A pressure ulcer risk prediction method and intelligent anti-pressure ulcer mattress

Abstract

The application discloses a pressure ulcer risk prediction method and an intelligent anti-pressure ulcer mattress, acquires pressure distribution data collected by a distributed pressure sensor array, carries out signal filtering and standardization processing on the pressure distribution data, and obtains a standardized pressure distribution matrix; a binary random forward algorithm is applied for feature selection, and an optimal feature subset is output; patient sign parameters are acquired, the patient sign parameters are fused with the optimal feature subset, an enhanced convolutional neural network with an activation function is used for training, and a pressure load distribution atlas is generated; a differential evolution algorithm with neighborhood mutation is applied to calculate risk assessment parameters of each body region, and a pressure ulcer risk score and development trend prediction are output; optimal air bag inflation and deflation parameters are calculated through an intelligent decision algorithm, execution instructions are generated to an air bag control system, and nursing intervention suggestions are generated simultaneously. The application realizes accurate prediction and active protection of pressure ulcer risks, and improves the effect of pressure ulcer prevention.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical care, in particular to a pressure ulcer risk prediction method and an intelligent pressure ulcer prevention mattress, which is used for real-time monitoring, prediction and active protection of pressure ulcer risk of patients who are long-term bedridden or have difficulty in moving. BACKGROUND

[0002] Pressure ulcer (pressure injury) refers to the damage to the skin and subcutaneous tissue caused by the continuous pressure on the local tissue due to the patient's long-term in the same position. This is particularly common in patients who are long-term bedridden or have difficulty in moving, which not only affects the quality of life of patients, but also may lead to serious complications and even threaten life.

[0003] The common pressure ulcer prevention techniques in the clinic mainly include passive prevention measures such as regular turning over and using air mattress. The traditional air mattress usually adopts an alternating inflation and deflation mode with fixed time interval, or a simple uniform pressure bearing design, which cannot be accurately adjusted according to the individual differences and real-time state of the patient. Some advanced medical institutions have adopted mattress systems with pressure detection function, but such systems can only detect pressure values and cannot make risk prediction and active intervention.

[0004] More advanced pressure ulcer prevention systems use distributed pressure sensing technology to collect pressure data on the body contact surface, and assess the risk area through simple threshold judgment or statistical analysis method. Although this technology realizes basic pressure monitoring, it has obvious shortcomings in the accuracy and intelligence of risk prediction, especially the lack of consideration of individual differences and long-term pressure accumulation effect of patients.

[0005] The existing technology has the following main defects: first, it lacks accurate individualized pressure ulcer risk prediction ability and cannot dynamically assess the risk according to the specific situation of the patient; second, most systems use preset inflation and deflation modes and cannot be intelligently adjusted according to real-time detection data; third, existing systems generally lack efficient machine learning algorithm support and are difficult to process and analyze a large amount of real-time pressure data and make accurate predictions. SUMMARY

[0006] The purpose of the present application is to provide a pressure ulcer risk prediction method and an intelligent pressure ulcer prevention mattress, which aims to solve the technical problems of lack of accurate individualized pressure ulcer risk prediction ability, inability to intelligently adjust air bags and lack of efficient machine learning algorithm support in the prior art.

[0007] To achieve the above purpose, the present application provides a pressure ulcer risk prediction method, which comprises the following steps:

[0008] The pressure distribution data collected by the distributed pressure sensor array is obtained, the pressure distribution data is subjected to signal filtering and standardization processing, and a standardized pressure distribution matrix is obtained.

[0009] Based on the standardized pressure distribution matrix, a binary random forward algorithm is applied for feature selection, and an optimal feature subset is output;

[0010] Obtain patient sign parameters, fuse the patient sign parameters with the optimal feature subset, train using an enhanced convolutional neural network with an activation function, and generate a pressure load distribution atlas;

[0011] Based on the pressure load distribution atlas, a differential evolution algorithm with neighborhood mutation is applied to calculate the risk assessment parameters of each body region, and the pressure ulcer risk score and development trend prediction are output;

[0012] According to the pressure ulcer risk score and development trend prediction, the optimal air bag inflation and deflation parameters are calculated by an intelligent decision algorithm, and the execution instructions are generated to the air bag control system, and nursing intervention suggestions are generated.

[0013] Preferably, the pressure distribution data is subjected to signal filtering and standardization processing to obtain a standardized pressure distribution matrix, including:

[0014] Based on the pressure distribution data, a combination of median filtering and Kalman filtering is applied to remove noise and outliers, and cleaned pressure data is generated;

[0015] The cleaned pressure data is subjected to min-max standardization processing, and the numerical value is mapped to the [0, 1] interval to obtain standardized data;

[0016] According to the sensor spatial arrangement information and the standardized data, a two-dimensional pressure distribution matrix is reconstructed to obtain a standardized pressure distribution matrix.

[0017] Preferably, based on the standardized pressure distribution matrix, a binary random forward algorithm is applied for feature selection, and an optimal feature subset is output, including:

[0018] Based on the standardized pressure distribution matrix, a multi-dimensional feature pool containing statistical features, morphological features and time series features is constructed to form an initial feature set;

[0019] The initial feature set is subjected to binary coding representation to construct a binary representation model for feature selection;

[0020] An optimal feature combination is iteratively selected from the binary representation model using a random forward algorithm to obtain an optimal feature subset.

[0021] Preferably, an enhanced convolutional neural network with an activation function is used for training to generate a pressure load distribution atlas, including:

[0022] The optimal feature subset and the patient sign parameters are subjected to multi-modal data fusion to construct an input tensor to form network input data;

[0023] The deep neural network architecture includes 5 convolutional layers, the first 3 layers use 3x3 convolutional kernels, and the last 2 layers use 5x5 convolutional kernels to extract multi-scale features from the input data and output multi-level feature representations;

[0024] Based on the multi-level feature representation, an 8x16 feature map is generated through the last convolutional layer to obtain the pressure load distribution map.

[0025] Preferably, the activation function α and β are learnable parameters that are automatically adjusted during the training process through backpropagation.

[0026] Preferably, based on the pressure load distribution map, a differential evolution algorithm with neighborhood mutation is applied to calculate the risk assessment parameters of each body region, output the pressure ulcer risk score and development trend prediction, including:

[0027] Based on the pressure load distribution map, combined with the pre-set body region pressure ulcer susceptibility weight, a multi-dimensional risk assessment parameter vector including pressure index, duration, and tissue tolerance is constructed to form an initial parameter space;

[0028] Latin hypercube sampling is used to initialize the initial parameter space to generate a diversified risk assessment parameter vector combination containing 30 individuals;

[0029] When the differential evolution algorithm falls into local optimum, a small range search is performed in the neighborhood space around the current optimal solution based on the diversified risk assessment parameter vector combination to generate an optimized parameter search strategy;

[0030] Based on the optimized parameter search strategy, adaptive crossover and selection strategies are implemented on the diversified risk assessment parameter vector combination to dynamically adjust the crossover probability according to the population diversity, optimize the parameter combination, and generate a new generation of risk assessment parameter vector combination;

[0031] Using the new generation of risk assessment parameter vector combination, the pressure ulcer risk score of each body region is calculated, and the risk change trend in the next 4 hours is estimated through a time series prediction model to obtain the pressure ulcer risk score and development trend prediction.

[0032] Preferably, the pressure ulcer risk score is calculated, including:

[0033] Based on the pressure index, duration, tissue tolerance, pressure change rate, and cumulative risk parameters of each body region, a multi-factor risk model is used for weighted calculation to output a pressure ulcer risk score in the range of 0-100;

[0034] According to the pressure ulcer risk score, the risk level is divided: 0-25 points for low risk, 26-50 points for low-medium risk, 51-75 points for medium-high risk, and 76-100 points for high risk, to generate a risk level classification result.

[0035] Preferably, according to the pressure ulcer risk score and development trend prediction, the optimal air bag inflation and deflation parameters are calculated by an intelligent decision algorithm, and an execution instruction is generated to the air bag control system, including:

[0036] Based on the pressure ulcer risk score and development trend prediction, high-risk areas and medium-risk areas are identified and the ideal pressure distribution target value of each area is calculated to determine the target body area that needs pressure redistribution;

[0037] A transfer function model of air bag pressure and mattress surface pressure is established, and the optimal inflation parameters and optimal deflation parameters of each air bag are calculated by a quadratic programming algorithm for the target body area and the ideal pressure distribution target value to generate a control instruction sequence;

[0038] According to the control instruction sequence, the micro air pump and electromagnetic valve are controlled by PWM signals to adjust the pressure of each air bag, and an execution instruction is generated to the air bag control system.

[0039] Preferably, nursing intervention suggestions are generated simultaneously, including:

[0040] Based on the pressure ulcer risk score and development trend prediction, high-risk areas that cannot be completely alleviated by air bag adjustment are identified, and key nursing areas that need manual intervention are determined;

[0041] Combined with the patient's historical data and the risk level classification result, the decision tree algorithm is used to determine the optimal intervention strategy for the key nursing areas to generate personalized nursing intervention suggestions;

[0042] When the pressure ulcer risk score of any area in the key nursing area is continuously monitored to be more than 85 points and the prediction trend of the development trend prediction is rising, a turning-over reminder is automatically generated based on the personalized nursing intervention suggestions, and the nursing staff is notified through sound and light prompts and mobile terminal push notifications to complete the nursing intervention notification.

