mattress finite element model. The human geometric model is constructed in layers according to the bone layer, muscle layer, fat layer, and skin layer. The material properties of each layer are defined using a preset biomechanical constitutive model. The physiological morphological parameters of the human geometric model are parametrically adjusted through a variety of adjustable individual parameters. The mattress geometric model is constructed according to preset geometric dimensions and material properties, and the type of the mattress geometric model can be parametrically selected through a variety of preset types.

[0068] The working condition configuration unit is used to generate a variety of simulation working conditions by combining and configuring body shape parameters, mattress geometry model type and body position parameters. The body position parameters include supine and lateral lying, and lateral lying includes a variety of preset tilt angles.

[0069] The calculation and pairing unit is used to perform finite element calculations for each simulation condition, extract the contact pressure distribution on the skin surface of the human geometric model as the body surface pressure distribution, and extract the maximum equivalent stress at the bone-muscle interface as the deep tissue equivalent stress. The body surface pressure distribution and the deep tissue equivalent stress are paired and stored as a sample record in the pre-training dataset.

[0070] Specifically, the human geometric model is constructed in layers: bone, muscle, fat, and skin. The bone layer is taken from a standard human skeleton model, using a linear elastic material with an elastic modulus ranging from 5 to 15 GPa and a Poisson's ratio ranging from 0.25 to 0.35. The muscle layer has a thickness ranging from 20 to 50 mm, uses a hyperelastic material, and adopts the Mooney-Rivlin constitutive model. The material constants in this constitutive model... and The shear properties and volume response properties of the material are characterized respectively, based on common knowledge in the field of biomechanics. The value ranges from 5 kPa to 20 kPa. The value ranges from 2 kPa to 10 kPa. The fat layer thickness ranges from 10 to 40 mm, also using Mooney-Rivlin superelastic material. The value ranges from 2 kPa to 10 kPa. The values ​​range from 1 kPa to 5 kPa. The skin layer thickness is set at 2 mm, using a linear elastic material with an elastic modulus ranging from 100 kPa to 500 kPa and a Poisson's ratio ranging from 0.4 to 0.49. The physiological morphological parameters of the human geometric model include skeletal geometric dimensions and spatial position, muscle layer thickness and distribution, fat layer thickness and distribution, and skin layer thickness. These physiological morphological parameters are adjusted according to individualized parameters, including height, weight, BMI, fat layer thickness, and degree of muscle atrophy. Height ranges from 150 cm to 190 cm, weight ranges from 45 kg to 120 kg, BMI is divided into five levels: less than 18.5, 18.5 to 24, 24 to 28, 28 to 32, and greater than 32, fat layer thickness is divided into four levels: thin, normal, thick, and very thick, and muscle atrophy is divided into three levels: none, mild, and severe.

[0071] The mattress geometric model is constructed according to preset geometric dimensions and material properties. The mattress model thickness is uniformly set to 100 mm, with a planar dimension of 200 cm × 90 cm, covering the entire contact area of ​​the human body geometric model. The mattress geometric models include high-density foam mattresses, air mattresses, and gel mattresses. High-density foam mattresses use compressible foam material, with a superfoam constitutive model. The initial shear modulus ranges from 30 kPa to 80 kPa, and the material constant ranges from 1.5 to 3.0. Air mattresses use an equivalent elastic support model, simulated through spring-damping elements. The spring stiffness ranges from 5 to 20 kN / m, and the damping coefficient ranges from 0.05 to 0.2 kN·s / m. Gel mattresses use viscoelastic material, and stress relaxation behavior is characterized by a Planck series. The relaxation time ranges from 0.05 s to 0.5 s, and the ratio of long-term modulus to instantaneous modulus ranges from 0.3 to 0.7.

[0072] The body position parameters in the work condition configuration unit include supine and lateral positions. Lateral positions include various angles such as 15 degrees left lateral, 30 degrees left lateral, 15 degrees right lateral, and 30 degrees right lateral. After combining all body shape parameters, mattress geometry model types, and body position parameters, approximately 3000 simulation work conditions are generated. For example, a typical work condition is: BMI index 23, normal fat layer thickness, no muscle atrophy, high-density foam mattress type, and left lateral position at 30 degrees.