[0043] The present application also provides an intelligent anti-pressure ulcer mattress, comprising: a mattress body on which a plurality of pressure sensors are arranged in an array; an air bag unit arranged in the mattress body; an air bag control system for controlling the inflation and deflation of the air bag unit; a control unit comprising a memory, a processor and a program stored on the memory and executable on the processor, wherein the processor implements the pressure ulcer risk prediction method as described above when executing the program.

[0044] The present application has the following advantages:

[0045] 1. Through a distributed pressure sensor array and an efficient data processing algorithm, the pressure distribution of each area of the patient's body is accurately monitored and analyzed;

[0046] 2. The feature selection mechanism based on the binary random forward algorithm significantly reduces the feature calculation energy consumption while ensuring the prediction accuracy, realizing the long-term stable operation of the system;

[0047] 3. The new activation function enhanced convolutional neural network architecture significantly improves the sensitivity of the model to small pressure changes and the recognition ability of the pressure distribution pattern;

[0048] 4. The differential evolution algorithm with neighborhood mutation mechanism is introduced to enhance the ability of the algorithm to jump out of the local optimum and improve the accuracy and efficiency of the pressure sore risk assessment model parameter optimization;

[0049] 5. Through the individualized risk assessment mechanism of multi-modal data fusion, the individual differences of different patients are accurately described, and the individualized level of risk prediction is improved;

[0050] 6. The intelligent airbag control system based on risk prediction realizes the accurate control of independent airbags and achieves the purpose of active prevention of pressure sores. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0052] Figure 1 The flowchart of the pressure sore risk prediction method of the present application;

[0053] Figure 2 The enhanced convolutional neural network architecture of the present application;

[0054] Figure 3 The structure diagram of the intelligent anti-pressure sore mattress of the present application. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below. In the following description, a lot of specific details are set forth in order to fully understand the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of the present application, therefore the present application is not limited by the specific embodiments disclosed below.

[0056] It is to be understood that when an element such as a layer, region or substrate is referred to as being "on" or "connected to" another element, it can be directly on or connected to the other element or intervening elements can be present. In contrast, when an element such as a layer, region or substrate is referred to as being "directly on" or "directly connected to" another element, there are no intervening elements present. It will be understood that, when a layer is referred to as being "on" or "connected to" another layer, it has a function of a known intermediate layer. The same can apply to the term "electrically connected". It will be understood that, for the purposes of this disclosure, "at least one", "one or more" and "multiple" include singular as well as plural.

[0057] In addition, the terms "first", "second", etc. are used herein only to describe different instances, and are not used to denote or imply relative importance or an order of magnitude. Thus, features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise explicitly specified and limited.

[0058] In the present application, unless otherwise explicitly specified and limited, the "on", "under", "above" and "over" of the first feature to the second feature can be that the first feature is in direct contact with the second feature, or the first feature is indirectly in contact with the second feature through an intermediate medium. Moreover, the "on", "above" and "over" of the first feature to the second feature can be that the first feature is directly above or obliquely above the second feature, or only indicates that the first feature is horizontally higher than the second feature. The "under", "below" and "under" of the first feature to the second feature can be that the first feature is directly below or obliquely below the second feature, or only indicates that the first feature is horizontally lower than the second feature.

[0059] Unless otherwise defined, all technical and scientific terms used in the specification of the present application have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The term "and / or" used in the specification of the present application includes any and all combinations of one or more related listed items.

[0060] The present application will be further described in detail below with reference to the accompanying drawings and examples.

[0061] As shown in Figure 1 The pressure sore risk prediction method provided by the present application comprises the following steps:

[0062] Step S1: acquiring pressure distribution data collected by a distributed pressure sensor array, performing signal filtering and standardization processing on the pressure distribution data, and obtaining a standardized pressure distribution matrix;

[0063] Step S2: based on the standardized pressure distribution matrix, applying a binary random forward algorithm for feature selection, and outputting an optimal feature subset;

[0064] Step S3: Obtain patient vital signs parameters, fuse the patient vital signs parameters with the optimal feature subset, train using an enhanced convolutional neural network with activation functions, and generate a pressure load distribution map;

[0065] Step S4: Based on the pressure load distribution map, apply the differential evolution algorithm with neighborhood mutation to calculate the risk assessment parameters of each body region, and output the pressure ulcer risk score and development trend prediction.

[0066] Step S5: Based on the pressure ulcer risk score and the development trend prediction, the optimal airbag inflation and deflation parameters are calculated using an intelligent decision-making algorithm, and an execution command is generated to the airbag control system. At the same time, nursing intervention suggestions are generated.

[0067] In step S1, pressure distribution data collected by a distributed pressure sensor array is first acquired. The high-precision flexible pressure sensor array built into the smart mattress adopts an 8×16 matrix arrangement, covering all pressure areas of the human body. Each sensor unit is 2cm×2cm in size, with a measurement range of 0-200mmHg and an accuracy of ±1mmHg. The sensor sampling frequency is set to 10Hz to achieve high-density real-time monitoring of the pressure distribution on the mattress surface, ensuring that minute pressure changes and distribution characteristics can be captured. The sensor array transmits the real-time collected pressure data to the control unit via Bluetooth Low Energy technology. The data includes sensor ID, timestamp, and pressure value triplet information, forming a raw pressure data stream. The patient's basic body position information (supine, lateral, prone, etc.) is recorded simultaneously as auxiliary annotation data, which helps in subsequent analysis of pressure distribution characteristics under different body positions. After acquiring the raw pressure data, signal filtering processing is performed to remove unavoidable noise and outliers during the acquisition process. Data cleaning was performed using a combination of median filtering and Kalman filtering. Median filtering primarily removed sudden spikes in noise, while Kalman filtering, by establishing state and observation equations and combining prior information with current observations, provided an optimal estimate of the state, effectively suppressing random fluctuations and noise. A sliding window size of 5 sampling points was set, and data points exceeding ±3σ were detected and corrected to ensure data continuity and reliability. The filtered pressure data underwent min-max normalization, mapping the values ​​to the [0,1] interval to eliminate dimensional influences and allow for fair comparison of data from different pressure levels. Finally, based on the spatial arrangement of the sensors and the normalized data, a two-dimensional pressure distribution matrix was reconstructed, forming an 8×16 normalized pressure distribution matrix. Each element in the matrix represents the normalized pressure value at the corresponding location, intuitively reflecting the pressure distribution on the mattress surface. Furthermore, based on the continuously acquired pressure distribution matrix, temporal features such as pressure duration and pressure change rate at each point were calculated, constructing a pressure feature set containing both spatial and temporal dimensions. This resulted in an enhanced dataset containing static pressure distribution and dynamic pressure change features, providing a comprehensive information foundation for subsequent pressure ulcer risk prediction.

[0068] In step S2, based on the standardized pressure distribution matrix, a binary random forward algorithm is applied for feature selection to output the optimal feature subset. First, a multi-dimensional feature pool containing statistical, morphological, and temporal features is constructed to form an initial feature set. Statistical features describe the basic statistical properties of pressure distribution, including global pressure statistics (such as average pressure, maximum pressure, minimum pressure, pressure standard deviation, pressure median, etc.), local statistics (local pressure statistics for key areas such as the sacrum, coccyx, heel, and scapula), pressure distribution ratio (the proportion of area covered by different pressure levels), and pressure concentration (quantifying the unevenness of pressure distribution using the Gini coefficient). Morphological features describe the spatial characteristics of pressure distribution, including the location of the pressure centroid (calculating the pressure-weighted centroid coordinates, reflecting the overall pressure center), pressure distribution shape characteristics (including parameters such as moment of inertia, dispersion, and eccentricity), high-pressure zone morphology (area, perimeter, shape complexity, principal axis direction, etc. of the high-pressure zone), and pressure gradient (the rate of pressure change between adjacent locations, reflecting the smoothness of the pressure distribution). Temporal features describe the dynamic characteristics of pressure changes over time, including duration of pressure (the time during which continuous pressure exceeds a threshold at each location), rate of pressure change (the magnitude of pressure change per unit time), frequency of pressure fluctuations (reflecting the frequency of minor postural adjustments by the patient), and pressure release patterns (the frequency and duration of temporary pressure releases, characterizing the patient's ability to relieve pressure independently). The initial feature pool contains over 200 feature dimensions, providing a rich pool of candidates for subsequent feature selection. Next, the initial feature set is represented using binary encoding to construct a binary representation model for feature selection. Each feature in the feature pool is assigned one bit, forming a binary vector of length N (N being the total number of features, approximately 200+). The value of each bit in this vector follows a rule: a value of 1 indicates selection of the feature, and a value of 0 indicates non-selection. This binary encoding representation transforms the complex feature selection problem into a binary optimization problem, facilitating algorithmic operation. Then, a binary random forward algorithm is used for feature selection. The algorithm first randomly selects approximately 10% of all features as an initial feature subset, constructs an initial evaluation model, and calculates performance metrics. In each iteration, a subset of candidate features is randomly selected from the unselected feature set. The value of each candidate feature is evaluated, and the feature that brings the greatest performance improvement is added to the feature subset. The model and performance metrics are then updated. This process continues until a termination condition is met. Finally, the selected feature subset is evaluated for energy consumption, and features with energy consumption ratios below a threshold are removed to ensure that energy consumption is minimized while maintaining prediction accuracy. The final output is an optimal feature subset containing 40-60 key features. These features are characterized by strong predictive power, high energy efficiency, good complementarity, and strong interpretability, providing high-quality input data for subsequent risk prediction.