[0073] Finite element analysis was performed using the Abaqus / Standard solver with automatic time step control for quasi-static nonlinear analysis. For each simulation condition, a frictional contact relationship was established between the human body geometric model and the mattress geometric model, with a friction coefficient of 0.3 (ranging from 0.2 to 0.5); a gravity load was applied, and a fixed constraint was applied to the bottom of the mattress.

[0074] Before finite element analysis, the human body geometric model and the mattress geometric model are discretized into a finite number of mesh elements, each consisting of several nodes, each node having a degree of displacement freedom. The solver iteratively solves the nonlinear equilibrium equation system K(u)u=F, with all nodal displacement components as unknowns, where K(u) is the tangent stiffness matrix dependent on displacement u, and F is the nodal force vector corresponding to the gravity load. Due to the existence of material, geometric, and contact nonlinearities, this equation system is solved using the Newton-Raphson iterative method. In each iteration step, the norm of the unbalanced force is checked to see if it is less than 0.5% of the norm of the current external load vector, as a convergence criterion. After iterative convergence, the nodal displacement field is obtained, and then the stress field of each mesh element is obtained. The equivalent stress (i.e., von Mises stress, hereinafter referred to as equivalent stress) is calculated according to the von Mises yield criterion.

[0075] After solving the problem, the contact pressure distribution on the skin surface of the human geometric model is extracted as the body surface pressure distribution. The body surface pressure distribution is two-dimensional data. The contact pressure and its spatial coordinates of each node on the skin surface (e.g., the finite element mesh nodes on the sacrococcygeal skin surface) are extracted from the finite element results. Then, they are mapped onto a regular mesh of 64 rows by 64 columns using bilinear interpolation. The pressure value at each mesh point is in kilopascals (kPa). Simultaneously, the maximum equivalent stress at the bone-muscle interface in high-incidence areas of pressure ulcers, such as the sacrococcygeal region, heel, and scapula, is extracted as the deep tissue equivalent stress. The equivalent stress is calculated according to the von Mises yield criterion and is in kPa. For example, for a BMI of 23, a 30-degree left lateral decubitus position, and a high-density foam pad, the calculated deep tissue equivalent stress at the bone-muscle interface of the sacrococcygeal region is approximately 28.5 kPa.

[0076] The above calculations were performed on all 3,000 simulation conditions. Based on each simulation condition, the body shape parameters were extended by linear interpolation (e.g., generating intermediate values ​​by linear interpolation between BMI index levels) and Gaussian noise with a standard deviation of 1% of the pressure sensor range (40 kPa) was added to the body surface pressure distribution for data augmentation. Finally, no less than 100,000 paired samples were generated to form a complete pre-training dataset.

[0077] In one embodiment of the present invention, the physical information deep learning inference module includes:

[0078] The model building unit is used to build a physical information deep learning model using a hybrid architecture of U-Net and gated recurrent units; wherein, U-Net is used to extract spatial features from the pressure distribution data, and the gated recurrent unit is used to extract temporal features from the stored historical pressure distribution data;

[0079] The pre-training unit is used to pre-train the physical information deep learning model based on the body surface pressure distribution in the pre-training dataset as input and the deep tissue equivalent stress as label, so as to obtain the pre-trained physical information deep learning model.

[0080] Specifically, U-Net employs an encoder-decoder symmetric architecture. The encoder contains 5 levels of downsampling, each consisting of two 3×3 depthwise separable convolutional layers, a batch normalization layer, and a linear rectifier activation function. The feature tensor size is then halved using max pooling with a stride of 2. The number of channels in the encoder increases progressively from 64 to 512. The decoder contains 5 levels of upsampling, each consisting of a transposed convolutional layer that skips to the feature map of the corresponding encoder layer, followed by two separable convolutional layers, a batch normalization layer, and a linear rectifier activation function, progressively restoring the number of channels to 64. The spatial feature tensor output by U-Net has a size of 64 rows × 64 columns and 64 channels.