[0069] In step S3, patient vital signs parameters are acquired and fused with the optimal feature subset. An enhanced convolutional neural network with activation functions is then used for training to generate a stress load distribution map. First, patient vital signs parameters are collected, including basic demographic characteristics (such as age, sex, weight, height, BMI, etc.), clinical assessment indicators (such as the six dimensions of the Braden score), underlying disease information (such as diabetes, vascular disease, neurological disease, etc.), laboratory test results (such as albumin level, hemoglobin, total protein, etc.), and medication use (such as steroids, sedatives, and other drugs that may affect tissue tolerance). These parameters are automatically acquired through a hospital information interface or manually entered by medical staff. Next, the patient vital signs parameters are fused with the optimal feature subset output from step S2 using multimodal data fusion to construct an input tensor. Stress features form the main spatial channels, maintaining their spatial relationships (8×16 grid), while patient vital signs parameters are incorporated into the model as additional feature channels, ensuring the model comprehensively considers stress distribution and individual differences. Then, a specialized neural network architecture is designed, containing a deep structure with five convolutional layers. The first three layers use 3×3 convolutional kernels to focus on capturing local stress patterns, while the last two layers use 5×5 convolutional kernels to capture a wider range of stress distribution features. The network employs a skip connection structure to enhance gradient flow and feature propagation, improving model training stability. An innovative activation function is used. The activation function, α and β, are learnable parameters. This function combines the linearity of ReLU with the nonlinear saturation of Sigmoid, effectively capturing nonlinear patterns and subtle changes in pressure distribution. Supervised learning is performed using a historical dataset containing known pressure ulcer risk ratings. A weighted cross-entropy loss function is employed to focus on high-risk samples. The Adam optimizer is used for parameter updates, and L2 regularization and Dropout are introduced to prevent overfitting. After training, the model can take into account the real-time pressure characteristics and vital signs of the current patient and generate an 8×16 pressure load distribution map through the last convolutional layer. The value at each position in this map represents the pressure ulcer risk index (between 0 and 1) for the corresponding area. Different risk levels are visually represented by color coding: low-risk areas are shown in green, low-to-medium risk areas in yellow, medium-to-high risk areas in orange, and high-risk areas in red. This visualization allows healthcare professionals to intuitively understand the pressure ulcer risk distribution across different areas of the patient's body, providing a clear reference for subsequent risk assessment and intervention decisions.

[0070] In step S4, based on the pressure load distribution map, a differential evolution algorithm with neighborhood mutation is applied to calculate the risk assessment parameters for each body region, outputting a pressure ulcer risk score and a predicted development trend. First, based on the pressure load distribution map and combined with pressure ulcer susceptibility weights for different body parts from medical literature (sacrum-coccyx 1.5, heel 1.3, hip 1.2, etc.), a multidimensional risk assessment parameter vector is constructed, incorporating factors such as pressure index, duration, and tissue tolerance. These parameters include: pressure intensity parameters (based on the risk index in the pressure load distribution map, combined with region weights), duration parameters (representing the duration of a specific pressure level), tissue tolerance parameters (reflecting the patient's tissue resistance to pressure), pressure change rate parameters (reflecting the rate of pressure change over time), and cumulative risk parameters (considering the cumulative effect of historical risk). Next, Latin hypercube sampling is used to initialize the initial parameter space, generating a diverse combination of risk assessment parameter vectors containing 30 individuals. Latin hypercube sampling ensures that samples are uniformly distributed across each dimension, improving the coverage efficiency of the parameter space. Then, an innovative neighborhood mutation operation is implemented to enhance the algorithm's ability to escape local optima. When the algorithm might get stuck in a local optimum, it performs a targeted local search within the "neighborhood" of the current optimal solution, balancing global exploration with local exploitation capabilities. An adaptive crossover strategy is also implemented, dynamically adjusting the crossover probability based on population diversity: a smaller crossover probability is used when diversity is high to maintain diversity, and a larger crossover probability is used when diversity is low to promote information exchange. The selection operation employs an elite retention mechanism to ensure that the best individuals in each generation are not lost, while tournament selection is introduced to increase population diversity. Finally, the optimized parameters are used to calculate pressure ulcer risk scores for different areas of the patient's body. For each body area, a multi-factor risk model is used for weighted calculation based on the pressure index, duration, tissue tolerance, pressure change rate, and cumulative risk parameters of that area, outputting a pressure ulcer risk score in the range of 0-100. Based on the risk score, each area is divided into different risk levels: 0-25 points for low risk, 26-50 points for low-to-medium risk, 51-75 points for medium-to-high risk, and 76-100 points for high risk. Furthermore, based on historical data, the system predicts the changing trends of risk scores in each area over the next four hours, providing healthcare professionals with forward-looking risk information to support timely and targeted interventions. This comprehensive risk assessment method, combining current risk scores with future trend predictions, significantly improves the foresight and effectiveness of pressure ulcer prevention.

[0071] In step S5, based on the pressure ulcer risk score and development trend prediction, the optimal airbag inflation / deflation parameters are calculated using an intelligent decision-making algorithm. Execution instructions are then sent to the airbag control system, and nursing intervention suggestions are generated simultaneously. First, based on the risk score and trend prediction, the system identifies high-risk areas (risk score > 70) and medium-risk areas (risk score 40-70), and calculates the target value for the ideal pressure distribution in each area. For high-risk areas, the goal is to significantly reduce surface pressure, with the ideal pressure value set at 50-60% of the current pressure. For medium-risk areas, the goal is to moderately reduce surface pressure, with the ideal pressure value set at 70-80% of the current pressure. For low-risk areas, pressure can be appropriately increased to compensate for the decompression in high-risk areas, but the increase should not exceed 20%, and the absolute pressure value should not exceed the safety threshold. Next, a transfer function model between airbag pressure and mattress surface pressure is established to describe how changes in airbag internal pressure affect the pressure distribution on the mattress surface. Based on this model, the pressure redistribution problem is transformed into an optimization problem: finding a set of airbag internal pressure values ​​that makes the actual pressure distribution on the mattress surface as close as possible to the target value for the ideal pressure distribution. This optimization problem is solved using a quadratic programming algorithm to calculate the optimal inflation and deflation parameters for each airbag, forming a complete control command sequence. Then, based on this sequence, a micro-pump and solenoid valve are controlled via PWM signals to precisely adjust the pressure of each airbag. A closed-loop control mechanism is employed, using pressure sensors built into the airbags to monitor the internal pressure in real time. The measured values ​​are compared with target values, and the PWM signal parameters are dynamically adjusted to ensure the airbag pressure accurately reaches the set value. The entire airbag control system achieves a minimum adjustment accuracy of ±1 mmHg and a response time of less than 2 seconds, enabling rapid response to changes in risk assessment results and timely adjustment of the pressure distribution on the mattress surface. Simultaneously, high-risk areas that cannot be fully relieved by airbag adjustment are identified, determining key care areas requiring manual intervention. For these areas, a decision tree algorithm is used to determine the optimal intervention strategy, generating personalized care intervention recommendations, based on patient history data and risk level classification results. These recommendations include turning procedures (suggested turning angles, frequencies, and specific operating methods), local decompression measures (suggestions for using specific positioning pads, air cushions, or other decompression devices), and skin care procedures (specific operational instructions for cleaning, moisturizing, massage, etc.). When the pressure ulcer risk score for any key care area continuously exceeds 85 points and the predicted trend is upward, a turning reminder is automatically generated and notified to caregivers via audio-visual prompts and mobile device push notifications. A priority management and escalation mechanism for notifications is also implemented to ensure timely handling of emergencies. This combination of machine assistance and human intervention provides a comprehensive and personalized pressure ulcer prevention solution, significantly improving prevention effectiveness and care efficiency.

[0072] Step S1 involves signal filtering and standardization of the pressure distribution data to obtain a standardized pressure distribution matrix, including:

[0073] Step S1.1: Based on the pressure distribution data, noise and outliers are removed by combining medium-range filtering and Kalman filtering to generate cleaned pressure data;

[0074] Step S1.2: Perform min-max normalization on the pressure data after cleaning, and map the values ​​to the [0,1] interval to obtain normalized data;

[0075] Step S1.3: Based on the sensor spatial arrangement information and the standardized data, reconstruct the two-dimensional pressure distribution matrix to obtain the standardized pressure distribution matrix.

[0076] In step S1.1, the raw pressure data collected from the distributed pressure sensor array is filtered to remove unavoidable noise and outliers during the acquisition process. The high-precision flexible pressure sensor array built into the smart mattress adopts an 8×16 matrix arrangement, covering the entire pressure area of ​​the human body. Each sensor unit is 2cm×2cm in size, with a measurement range of 0-200mmHg, an accuracy of ±1mmHg, and a sampling frequency set to 10Hz. These sensors collect pressure distribution data on the mattress surface in real time and transmit the data to the control unit via Bluetooth Low Energy technology, forming a raw pressure data stream containing a triplet of sensor ID, timestamp, and pressure value information. However, due to environmental interference, sensor fluctuations, and interference during data transmission, the raw data usually contains noise and outliers, requiring filtering to improve data quality. A combination of median filtering and Kalman filtering is used for data cleaning. First, median filtering is applied to remove sudden spikes in noise and outliers. The principle of median filtering is to use a sliding window (in this case, the window size is set to 5 sampling points) to sort the data within the window and take the median value as the filtering result. This nonlinear filtering method is particularly effective in removing impulse noise while preserving the edge characteristics of the data. Next, Kalman filtering is applied to further smooth the data and estimate the true state. Kalman filtering is a recursive estimator that establishes state equations and observation equations, combining prior information and current observations to make an optimal estimate of the state. For pressure data, the state can be the true pressure value at the sensor location, and the observation value is the sensor reading. Kalman filtering continuously optimizes the estimate of the true pressure through a prediction-update loop, effectively suppressing random fluctuations and noise. In addition, an outlier detection mechanism is set up to mark and correct data points that exceed ±3σ (σ is the local standard deviation) to ensure the continuity and reliability of the data. By combining median filtering and Kalman filtering, noise and outliers in the original pressure data can be effectively removed, generating cleaned, high-quality pressure data, providing a reliable foundation for subsequent data analysis and risk assessment.