[0081] The historical pressure distribution data sequence is a series of pressure distribution data collected over the past 3 minutes at a sampling rate of 5 Hz, totaling 15 frames. Temporal feature extraction was performed using a gated recurrent unit (GRU). Each frame of historical pressure distribution data was mapped to a feature tensor through a lightweight convolutional network consisting of two 3×3 convolutional layers, resulting in 15 feature tensors, each with a dimension of 32×64×64. These feature tensors were then subjected to global average pooling in terms of spatial dimensions, yielding 15 feature vectors of dimension 32, forming a sequence of length 15 with each element having a dimension of 32. This sequence was then input into the GRU network, which consists of two layers with a hidden layer dimension of 256. The hidden state of the last time step output by the GRU network was used as the global temporal feature vector, with a dimension of 256.

[0082] The spatial feature tensor output by U-Net is fused with the global temporal feature vector output by GRU. The fusion method is as follows: the global temporal feature vector is mapped to a 64-dimensional bias vector through a fully connected layer, and then broadcast to each spatial location of the spatial feature tensor. The bias vector is added to the spatial feature tensor channel by channel to obtain the fused feature tensor. Finally, a 1×1 convolutional layer is used to map the number of channels to 1, outputting a 64-row × 64-column deep tissue equivalent stress distribution tensor, where each element represents the deep tissue equivalent stress at the location of the corresponding pressure sensor in the flexible pressure sensing blanket.

[0083] Each sample in the pre-training dataset contains a pressure distribution data point and its corresponding deep tissue equivalent stress label, which represents the maximum equivalent stress at the bone-muscle interface under that working condition. The pre-training dataset is divided into training, validation, and test sets in an 8:1:1 ratio. During training, mean squared error is used as the loss function, AdamW is selected as the optimizer, and the initial learning rate is typically 0.001, ranging from 0.0001 to 0.01. A cosine annealing decay strategy is employed. The number of training samples used for each model parameter update ranges from 16 to 64, with a maximum of 200 training epochs. Training is terminated early if the validation set loss does not decrease for 20 consecutive epochs. Online data augmentation is performed on the input data during training, including random horizontal flipping, random rotation by ±10 degrees, and adding Gaussian noise with a standard deviation of 0.02 to the normalized pressure distribution data (range 0 to 1). After pre-training, the model parameters with the lowest validation set loss are saved, resulting in the pre-trained physical information deep learning model.

[0084] In one embodiment of the present invention, the physical information deep learning inference module further includes:

[0085] The physical constraint unit is used to add force balance constraint terms and constitutive relation constraint terms to the loss function during the pre-training process of the physical information deep learning model. The force balance constraint term is the difference between the integral value of the body surface pressure distribution obtained by converting the deep tissue equivalent stress distribution tensor output by the untrained physical information deep learning model through the preset mechanical constitutive model and the patient's weight. The constitutive relation constraint term is the deviation between the directional difference between the gradient of the deep tissue equivalent stress distribution tensor output by the untrained physical information deep learning model and the gradient of the input body surface pressure distribution and the value calculated by the preset mechanical constitutive model.

[0086] The real-time inference unit is used to output the deep tissue equivalent stress distribution tensor by fusing spatial and temporal features with the pressure distribution data and temperature distribution data as inputs, based on the pre-trained physical information deep learning model.

[0087] Specifically, for the deep tissue equivalent stress distribution tensor output by the physical information deep learning model that was not fully trained during the pre-training process, it is converted into an estimate of the body surface pressure distribution using a mechanical transfer matrix pre-established by offline finite element simulation. The mechanical transfer matrix is ​​constructed as follows: using a large number of paired samples generated by the offline finite element simulation module, the deep tissue equivalent stress distribution under each simulation condition is discretized into a vector s, and the corresponding body surface pressure distribution is discretized into a vector p. A linear mapping p=Ks is fitted using the least squares method, where K is the mechanical transfer matrix. The discretized deep tissue equivalent stress distribution tensor output by the physical information deep learning model is multiplied by K to obtain the estimated body surface pressure distribution. This estimated value is then multiplied by the corresponding area micro-element (the actual physical area corresponding to each pressure sensor is 12.5mm×12.5mm) at each sensor node location and summed to obtain the model-predicted total support force; this total support force should be equal to the patient's weight (in Newtons). The force balance constraint is the difference between the total support force predicted by the model and the patient's weight; for example, a patient weighing 70 kg has a weight of 686 Newtons. If the model predicts a total support force of 650 Newtons, then the force balance constraint is 36 Newtons.