[0077] In step S1.2, the cleaned pressure data is standardized to map data of different magnitudes and units to a uniform numerical range, facilitating subsequent processing and analysis. The original pressure data ranges from 0 to 200 mmHg. Pressure distribution varies significantly among different patients and in different positions. Directly using the original data may lead to an overemphasis on high-pressure values ​​in the model, while ignoring subtle changes in the low-pressure region. Therefore, a min-max standardization method is used to linearly map the cleaned pressure data to the [0,1] interval. For each sensor location... pressure value The standardized formula is: .

[0078] in, and These are the minimum and maximum non-zero values ​​in the current pressure dataset, respectively. Choosing the minimum non-zero value instead of absolute zero as the lower bound avoids a large number of pressure-free regions (values ​​of 0) remaining at 0 after standardization, leading to information loss. This standardization process ensures that all pressure data are mapped to the same numerical range, eliminating the influence of dimensions and allowing for fair comparison of data from different pressure levels. Furthermore, standardization helps improve the convergence speed and stability of subsequent machine learning algorithms, as most machine learning algorithms perform better when processing standardized data. For abnormally large pressure values ​​(potentially due to sensor malfunction or temporary interference), truncation is performed before standardization, limiting values ​​exceeding a reasonable range (e.g., >200 mmHg) to a maximum threshold to prevent extreme values ​​from distorting the overall standardization results. After standardization, the original physical pressure values ​​are converted into dimensionless values ​​between 0 and 1, where 0 represents no pressure or minimum pressure, 1 represents maximum pressure, and intermediate values ​​are distributed proportionally. This standardized data preserves the relative relationships and patterns of the original pressure distribution while exhibiting better numerical stability and comparability.

[0079] In step S1.3, a two-dimensional pressure distribution matrix is ​​reconstructed based on the spatial arrangement information of the sensors and standardized data, forming a complete standardized pressure distribution matrix. The sensor array adopts an 8×16 matrix arrangement, covering the main stress areas of the mattress surface, with each sensor corresponding to a specific location on the mattress surface. During data transmission, sensor data is transmitted in the form of sensor ID, timestamp, and pressure value. Based on the mapping relationship between sensor ID and its physical location, the discrete sensor readings need to be reconstructed into a two-dimensional matrix reflecting the spatial distribution. Maintaining a sensor ID to matrix location... A mapping table is used to place the standardized pressure value of each sensor into the corresponding position in a two-dimensional matrix, forming an 8×16 standardized pressure distribution matrix P. Each element in the matrix... Indicates position The standardized pressure values ​​are located between [0,1]. For potential data gaps (e.g., temporary sensor malfunction or data loss), interpolation is used to fill in the missing data. Isolated missing points are filled using the average of surrounding valid points; for continuous missing regions, bilinear interpolation is used to estimate the pressure value at the missing location based on surrounding valid data. After reconstruction, the entire matrix is ​​smoothed using a Gaussian filter to reduce abrupt changes between adjacent locations, generating a continuous and smooth matrix that better reflects the actual pressure distribution. This smoothing process considers the physical characteristics of the actual pressure distribution, where pressure typically varies continuously rather than abruptly in space. The final output 8×16 standardized pressure distribution matrix is ​​a complete, continuous, and standardized two-dimensional data structure that intuitively reflects the pressure distribution on the mattress surface. This matrix can be used for visualization (presenting the pressure distribution intuitively through heatmaps, etc.) and serves as the foundational data for subsequent feature extraction and risk assessment. In addition, based on the continuously collected pressure distribution matrix, the temporal characteristics such as pressure duration and pressure change rate at each point are calculated to construct a pressure feature set containing spatial and temporal dimensions. The output is an enhanced dataset containing static pressure distribution and dynamic pressure change characteristics, providing more comprehensive information for subsequent pressure ulcer risk prediction.

[0080] In step S2, based on the standardized pressure distribution matrix, a binary random forward algorithm is applied to perform feature selection, outputting the optimal feature subset, including:

[0081] Step S2.1: Based on the standardized pressure distribution matrix, construct a multi-dimensional feature pool containing statistical features, morphological features, and temporal features to form an initial feature set;

[0082] Step S2.2: Represent the initial feature set using binary encoding to construct a binary representation model for feature selection;

[0083] Step S2.3: Use a random forward algorithm to iteratively select the optimal feature combination from the binary representation model to obtain the optimal feature subset.

[0084] In step S2.1, a comprehensive feature pool is constructed based on the standardized pressure distribution matrix to capture key information about the pressure distribution from different dimensions. First, three core features are extracted: statistical features, morphological features, and temporal features. Statistical features describe the basic statistical characteristics of the pressure distribution, including global pressure statistics (such as average pressure, maximum pressure, minimum pressure, pressure standard deviation, and pressure median), local statistics (statistical descriptions of local pressure in key areas such as the sacrum, heel, and scapula), pressure distribution proportions (the percentage of area covered by different pressure levels), and pressure concentration (quantifying the unevenness of the pressure distribution using the Gini coefficient). Morphological features describe the spatial characteristics of the pressure distribution, including the location of the pressure centroid (calculating the pressure-weighted centroid coordinates to reflect the overall pressure center), pressure distribution shape characteristics (including parameters such as moment of inertia, dispersion, and eccentricity), high-pressure zone morphology (area, perimeter, shape complexity, and principal axis direction of the high-pressure zone), and pressure gradient (the rate of pressure change between adjacent locations, reflecting the smoothness of the pressure distribution). Temporal features describe the dynamic characteristics of pressure changes over time, including duration of pressure (the duration for which continuous pressure at each location exceeds a threshold), rate of pressure change (the magnitude of pressure change per unit time), frequency of pressure fluctuations (reflecting the frequency of minor postural adjustments by the patient), and pressure release patterns (the frequency and duration of temporary pressure releases, characterizing voluntary decompression ability). These features are extracted from the raw pressure data through mathematical calculations and signal processing methods, with the initial feature pool containing over 200 feature dimensions. For example, for a specific region r, its pressure duration feature can be represented as: Where I is an indicator function, representing the pressure in region r at time t. The value is 1 when the threshold is exceeded, and 0 otherwise. Constructing this comprehensive feature pool provides a rich pool of candidates for subsequent feature selection, ensuring that various pressure patterns associated with pressure ulcer risk can be captured.

[0085] In step S2.2, the feature selection problem is transformed into a mathematical form suitable for algorithmic processing. Each feature in the feature pool constructed in step S2.1 is assigned one binary bit, forming a binary vector of length N (N being the total number of features, approximately 200+). This vector B is represented as... Each binary bit The rules for determining the value are: This indicates selecting the i-th feature. This indicates that the i-th feature is not selected. For example, for a feature pool containing 200 features, a possible feature selection scheme can be represented as follows: This indicates that features 1, 3, and 6 were selected, while features 2, 4, and 5 were discarded. The advantages of this binary encoding representation are: simplified problem expression (transforming the complex feature selection problem into a binary optimization problem), ease of algorithm operation (binary vectors are easy to manipulate, such as flipping and crossing), and a clear state space (a binary vector of length N can represent 2^N different feature combinations). A mapping table is established between feature IDs and binary bit positions, with each feature selection scheme (i.e., a binary vector) corresponding to a possible feature subset. When evaluating the performance of a feature subset, the binary vector is used as a mask, extracting only the features corresponding to positions with a value of 1 for model training and evaluation. This binary encoding method provides a good representational foundation for subsequent stochastic forward algorithms, enabling the algorithm to efficiently search for optimal feature combinations in a solution space of 2^N.

[0086] In step S2.3, the random forward algorithm is used. Feature selection is performed. First, k features are randomly selected from all N features as an initial feature subset. (k is set to 10% of the total number of features; for example, if the total number of features is 200, approximately 20 features will be initially selected). The random selection process uses a uniformly distributed pseudo-random number generator to ensure that each feature has an equal probability of being selected. Next, based on the initial feature subset... Constructing an initial evaluation model (This can be a simple logistic regression or decision tree model), and calculate the initial model. performance indicators This performance metric is a weighted combination of information gain and computational complexity, defined as:

[0087]

[0088] in, The predictive power of a model can be measured by AUC or F1 score. The complexity of the feature subset can be measured by the number of features or the energy required to compute them; α and β are weighting parameters that balance prediction performance and complexity. Then, the iterative process begins: m candidate features are randomly selected from the unselected feature set (m is set to 20% of the remaining features); the value of each candidate feature is evaluated sequentially, and the incremental performance improvement is calculated. Select the features that deliver the greatest performance improvement. If the performance improvement brought by the best candidate feature is positive, add it to the feature subset and update the model and performance metrics. Check the termination condition (no significant performance improvement in t consecutive iterations, reaching the preset maximum number of iterations, or the feature subset size reaching the preset upper limit). If the condition is met, stop; otherwise, continue to the next iteration. Finally, perform energy consumption optimization on the selected feature subset and calculate the information gain to energy consumption ratio for each feature. Remove Features below a preset threshold are selected to ensure that energy consumption is minimized while maintaining prediction accuracy. The final output contains an optimal feature subset of 40-60 key features, which are characterized by strong predictive power, high energy efficiency, good complementarity, and strong interpretability.

[0089] like Figure 2 As shown, step S3 uses an enhanced convolutional neural network with activation functions for training to generate a pressure load distribution map, including:

[0090] Step S3.1: Perform multimodal data fusion of the optimal feature subset and the patient's vital signs parameters to construct an input tensor, forming network input data;

[0091] Step S3.2: Design a deep neural network architecture with 5 convolutional layers. The first 3 layers use 3×3 convolutional kernels and the last 2 layers use 5×5 convolutional kernels. Perform multi-scale feature extraction on the network input data and output multi-level feature representation.

[0092] Step S3.3: Based on the multi-level feature representation, an 8×16 feature map is generated through the last convolutional layer to obtain the pressure load distribution map.

[0093] In step S3.1, multimodal data fusion is achieved by combining stress features with patient vital signs parameters to create a comprehensive input data tensor. Two types of input data are collected and prepared: stress feature data (from the optimal feature subset output in step S2, containing 40-60 selected key features, such as stress values, duration, and rate of change in each region, organized into a structured matrix according to the body regions they describe) and patient vital signs parameters (including basic demographic characteristics such as age, sex, weight, height, and BMI; clinical assessment indicators such as the six dimensions of the Braden score; underlying disease information such as diabetes, vascular disease, and neurological disease; laboratory test results such as albumin levels, hemoglobin, and total protein; and medication use such as steroids and sedatives, which may affect tissue tolerance). These two types of data are preprocessed and standardized: numerical features are standardized to a similar numerical range using Z-score or Min-Max methods; categorical features are converted to numerical representations using one-hot encoding; and missing values ​​are filled using multiple imputation or model-based methods. Then, a multimodal input data tensor is constructed. For a convolutional neural network architecture, the data is organized into basic input dimensions. Where B is the batch size, C is the number of channels, and H and W are the spatial dimensions. Stress features form the main spatial channels, maintaining their spatial relationships (8×16 grid). Patient vital signs are incorporated into the model in two ways: as additional feature channels (extended to the same spatial dimension as the stress features via broadcasting) and fused with the stress features extracted from the convolutional layers in subsequent layers after being processed by separate fully connected layers. The advantages of this multimodal data fusion method are: capturing the interaction between stress data and individual patient characteristics, allowing the model to learn differentiated risk patterns across different patient groups, and improving the personalization of predictions (the same stress distribution produces different risk predictions for different patients).

[0094] In step S3.2, a specialized neural network architecture was designed and optimized for the pressure ulcer risk prediction task. The network as a whole adopts a deep convolutional structure, containing 5 convolutional layers. The core idea of ​​the architecture design is to extract pressure features from local to global layers layer by layer. The input layer receives the multimodal input data tensor from step S3.1, with the shape of... The local feature extraction layers (the first three convolutional layers) use a small receptive field (3×3) to focus on capturing local pressure patterns, such as high pressure points, pressure gradients, and other detailed features: the first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1 and padding of 1, followed by batch normalization and a novel activation function; the second convolutional layer uses 64 3×3 convolutional kernels with a stride of 1 and padding of 1, followed by batch normalization and a novel activation function; the third convolutional layer uses 128 3×3 convolutional kernels with a stride of 1 and padding of 1, followed by batch normalization and a novel activation function. The global feature extraction layer (the last two convolutional layers) uses a larger receptive field (5×5) to capture a wider range of stress distribution patterns, such as stress center shifts and overall distribution shapes. The fourth convolutional layer uses 128 5×5 convolutional kernels with a stride of 1 and padding of 2, followed by batch normalization and a novel activation function. The fifth convolutional layer uses 64 5×5 convolutional kernels with a stride of 1 and padding of 2, followed by batch normalization and a novel activation function. The network also includes skip connection structures (skip connections are added between layers 1 and 5, and between layers 2 and 4, using identity mapping or 1×1 convolutions to achieve feature dimension matching), directly transferring shallow features to deeper layers, effectively mitigating the gradient vanishing problem, preserving detailed information, and improving training stability. The feature fusion layer uses an attention mechanism to adaptively weight and fuse features from different levels, achieving channel attention and spatial attention, dynamically adjusting the weights of different features according to their importance. The advantages of this enhanced convolutional neural network architecture are: multi-scale feature extraction (capturing local and global stress patterns through convolutional kernels of different sizes), gradient flow optimization (skip connections improve gradient flow, facilitating the training of deeper networks), feature preservation (preserving all useful information from shallow to deep layers), and attention enhancement (highlighting important features and suppressing irrelevant information).

[0095] In step S3.3, a pressure load distribution map is generated based on multi-level feature representation. Real-time data input for the current patient (the latest pressure data collected from distributed pressure sensors and the patient's vital signs) is acquired. A multimodal input data tensor is constructed according to the format in step S3.1, and the input data is fed into the trained model for forward propagation. The output shape of the last convolutional layer of the model is... The feature map is given by B, where B is the batch size (usually 1 for real-time prediction), and the value of each spatial location represents the pressure ulcer risk index of the corresponding area, ranging from 0 to 1. The raw output is post-processed to transform into a more informative stress load distribution map: risk level classification (classifying continuous risk indices into different risk levels based on preset thresholds: 0.0-0.25 for low risk / green, 0.25-0.5 for low-to-medium risk / yellow, 0.5-0.75 for medium-to-high risk / orange, and 0.75-1.0 for high risk / red); spatial interpolation enhancement (upgrading to a higher resolution, such as 64×128, through bicubic interpolation to create a smooth, high-resolution risk distribution map); body region mapping (aligning the risk distribution map with standard human anatomical atlases, identifying corresponding body regions such as the sacrum, coccyx, heels, and scapula, and adding region labels); heatmap rendering (using color mapping to render risk values ​​into an intuitive heatmap, with high-risk areas displayed in red and low-risk areas in green); and dynamic time window comparison (comparing the current risk distribution map with distribution maps from past time windows to generate a risk change trend map, showing areas where risk has increased or decreased). The final output pressure load distribution map is a comprehensive visualization that combines the output of the risk prediction model, spatial interpolation enhancement, body region mapping, and heatmap rendering. It visually displays the risk of pressure ulcers in various areas of the patient's body. This map is presented to healthcare professionals in real-time on a monitor and simultaneously stored in the patient's electronic medical record as a basis for subsequent nursing decisions and risk monitoring. The value of this map lies in transforming complex model predictions into intuitive visual information, allowing healthcare professionals to quickly identify body areas requiring focused attention and implement timely interventions without needing to understand the underlying algorithms. The activation function mentioned... α and β are learnable parameters that are automatically adjusted through backpropagation during training.

[0096] Specifically, traditional activation functions such as ReLU, Sigmoid, and Tanh each have their advantages and disadvantages: ReLU has linear response characteristics, is computationally simple, and can alleviate the gradient vanishing problem, but it completely suppresses negative inputs, which may lead to neuron "death"; the Sigmoid function has good nonlinear saturation characteristics, with its output range stable between (0,1), but the gradient is close to zero in the saturation region, which can easily lead to gradient vanishing; the Tanh function has an output range of (-1,1), and its zero-centeredness is beneficial for optimization, but it also suffers from saturation problems. For the specific task of pressure ulcer risk prediction, it is necessary to simultaneously capture minute pressure changes (linear response) and simulate the threshold effect of tissue damage (nonlinear saturation). Therefore, a parameterized activation function was designed. .

[0097] The mathematical structure of the activation function incorporates linear terms. and nonlinear modulation terms α controls the strength of the linear response, and β controls the steepness of the nonlinear saturation. When the input signal x is small, The function is approximately equal to It exhibits an approximately linear response, effectively capturing minute pressure changes; when the input signal x is large, As it approaches 1, the function approximates... Maintain linear growth but control the growth rate to avoid excessively large output values; when the input signal x is negative and has a large absolute value... Approaching zero, the function output is close to zero, but it does not completely suppress negative inputs, retaining some information. This design allows the activation function to exhibit different behavioral characteristics across different input ranges, making it particularly suitable for processing stress data and other signals that require both preserving subtle changes and handling threshold effects.

[0098] The initial values ​​of the α and β parameters have a significant impact on network training. Through experimental analysis, the system initially sets α to 1.0 and β to 0.5. This setting ensures that the activation function has moderate nonlinearity in the early stages of training, avoiding both excessive linearity leading to insufficient expressive power and excessive nonlinearity causing training instability. During network training, α and β, as learnable parameters, are updated along with other network parameters through the backpropagation algorithm.

[0099] Specifically, during the forward propagation phase, the activation value of each neuron is calculated. During the backpropagation phase, the gradients of the loss function L with respect to α and β are calculated: and Then, the parameters are updated using gradient descent: , where η is the learning rate.