[0088] Constitutive constraint terms are used to ensure that the deep tissue equivalent stress distribution tensor output by the physical information deep learning model satisfies a preset mechanical constitutive relation with the input pressure distribution data. Specifically, according to elasticity theory, there is a definite physical relationship between deep tissue equivalent stress and surface pressure, manifested as the consistency of the directions of the stress gradient and pressure gradient. In this embodiment, the gradient fields of the deep tissue equivalent stress distribution tensor and the pressure distribution data in space are calculated separately. Then, the directional difference between the two gradient fields at each spatial location is calculated. The absolute values ​​of the directional differences at all spatial locations are summed and divided by the total number of spatial locations to obtain the constitutive constraint term. The directional difference is measured by the absolute value of the angle between the two gradient vectors. This constraint term can guide the model to learn a mapping relationship that conforms to physical laws. For example, at a certain spatial location, if the pressure gradient direction is horizontal to the right, and the gradient direction of the output deep tissue equivalent stress distribution tensor is horizontally deviated to the right by 10 degrees, then the directional difference is 10 degrees.

[0089] The force balance constraint term and constitutive relation constraint term are multiplied by their respective weight coefficients, and then added to the mean squared error (loss function) from the pre-training process to form the total loss function. The typical value of the weight coefficient for the force balance constraint term is 0.1, with a range of 0.05 to 0.5; the typical value of the weight coefficient for the constitutive relation constraint term is 0.05, with a range of 0.01 to 0.2.

[0090] The real-time inference unit deploys the pre-trained physical information deep learning model to the edge computing box. It receives the current pressure distribution data and historical pressure distribution data sequences from the past 3 minutes, along with temperature distribution data. The temperature distribution data is expanded to 64 rows by 64 columns using bilinear interpolation, and then concatenated with the pressure distribution data along the channel dimension as input to the physical information deep learning model. The pre-trained physical information deep learning model performs calculations following the process of U-Net extracting spatial features, GRU extracting temporal features, feature fusion, and finally convolutional output, resulting in a 64-row × 64-column deep tissue equivalent stress distribution tensor. The real-time inference unit repeats the above process at a frequency of 5 Hz.

[0091] In one embodiment of the present invention, the risk evolution calculation module includes:

[0092] The calculation unit is used to calculate the equivalent stress distribution tensor of the deep tissue, the temperature distribution data, and the duration of continuous pressure in each region. The cumulative damage risk index and risk evolution rate of each sensing node of the flexible pressure sensing blanket are calculated based on the tissue damage risk evolution law.

[0093] The evolutionary pattern of tissue damage risk is as follows:

[0094]

[0095] in, Where D is the risk evolution rate, and D is the cumulative damage risk index. For deep tissue equivalent stress, The tissue's stress tolerance threshold, The stress sensitivity index is T, where T is temperature. Baseline body temperature As a temperature acceleration factor, Based on the damage rate constant, This is the reference time constant.

[0096] Specifically, The cumulative damage risk index is the rate of change over time, i.e., the risk evolution rate, in units of... D is the cumulative damage risk index, ranging from 0 to 1, where 0 indicates no risk of damage and 1 indicates irreversible tissue damage has occurred. The equivalent stress of deep tissues is expressed in kilopascals. The stress threshold for tissue tolerance is expressed in kPa, with a typical value of 32 kPa and a range of 25 to 40 kPa. This threshold is determined by the critical stress value at which irreversible damage begins to occur in the tissue during offline finite element simulation. The critical stress value is set based on soft tissue mechanical experimental data. The stress sensitivity index is dimensionless, with a typical value of 2.5 and a range of 2.0 to 3.0. It reflects the superlinear relationship between the damage rate and stress, and is obtained by fitting soft tissue fatigue experimental data. Its value range is consistent with the power law exponent in fatigue damage theory. T is the body surface temperature, provided by the flexible pressure sensing blanket module, in degrees Celsius. The baseline body temperature is 37 degrees Celsius, with a range of 36.5 to 37.5 degrees Celsius. Temperature acceleration factor, unit: The typical value is 0.07, and the range is usually less than 0.1. It is calibrated by thermo-mechanical coupling finite element simulation, and its value is based on the range of the influence of temperature on the reaction rate of biological tissue in the Arrhenius equation. The basic damage rate constant, in units of Typical value is 5× The value range is 2× Up to 1× The value is obtained by fitting the inverse solution of the time required for the cumulative damage risk index to reach 1 in the finite element simulation, ensuring that the damage rate is within a reasonable range under typical working conditions. Its specific value can be further optimized based on clinical data. This is a reference time constant, measured in seconds, with a typical value of 1800 seconds and a range of 1200 to 2400 seconds. It is based on the "Prevention and Treatment of Pressure Injuries: Clinical Practice Guidelines (2019 Edition)" jointly published by the European Pressure Ulcer Advisory Committee (EPUAP), the National Pressure Ulcer Advisory Committee (NPIAP) of the United States, and the Pan-Pacific Pressure Injury Alliance (PPPIA). Continuous pressure exceeding 30 minutes is a significant time threshold leading to tissue ischemia and injury, and is commonly used as an important reference point for turning interventions in clinical practice. Setting it to 1800 seconds will allow the time item to... The contribution value was 1 after 30 minutes, which is consistent with clinical understanding.

[0097] In one embodiment of the present invention, the risk evolution calculation module further includes:

[0098] An integration unit is used to numerically integrate the risk evolution rate and update the cumulative damage risk index.

[0099] Boundary processing unit, used to limit the updated cumulative damage risk index to the range of 0 to 1;

[0100] The safety time calculation unit is used to calculate the remaining safety time at the corresponding position of each sensing node of the flexible pressure sensing blanket based on the current cumulative damage risk index and the risk evolution rate, wherein the remaining safety time is the time required for the current cumulative damage risk index to rise to 1.

[0101] Specifically, the integration unit uses the forward Euler integration method to numerically integrate the risk evolution rate and update the cumulative damage risk index, i.e. ,in, This represents the cumulative damage risk index from the previous moment. The time step is set to 0.2 seconds (corresponding to a 5 Hz sampling frequency for the pressure distribution data of the flexible pressure sensing blanket module). The integration unit repeats the above update operation at a frequency of 5 Hz. For example, if the cumulative damage risk index of a certain sensing node (the location of each pressure sensor in the flexible pressure sensing blanket) was 0.25 at the previous moment, the risk evolution rate at the current moment would be 0.002. Then, after 0.2 seconds, it will be updated to 0.2501.

[0102] The boundary processing unit limits the updated cumulative damage risk index to a range of 0 to 1. Specifically, if the updated cumulative damage risk index of a certain sensor node location is less than 0, the boundary processing unit sets it to 0; if it is greater than 1, it sets it to 1. For example, if the cumulative damage risk index calculated for a certain location is -0.003, the boundary processing unit corrects it to 0; if it is 1.02, it corrects it to 1.

[0103] The remaining safe time in the safe time calculation unit is defined as the time required for the current cumulative damage risk index to rise to 1. Since the risk evolution rate changes dynamically over time, two calculation methods are provided: First, a linear approximation based on the current risk evolution rate is used, with the following formula: ,in The cumulative damage risk index output by the boundary processing unit (limited to 0-1). It is a very small positive number (typical value is ). The first step is to avoid division by zero. This linear approximation assumes that the risk evolution rate remains constant over a short period, providing a conservative reference for intervention time in clinical practice. The second step uses numerical integration to solve the risk evolution differential equation in reverse, obtaining a more accurate remaining safe time. Specifically, given the current cumulative damage risk index and duration of continuous pressure, the remaining safe time is obtained by gradually increasing the time step and recalculating the risk evolution rate until the cumulative damage risk index reaches 1. If the current risk evolution rate is zero, the remaining safe time is set to a preset maximum value, typically 7200 seconds (2 hours). In practical applications, either linear approximation or numerical integration can be selected based on clinical needs. For example, the cumulative damage risk index after boundary processing of a certain sensor node is 0.7, and the risk evolution rate is 0.0001. If the risk evolution rate is zero, then the remaining safe time is 3000 seconds (50 minutes); if the risk evolution rate is zero, then the remaining safe time is directly taken as 7200 seconds.