[0100] To prevent the α and β parameters from diverging to unreasonable ranges, a parameter constraint mechanism was implemented: a range constraint was applied to the α parameter, limiting it to a certain range. Within the interval, ensure the activation function maintains a moderately linear response strength; apply range constraints to the β parameter, limiting it to... Within the interval, to prevent the nonlinear modulation term from becoming too steep or too flat. In addition, L2 regularization was applied to these two parameters to prevent them from taking too large values ​​and causing overfitting.

[0101] As training progresses, the α and β parameters of different layers converge to different values, reflecting the varying demands of each layer for linear and nonlinear characteristics. Typically, shallower networks (closer to the input) have smaller α values ​​and larger β values, indicating that these layers are more focused on the nonlinear transformation of the input; deeper networks (closer to the output) have larger α values ​​and moderate β values, indicating that these layers need to retain more linear information to generate accurate risk predictions. Experimental results show that, compared to traditional activation functions with fixed parameters, this learnable activation function improves the model's accuracy on the pressure ulcer risk prediction task by 8.3%, particularly demonstrating greater sensitivity in identifying early risk signs and capturing subtle changes in pressure patterns that might be overlooked by traditional methods.

[0102] This innovative activation function design demonstrates the technical depth of this invention. By introducing learnable parameters and allowing the network to automatically adjust the shape of the activation function, the system can better adapt to the needs of the specific task of pressure ulcer risk prediction, improve the accuracy and sensitivity of prediction, and provide more reliable decision support for clinical practice.

[0103] In step S4, based on the pressure load distribution map, a differential evolution algorithm with neighborhood mutation is applied to calculate the risk assessment parameters for each body region, outputting a pressure ulcer risk score and a development trend prediction, including:

[0104] Step S4.1: Based on the pressure load distribution map and combined with the preset body area pressure ulcer susceptibility weights, construct a multidimensional risk assessment parameter vector including pressure index, duration, and tissue tolerance to form an initial parameter space;

[0105] Step S4.2: Initialize the initial parameter space using Latin hypercube sampling to generate a diverse combination of risk assessment parameter vectors containing 30 individuals;

[0106] Step S4.3: When the differential evolution algorithm gets stuck in a local optimum, a small-range search is performed in the neighborhood space near the current optimal solution based on the diversified risk assessment parameter vector combination to generate an optimized parameter search strategy;

[0107] Step S4.4: Based on the optimized parameter search strategy, implement an adaptive crossover and selection strategy for the diversified risk assessment parameter vector combination, dynamically adjust the crossover probability according to the population diversity, optimize the parameter combination, and generate a new generation of risk assessment parameter vector combination;

[0108] Step S4.5: Calculate the pressure ulcer risk score for each body region using the new generation risk assessment parameter vector combination and estimate the risk change trend in the next 4 hours using a time series prediction model to obtain the pressure ulcer risk score and the development trend prediction.

[0109] In step S4.1, based on the pressure load distribution map output in step S3, a comprehensive risk assessment parameter vector is constructed to provide a foundation for subsequent differential evolution algorithm optimization. First, the following information is collected as the basis for constructing the parameter vector: the pressure load distribution map (an 8×16 risk index matrix obtained from step S3.5, with each element ranging from 0 to 1, representing the preliminary risk assessment of the corresponding location) and the body region pressure ulcer susceptibility weights from medical literature (different body parts have significantly different pressure tolerances, typically including the following weight coefficients: sacrum and coccyx weight 1.5, one of the most prone areas to pressure ulcers; heel weight 1.3, bony prominence area with thin tissue layer; hip weight 1.2, area of ​​concentrated pressure when lying on one's side; scapular region weight 1.1, area of ​​pressure when lying prone; occipital region weight 1.1, head support area; other areas weight 1.0, relatively low risk). Next, the original pressure map is mapped to body regions to determine the body region to which each pixel belongs. This mapping is based on a predefined anatomical template, taking into account the patient's height and position, and is adjusted through affine transformation to match the current patient. Then, a multidimensional risk assessment parameter vector is constructed, containing the following dimensions: pressure intensity parameters. (Based on the risk index in the pressure load distribution map, combined with regional weights:) (where r represents a specific region); duration parameter (This indicates the duration of a specific pressure level, based on continuous monitoring data.) Where I is an indicator function, representing the pressure in region r at time t. When the threshold is exceeded, the value is 1. (Sampling interval); tissue tolerance parameters (Reflects the patient's tissue resistance to pressure, calculated based on patient characteristics:) (where each factor is an adjustment coefficient between 0 and 1); pressure change rate parameter (This reflects the rate of change of pressure over time and is used to assess the recovery capacity of microcirculation.) ); Cumulative risk parameters (Considering the cumulative effect of historical risks, a decaying summation model is adopted:) (where λ is the time decay coefficient). These parameters are combined into a high-dimensional risk assessment parameter vector V:

[0110]

[0111] The subscripts represent different body regions. For the eight main body regions and five parameter categories, a 40-dimensional parameter vector is obtained. Additionally, a set of weight parameters W is included, which adjusts the importance of different parameters in the final risk assessment.

[0112]

[0113] The final risk assessment function can be expressed as a function of the parameter vector V and the weights W: This risk assessment parameter vector encompasses several key factors in pressure ulcer formation, providing an optimization target space for the differential evolution algorithm.

[0114] In step S4.2, a diverse parameter population is initialized to provide a good starting point for the subsequent evolutionary process. The core of the differential evolution algorithm is to maintain a population containing multiple candidate solutions (individuals), each individual representing a set of possible risk assessment parameter combinations. First, the population size and parameter vector dimension are determined: the population size NP is set to 30 individuals, which strikes a balance between computational efficiency and search capability; the parameter vector dimension D is based on the risk assessment parameter vector constructed in step S4.1, including weight coefficients and adjustment parameters, with a total dimension of approximately 20-40. Next, the range of parameter values ​​is defined: weight parameters (such as...) , The range of parameters (e.g., [0, 2]) is typically set to [0, 2], allowing the algorithm to explore the importance of different parameters; the time decay coefficient (λ) is set to [0.01, 0.5] to control the decay rate of historical risk; the threshold parameter is set to a reasonable range based on the specific parameter characteristics. Then, Latin hypercube sampling is used. The method initializes the population. Compared to simple random sampling, LHS ensures a uniform distribution of samples across each dimension, improving the efficiency of parameter space coverage. The basic principle of Latin hypercube sampling is to divide each dimension into NP equally spaced intervals, ensuring that each interval contains exactly one sample point. This allows for efficient parameter space coverage even with high dimensionality, requiring relatively few samples. After initialization, the system evaluates the fitness of each individual. The fitness function, based on a historical dataset, measures the accuracy of risk assessment using this set of parameters. The historical dataset contains patient cases with known pressure ulcer outcomes. Each case includes pressure data and other relevant features, as well as a marker indicating whether the patient developed pressure ulcers within a specific time window. After initialization, all individuals and their fitness values ​​are stored in a population data structure, preparing for subsequent mutation, crossover, and selection operations. Furthermore, the system records the initial population diversity index for monitoring population diversity changes during evolution, preventing premature convergence to local optima. The population initialized using the Latin hypercube sampling method exhibits good diversity and uniformity, providing a high-quality search starting point for differential evolution algorithms.

[0115] In step S4.3, an innovative neighborhood mutation operation is implemented to enhance the algorithm's ability to escape local optima and accelerate the convergence process. The core idea of ​​neighborhood mutation is that when the algorithm might get stuck in a local optimum, instead of performing a completely random global search, it performs a targeted local search within the "neighborhood" of the current optimal solution, balancing global exploration with local exploitation capabilities. First, the basic mutation operation of standard difference evolution is defined. For each individual in the population... Generate test vectors Standard The strategy is: ,in , , These are three distinct individuals randomly selected from the population, where F is the scaling factor (usually in the range of 1000-1000). Within the range). The neighborhood mutation operation enhances the standard mutation in the following ways: neighborhood structure definition (defining the local neighborhood structure for the parameter space, for the current best individual). its neighborhood Defined as: ,in Is X and The distance between them, where r is the neighborhood radius; adaptive neighborhood radius (the neighborhood radius r is not fixed, but dynamically adjusted according to the evolutionary stage: Where t is the current iteration number, T is the maximum iteration number, and α is a parameter controlling the decay rate; Neighborhood mutation triggering mechanism (design an adaptive triggering mechanism that triggers neighborhood mutation when one of the following conditions is met: the improvement of the optimal solution after k consecutive iterations is less than a threshold ε, the population diversity index is lower than a threshold σ, and random probability...) ); Neighborhood mutation operation (when a neighborhood mutation is triggered, for an individual Its test vector The generation method has been modified as follows:

[0116]

[0117] in, It is the best individual in the current population. and From Two individuals randomly selected from the neighborhood of , It is the main scaling factor. It is a noise scaling factor); local search enhancement (under specific conditions, such as the optimal solution remaining unchanged for a long time, in (A more intensive local search is performed in the vicinity). The key advantages of the neighborhood mutation operation are: enhanced local exploitation capability (more concentrated search in promising areas), maintained exploration capability (maintaining population diversity through random perturbation and dynamic neighborhood radius), adaptive behavior (automatically adjusting the search strategy according to the evolutionary stage and population state), and accelerated convergence (focusing on exploring the vicinity of the optimal solution to improve algorithm efficiency). This neighborhood mutation mechanism is particularly suitable for problems such as pressure ulcer risk assessment, which have complex parameter spaces and may have multiple local optima.