[0104] In one embodiment of the present invention, the visual warning module includes:

[0105] The visualization unit is used to map the risk index of each sensing node of the flexible pressure sensing blanket in the cumulative damage risk index to a color gradient, generate a risk heat map, and overlay the risk heat map onto the human body model for display.

[0106] The early warning unit is used to determine the early warning level of each sensor node based on the risk index and its rate of change of each sensor node corresponding to the location of the flexible pressure sensing blanket in the cumulative damage risk index.

[0107] The instruction unit is used to generate intervention instructions based on the spatial distribution of the warning level and the cumulative injury risk index. The intervention instructions include the turning direction, turning angle, and decompression measures.

[0108] Specifically, the visualization unit maps the risk index of each sensor node in the cumulative injury risk index matrix to a color gradient, generating a risk heatmap, which is then overlaid on the human body model for display. The color mapping rules are: a cumulative injury risk index of 0 corresponds to blue, 0.3 to yellow-green, 0.6 to yellow, 0.8 to orange, and 1 to red. For risk indices between these thresholds, linear interpolation is used to calculate the RGB color components. For example, a cumulative injury risk index of 0.5 falls between 0.3 (yellow-green) and 0.6 (yellow), and after interpolation, it appears as a yellow-green with a yellowish tint. The display refreshes at a frequency of 1 Hz, showing the generated heatmap in real time on a terminal (such as a large screen at the nurse station). The heatmap is overlaid on a standard human back model, with high-risk areas for pressure ulcers, such as the sacrum, heel, and scapula, aligned with the sensor blanket layout. Users can click on any location on the heatmap to view the specific risk index value and remaining safe time for that point.

[0109] The early warning unit first calculates the rate of change of the risk index at each sensor node location, which is the difference between the current cumulative damage risk index and the cumulative damage risk index 15 minutes ago, divided by 900 seconds. Then, a three-level early warning mechanism is employed, with each threshold dynamically adjusted based on clinical data. The following are exemplary thresholds: Level 1 (Attention) warning trigger condition: the cumulative damage risk index at any sensor node location (e.g., the sensor node location corresponding to the center of the sacrococcygeal region) is greater than 0.5 and the rate of change is greater than 0.0001 (i.e., an increase of more than 0.09 within 15 minutes); Level 2 (High Risk) warning trigger condition: the cumulative damage risk index at any sensor node location is greater than 0.7, or the cumulative damage risk index is predicted to exceed 0.9 within the next 20 minutes (predicted value = current index + rate of change × 1200 seconds); Level 3 (Emergency) warning trigger condition: the cumulative damage risk index at any sensor node location is greater than 0.85, or the cumulative damage risk index is predicted to exceed 1.0 within the next 10 minutes. All locations are continuously monitored, and when any location meets the trigger condition, the highest warning level for that location is determined, and the corresponding alarm is triggered. For example, a patient's current cumulative injury risk index at the center of the sacrococcygeal region is 0.75, compared to 0.65 15 minutes ago, a change rate of approximately 0.00011. This meets the criteria for a level-two warning, triggering a level-two alarm. To prevent frequent alarms due to momentary fluctuations, the trigger condition must persist for at least 10 seconds before a formal alarm signal is output.

[0110] The intervention instructions in the instruction unit include the direction of turning, the angle of turning, and decompression measures. The principle for determining the direction of turning is: compare the average cumulative injury risk index of the left side (column indices 1 to 32) and the right side (column indices 33 to 64). If the average cumulative injury risk index of the left side is higher than that of the right side, turning to the right is recommended; otherwise, turning to the left is recommended. The selection of the turning angle is based on clinical nursing routines. The following are exemplary recommendations: 15 degrees for Level 1 warning, 30 degrees for Level 2 warning, and 30 degrees with the addition of a decompression pad for Level 3 warning; in actual application, adjustments can be made according to the patient's body type, tolerance, and clinical experience. Decompression measures are generated based on the area of ​​the highest risk index: if the highest risk area is located in the sacrococcygeal region, "use a sacrococcygeal decompression pad" is recommended; if it is located in the heel area, "use a heel suspension pad" is recommended. For example, if a patient's cumulative risk index for sacrococcygeal injury is 0.82 (Level 3 warning), the average risk index for the left side of the body is 0.6, and the average risk index for the right side of the body is 0.4, the instruction unit will generate the instruction: "Level 3 warning, immediately turn to the right 30 degrees and use a sacrococcygeal decompression pad." This instruction is displayed in text form on the terminal, and a voice broadcast is triggered simultaneously when a Level 3 warning is issued. After executing the instruction, the user can confirm "executed" on the terminal, and the system will automatically record the intervention.

[0111] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A real-time pressure ulcer risk assessment system based on finite element analysis and deep learning, characterized in that, Includes the following modules: The flexible pressure sensing blanket module includes a flexible pressure sensing blanket, which integrates a pressure sensing array and a temperature sensing node for real-time acquisition of pressure distribution data, temperature distribution data and duration of continuous pressure on each area of ​​the patient's body surface. The offline finite element simulation module is used to construct a parametric human body-mattress finite element model. Through multi-condition simulation, it generates paired data of surface pressure distribution and deep tissue equivalent stress to form a pre-training dataset. The physical information deep learning inference module is used to pre-train the built-in physical information deep learning model based on the pre-trained dataset; and to obtain the deep tissue equivalent stress distribution tensor based on the pressure distribution data and temperature distribution data through the pre-trained physical information deep learning model. The risk evolution calculation module is used to calculate the cumulative damage risk index and remaining safe time at the corresponding positions of each sensing node of the flexible pressure sensing blanket based on the equivalent stress distribution tensor of the deep tissue, combined with the temperature distribution data and the duration of continuous pressure on each region, and based on the tissue damage risk evolution law. The visualization and early warning module is used to generate and display a risk heat map based on the risk index of each sensing node of the flexible pressure sensing blanket in the cumulative damage risk index, and to trigger graded early warning and intervention instructions based on the cumulative damage risk index and its rate of change.

2. The pressure ulcer risk real-time assessment system based on finite element analysis and deep learning according to claim 1, characterized in that, The flexible pressure sensing blanket module includes: The pressure acquisition unit, composed of the pressure sensor array, is used to acquire pressure values ​​at various spatial points on the patient's body surface in real time at a preset sampling frequency, forming pressure distribution data. The temperature acquisition unit is composed of the temperature sensing nodes, which are distributed in the flexible pressure sensing blanket at a density lower than that of the pressure sensing array. It is used to collect the temperature values ​​of various areas of the patient's body surface in real time and form temperature distribution data. The judgment unit is used to determine whether a pressure unloading event has occurred in the body surface area corresponding to each pressure sensor based on the real-time pressure value of each pressure sensor in the pressure sensor array, and to maintain the duration of continuous pressure on the area corresponding to each pressure sensor; wherein, a pressure unloading event is defined as the time when the real-time pressure value of the pressure sensor drops from above a preset pressure unloading judgment threshold to below a preset pressure unloading judgment threshold and remains below the preset pressure unloading judgment threshold for a period of time exceeding a preset time window; the duration of continuous pressure is defined as the time elapsed from the moment the most recent pressure unloading event occurred to the current moment.

3. The pressure ulcer risk real-time assessment system based on finite element analysis and deep learning according to claim 1, characterized in that, The offline finite element simulation module includes: The model building unit is used to construct a parametric human-mattress finite element model. The human geometric model is constructed in layers according to the bone layer, muscle layer, fat layer, and skin layer. The material properties of each layer are defined using a preset biomechanical constitutive model. The physiological morphological parameters of the human geometric model are parametrically adjusted through a variety of adjustable individual parameters. The mattress geometric model is constructed according to preset geometric dimensions and material properties, and the type of the mattress geometric model can be parametrically selected through a variety of preset types. The working condition configuration unit is used to generate a variety of simulation working conditions by combining and configuring body shape parameters, mattress geometry model type and body position parameters. The body position parameters include supine and lateral lying, and lateral lying includes a variety of preset tilt angles. The calculation and pairing unit is used to perform finite element calculations for each simulation condition, extract the contact pressure distribution on the skin surface of the human geometric model as the body surface pressure distribution, and extract the maximum equivalent stress at the bone-muscle interface as the deep tissue equivalent stress. The body surface pressure distribution and the deep tissue equivalent stress are paired and stored as a sample record in the pre-training dataset.