[0118] In step S4.4, an adaptive crossover strategy and selection mechanism are implemented to optimize the algorithm's convergence characteristics and population diversity. The crossover operation in the differential evolution algorithm combines information from original and mutated individuals to generate experimental individuals, while the selection operation determines which individuals will enter the next generation. An adaptive crossover strategy is implemented, dynamically adjusting the crossover probability based on population diversity: the population diversity index D is calculated periodically (using the average pairwise distance, the average distance to the population center, or the mean of the parameter standard deviations); the crossover probability is adaptively adjusted (a smaller value is used when diversity is high). To maintain diversity, use larger values ​​when diversity is low. To facilitate information exchange, typical parameter settings are: Individual-level crossover probability (Each individual can have its own crossover probability, adjusted according to its performance in the population; high-performing individuals use a lower crossover probability). Retaining more original information, poorer performers use larger amounts of data. Accept more mutations); perform crossover (using adaptive) The values ​​are subjected to binomial crossover to generate experimental individuals. A comprehensive selection strategy is also implemented, combining elite preservation and tournament selection: an elite preservation mechanism (ensuring that the current best individual is not lost during evolution); and greedy selection (standard...). The adaptive crossover and selection strategy employs several methods: one-to-one selection (comparing the fitness of the original individuals with their corresponding trial individuals); tournament selection (periodically introducing tournament selection to increase population diversity); and diversity protection mechanisms (monitoring population diversity and triggering protective measures when diversity falls below a threshold, such as retaining some optimal individuals and applying random perturbation to the remaining individuals). The advantages of this adaptive crossover and selection strategy are: balancing exploration and development (dynamically adjusting parameters based on population state to achieve optimal balance at different stages of the algorithm), maintaining diversity (preventing premature convergence and loss of population diversity through multiple mechanisms), preserving elites (ensuring no loss of optimal solutions during evolution), and enhancing adaptability (automatically adjusting strategies based on problem characteristics and optimization progress). Through these adaptive strategies, the differential evolution algorithm can more efficiently explore the parameter space for pressure ulcer risk assessment and find more accurate parameter combinations.

[0119] In step S4.5, the optimized parameters are used to accurately assess the patient's pressure ulcer risk and predict future risk trends. The optimized parameters are used to calculate pressure ulcer risk scores for each region of the patient's body: obtaining the current optimized parameters (the optimal individual output by the differential evolution algorithm includes weight coefficients and threshold parameters, which define the importance of different factors in risk assessment); calculating regional risk scores (for each body region r, a multi-factor risk model is used to calculate the risk score, extracting feature data such as pressure index, duration, tissue tolerance, rate of change, and cumulative risk; extracting optimized parameters such as the weights of each feature; calculating the risk score and standardizing it to the 0-100 range); risk level classification (continuous risk scores are divided into different risk levels for easier clinical interpretation); and overall risk score (based on the risk scores of each region, an overall risk score is calculated, considering the dominant role of high-risk regions). Next, based on historical data, the trend of risk scores for each region over the next 4 hours is predicted: Time series data preparation (collecting time series data of risk scores for each region over the past 24 hours, sampling at fixed time intervals such as 15 minutes); trend analysis method selection (selecting appropriate time series prediction methods based on data characteristics, such as linear regression, exponential smoothing, ARIMA models, or recurrent neural networks); risk trend prediction implementation (using Holt exponential smoothing as an example, predicting risk trends using the Holt exponential smoothing method, including initialization, exponential smoothing calculation, and predicting future values); risk trend classification (classifying predicted trends for clinical interpretation and intervention decisions, such as rapid increase, moderate increase, slight increase, stable, slight decrease, moderate decrease, and rapid decrease); risk warning threshold setting (determining whether to trigger a warning based on a comprehensive assessment of risk scores and trends, and adjusting the warning threshold according to the trend). Finally, the risk scores and trend prediction results are integrated to generate a comprehensive risk report, including the overall risk score and risk level, detailed risk scores and trends for each body region, a list of high-risk areas and their risk status, risk warnings requiring immediate intervention, and risk trend predictions for the next 4 hours. This risk report provides healthcare professionals with comprehensive risk assessment information, supporting timely and targeted interventions to improve the effectiveness of pressure ulcer prevention.

[0120] The calculation of the pressure ulcer risk score includes:

[0121] Step S4.5.1: Based on the pressure index, duration, tissue tolerance, pressure change rate and cumulative risk parameters of each body region, use a multi-factor risk model to perform weighted calculation and output the pressure ulcer risk score in the range of 0-100.

[0122] Step S4.5.2: Divide the risk level according to the pressure ulcer risk score: 0-25 points are low risk, 26-50 points are low-to-medium risk, 51-75 points are medium-to-high risk, and 76-100 points are high risk, and generate the risk level classification result.

[0123] In step S4.5.1, a multi-factor risk model is used to calculate pressure ulcer risk scores for each body region. First, weight parameters are extracted from the optimal individual obtained through differential evolution algorithm optimization. These parameters determine the importance of different risk factors in the final score. For each body region r, the following feature data are extracted: pressure index (representing the pressure intensity of the region, already considering the region's susceptibility weight), duration (representing the duration of continuous pressure application), tissue tolerance (representing the tissue's resistance to pressure, determined by factors such as the patient's age, nutritional status, and humidity), pressure change rate (representing the rate of pressure change over time; positive values ​​indicate increased pressure, negative values ​​indicate decreased pressure), and cumulative risk (representing the cumulative effect of historical risk, considering time decay). Then, the system uses a weighted summation model to calculate the risk score:

[0124]

[0125] in, , , , and These are the optimized weight parameters. Note that the tissue tolerance term uses... Because lower tolerance means higher risk; the pressure change rate term uses... This is because only an increase in pressure (positive rate of change) increases risk. The calculated raw risk score is standardized to a range of 0-100 using a linear mapping: ,in The scaling factor is determined based on historical data, ensuring that most scores fall within a meaningful range. The advantages of this multifactor risk model are: it comprehensively considers multiple key factors influencing pressure ulcer formation; the weight parameters are automatically optimized through machine learning rather than being manually set; it can adapt to the characteristics of different patient groups; and the scoring results are intuitive and clinically meaningful.

[0126] In step S4.5.2, the continuous risk scores are divided into different risk levels to facilitate clinical interpretation and intervention decisions. A four-level risk classification standard is adopted: 0-25 points indicates low risk. A score of 26-50 indicates a low likelihood of developing pressure ulcers, usually requiring only routine care and regular monitoring; a score of 26-50 indicates a low to medium risk. A score of 51-75 indicates a certain risk of pressure ulcers, requiring increased monitoring frequency and consideration of preventative measures; a score of 51-75 indicates a medium to high risk. A score of 76-100 indicates a significant risk of pressure ulcers, requiring proactive preventative measures such as increasing the frequency of turning and using pressure-reducing devices; a score of 76-100 indicates a high risk. This indicates a very high likelihood of pressure ulcers, requiring immediate and comprehensive prevention and intervention measures, potentially involving a professional nursing team. Based on the risk score for each body region, a corresponding risk level classification is automatically generated and displayed visually on the user interface using different colors: low-risk areas are shown in green, low-to-medium risk areas in yellow, medium-to-high risk areas in orange, and high-risk areas in red. Furthermore, the overall risk level is calculated, taking into account the weighted average risk of each region and the impact of the highest-risk area.

[0127]

[0128] The weighted average considers both area and importance. The advantages of this risk level classification are: it simplifies complex numerical scoring, enabling healthcare professionals to quickly understand risk status; standardized classification facilitates the development of nursing guidelines and intervention procedures; intuitive color coding improves the efficiency of risk identification; and it considers the balance between overall risk and localized high risk, avoiding the problem of high-risk areas being diluted by averaging. The risk level classification results serve as an important basis for subsequent nursing intervention decisions, guiding healthcare professionals to allocate resources rationally and prioritize high-risk patients and high-risk body areas.

[0129] In step S5, based on the pressure ulcer risk score and the development trend prediction, the optimal airbag inflation / deflation parameters are calculated using an intelligent decision-making algorithm, and an execution command is generated and sent to the airbag control system, including:

[0130] Step S5.1: Based on the pressure ulcer risk score and the development trend prediction, identify high-risk and medium-risk areas and calculate the target value of ideal pressure distribution for each area to determine the target body areas that need pressure redistribution;

[0131] Step S5.2: Establish a transfer function model between airbag pressure and mattress surface pressure; calculate the optimal inflation and deflation parameters for each airbag using a quadratic programming algorithm for the target body area and the target value of the ideal pressure distribution; and generate a control command sequence.

[0132] Step S5.3: According to the control command sequence, the pressure of each airbag is adjusted by controlling the micro air pump and solenoid valve through the PWM signal, and the execution command is sent to the airbag control system.