4. The pressure ulcer risk real-time assessment system based on finite element analysis and deep learning according to claim 1, characterized in that, The physical information deep learning inference module includes: The model building unit is used to build a physical information deep learning model using a hybrid architecture of U-Net and gated recurrent units; wherein, U-Net is used to extract spatial features from the pressure distribution data, and the gated recurrent unit is used to extract temporal features from the stored historical pressure distribution data; The pre-training unit is used to pre-train the physical information deep learning model based on the body surface pressure distribution in the pre-training dataset as input and the deep tissue equivalent stress as label, so as to obtain the pre-trained physical information deep learning model.

5. The pressure ulcer risk real-time assessment system based on finite element analysis and deep learning according to claim 4, characterized in that, The physical information deep learning inference module also includes: The physical constraint unit is used to add force balance constraint terms and constitutive relation constraint terms to the loss function during the pre-training process of the physical information deep learning model. The force balance constraint term is the difference between the integral value of the body surface pressure distribution obtained by converting the deep tissue equivalent stress distribution tensor output by the untrained physical information deep learning model through the preset mechanical constitutive model and the patient's weight. The constitutive relation constraint term is the deviation between the directional difference between the gradient of the deep tissue equivalent stress distribution tensor output by the untrained physical information deep learning model and the gradient of the input body surface pressure distribution and the value calculated by the preset mechanical constitutive model. The real-time inference unit is used to output the deep tissue equivalent stress distribution tensor by fusing spatial and temporal features with the pressure distribution data and temperature distribution data as inputs, based on the pre-trained physical information deep learning model.

6. The pressure ulcer risk real-time assessment system based on finite element analysis and deep learning according to claim 1, characterized in that, The risk evolution calculation module includes: The calculation unit is used to calculate the equivalent stress distribution tensor of the deep tissue, the temperature distribution data, and the duration of continuous pressure in each region. The cumulative damage risk index and risk evolution rate of each sensing node of the flexible pressure sensing blanket are calculated based on the tissue damage risk evolution law. The evolutionary pattern of tissue damage risk is as follows: in, Where D is the risk evolution rate, and D is the cumulative damage risk index. For deep tissue equivalent stress, The tissue's stress tolerance threshold, The stress sensitivity index is T, where T is temperature. Baseline body temperature As a temperature acceleration factor, Based on the damage rate constant, This is the reference time constant.

7. The pressure ulcer risk real-time assessment system based on finite element analysis and deep learning according to claim 6, characterized in that, The risk evolution calculation module also includes: An integration unit is used to numerically integrate the risk evolution rate and update the cumulative damage risk index. Boundary processing unit, used to limit the updated cumulative damage risk index to the range of 0 to 1; The safety time calculation unit is used to calculate the remaining safety time at the corresponding position of each sensing node of the flexible pressure sensing blanket based on the current cumulative damage risk index and the risk evolution rate, wherein the remaining safety time is the time required for the current cumulative damage risk index to rise to 1.

8. The pressure ulcer risk real-time assessment system based on finite element analysis and deep learning according to claim 1, characterized in that, The visual early warning module includes: The visualization unit is used to map the risk index of each sensing node of the flexible pressure sensing blanket in the cumulative damage risk index to a color gradient, generate a risk heat map, and overlay the risk heat map onto the human body model for display. The early warning unit is used to determine the early warning level of each sensor node based on the risk index and its rate of change of each sensor node corresponding to the location of the flexible pressure sensing blanket in the cumulative damage risk index. The instruction unit is used to generate intervention instructions based on the spatial distribution of the warning level and the cumulative injury risk index. The intervention instructions include the turning direction, turning angle, and decompression measures.

Images