[0133] In step S5.1, based on pressure ulcer risk scores and development trend predictions, body areas requiring pressure redistribution are identified, and ideal pressure distribution target values ​​are calculated. First, the risk status of each area of ​​the patient's body is classified: areas with a risk score greater than 70 are identified as high-risk areas, areas with a risk score between 40 and 70 are identified as medium-risk areas, and areas with a risk score below 40 are considered low-risk areas. For high-risk and medium-risk areas, their risk trends are further analyzed: if the risk score shows an upward trend (expected to increase by more than 5 points in the next 4 hours), it is marked as "urgently requiring intervention"; if the risk score is stable or slowly increasing (expected to increase by 1-5 points), it is marked as "requiring attention"; if the risk score shows a downward trend, it is marked as "continuous monitoring". Next, the ideal pressure distribution target value for each area is calculated. For high-risk areas, the goal is to significantly reduce surface pressure, with the ideal pressure value set at 50-60% of the current pressure. For medium-risk areas, the goal is to moderately reduce surface pressure, with the ideal pressure value set at 70-80% of the current pressure. For low-risk areas, pressure can be appropriately increased to compensate for the decompression in high-risk areas, but the increase should not exceed 20%, and the absolute pressure value should not exceed the safety threshold (usually 40 mmHg). The overall body support balance requirement is also considered to ensure that pressure redistribution does not lead to patient instability or excessive pressure concentration in localized areas. Furthermore, the system analyzes the pressure gradient between adjacent areas to avoid excessive pressure differences, which could lead to increased shear force and consequently increase the risk of tissue damage. Based on these analyses, an ideal pressure distribution target matrix is ​​generated, representing the target pressure value to be achieved at each location compared to the current actual pressure distribution matrix. This ideal pressure distribution target matrix serves as the basis for subsequent optimization of airbag control parameters, guiding the system on how to adjust the pressure of each airbag to achieve optimal pressure redistribution, reduce pressure load in high-risk areas, and prevent the occurrence and development of pressure ulcers.

[0134] In step S5.2, a transfer function model of the airbag pressure and mattress surface pressure is established, and the optimal airbag control parameters are calculated using a quadratic programming algorithm. The smart mattress contains 64 independently controlled airbag units, corresponding to an 8×16 pressure sensor matrix. Each airbag can individually adjust its inflation volume and internal pressure, thereby changing the support characteristics and pressure distribution of the mattress surface. First, a transfer function model of the pressure inside the airbags and the pressure on the mattress surface is established. This model describes how changes in the internal pressure of the airbags affect the pressure distribution on the mattress surface, considering the interaction between airbags, the elastic properties of the mattress material, and the mechanical response of human tissue. The transfer function can be expressed in matrix form: ,in It is the pressure distribution matrix on the mattress surface. Let F be the internal pressure vector of the airbag, and let F be the nonlinear transfer function. This transfer function is obtained through a combination of experimental data and physical modeling, and then optimized and calibrated using machine learning methods. With the transfer function model, the pressure redistribution problem is transformed into an optimization problem: finding a set of internal pressure values ​​for the airbags such that the actual pressure distribution on the mattress surface is as close as possible to the ideal pressure distribution target value determined in step S5.1. This optimization problem can be expressed as: minimizing... ,in The objective function is the target pressure distribution matrix, while also considering constraints such as upper and lower limits of airbag pressure and the smoothness requirements of adjacent airbag pressures. The system employs a quadratic programming algorithm to solve this optimization problem because the objective function is a quadratic form in terms of airbag pressure, and the constraints can be expressed as linear inequalities. The quadratic programming algorithm can efficiently find the optimal solution that satisfies all constraints, calculating the optimal inflation and deflation parameters for each airbag. The optimization results include the target pressure value, inflation / deflation rate, and sequence for each airbag, forming a complete control command sequence. This control command sequence considers factors such as patient comfort, system response time, and energy efficiency, ensuring a smooth, efficient, and comfortable airbag pressure adjustment process while achieving the best pressure redistribution effect.

[0135] In step S5.3, the system precisely adjusts the pressure of each airbag according to the control command sequence via hardware actuators. The airbag control system of the smart mattress includes a high-precision miniature air pump, an array of solenoid valves, pressure sensors, and control circuitry. The control unit receives the control command sequence generated in step S5.2 and converts it into specific hardware control signals. First, the system determines the target pressure state of each airbag according to the control commands, calculates the difference between the current pressure and the target pressure, and decides whether to inflate, deflate, or keep it unchanged. For airbags that need to be inflated, the corresponding miniature air pump and inlet solenoid valve are activated; for airbags that need to be deflated, the exhaust solenoid valve is opened; for airbags whose pressure is close to the target value, all valves are closed to maintain stable pressure. To achieve precise pressure control, PWM (Pulse Width Modulation) signals are used to control the working state of the air pump and solenoid valves. The duty cycle of the PWM signal determines the output power of the air pump and the opening time of the solenoid valves. By adjusting the duty cycle, the airflow rate can be precisely controlled to achieve smooth pressure adjustment. For example, for airbags requiring a significant increase in pressure, a high duty cycle PWM signal drives the air pump; for airbags requiring fine-tuning, a low duty cycle signal achieves slow and precise adjustments. A closed-loop control mechanism is also implemented, using pressure sensors built into the airbags to monitor the internal pressure in real time. The measured value is compared with the target value, and the PWM signal parameters are dynamically adjusted to ensure the airbag pressure accurately reaches the set value. This closed-loop control strategy compensates for the effects of system delays and changes in airflow resistance, improving control accuracy and response speed. The minimum adjustment accuracy of the entire airbag control system reaches ±1 mmHg, with a response time of less than 2 seconds, enabling rapid response to changes in risk assessment results and timely adjustment of the pressure distribution on the mattress surface. Furthermore, priority sequencing and timing control of airbag pressure adjustment are implemented, ensuring that pressure adjustments in high-risk areas are executed first, while avoiding patient discomfort caused by simultaneous large-scale inflation and deflation of multiple airbags. Through this precise and efficient airbag control mechanism, the system can proactively adjust the pressure distribution on the mattress surface based on pressure ulcer risk assessment results, providing patients with personalized pressure redistribution plans and effectively preventing the occurrence and development of pressure ulcers.

[0136] The simultaneous generation of nursing intervention recommendations includes:

[0137] Step S5.4: Based on the pressure ulcer risk score and the development trend prediction, identify high-risk areas that cannot be completely relieved by airbag adjustment, and determine key care areas that require manual intervention;

[0138] Step S5.5: Combining the patient's historical data and the risk level classification results, use a decision tree algorithm to determine the optimal intervention strategy for the key care areas and generate personalized nursing intervention suggestions;

[0139] Step S5.6: When the pressure ulcer risk score of any area in the key care area exceeds 85 points and the predicted trend of the development trend increases, a turning reminder is automatically generated based on the personalized nursing intervention suggestion and the nursing staff is notified through sound and light prompts and mobile terminal push notifications to complete the nursing intervention notification.

[0140] In step S5.4, high-risk areas that cannot be fully relieved by airbag adjustment are identified, determining key care areas requiring manual intervention. Although the intelligent airbag control system can effectively redistribute pressure on the mattress surface, in some cases, airbag adjustment alone may not be sufficient to reduce pressure in all high-risk areas to a safe level. These situations include: areas where the patient's weight exceeds the safe threshold even with maximum decompression; bony prominences that inevitably form high-pressure points under specific postures (such as the sacrum, coccyx, heels, and hip bones); areas with extremely low tissue tolerance due to specific patient conditions or wounds; and areas where prolonged periods of pressure are sustained due to maintaining the same posture. The system identifies these still high-risk areas by analyzing the expected pressure distribution after airbag adjustment against the safe threshold. Specifically, the expected risk score for each area after optimal airbag adjustment is calculated. If the expected risk score for an area still exceeds 70 points, or if the risk score, although below 70 points, shows a rapid upward trend (expected to increase by more than 10 points within 4 hours), it is marked as a "key care area requiring manual intervention." Furthermore, the duration of risk was considered. Areas with risk scores exceeding the threshold for more than 2 hours were included in the priority care scope, even if the risk score was relatively low, because prolonged continuous pressure is one of the key factors in pressure ulcer formation. These priority care areas were ranked according to risk level to form a priority list to guide subsequent nursing intervention decisions. This combination of machine-assisted and human intervention fully leverages the automatic adjustment capabilities of the intelligent system and the experience judgment of professional nurses, forming complementary advantages and improving the overall effectiveness of pressure ulcer prevention. The identified priority care areas will serve as the basis for generating personalized nursing intervention recommendations in step S5.5, ensuring that nursing resources are focused on the body areas that require the most attention.

[0141] In step S5.5, combining patient historical data and risk level classification results, a decision tree algorithm is used to generate personalized nursing intervention suggestions for key care areas. First, the system collects patient information related to nursing interventions, including: basic information (age, gender, weight, height, etc.), clinical condition (major diseases, mobility, level of consciousness, etc.), pressure ulcer risk assessment results (Braden score dimensions, historical risk trends, etc.), past nursing records (previously effective interventions, patient responses to different intervention methods, etc.), and currently used medical equipment and assistive devices. Then, the decision tree algorithm is used to analyze the relationship between this information and the effectiveness of nursing interventions. Decision trees are an intuitive and efficient machine learning method that can learn "if-then" rules from historical data, making them suitable for scenarios requiring interpretability, such as nursing decisions. The decision tree model is trained based on a large amount of historical nursing records to learn the most effective nursing intervention strategies under different patient characteristics and risk conditions. For each key care area, the decision tree model recommends the most suitable intervention based on the specific situation of that area (risk score, risk trend, anatomical location, etc.) and patient characteristics. These intervention recommendations include, but are not limited to: turning protocols (suggested turning angles, frequencies, and specific procedures), local decompression measures (suggestions for using specific positioning pads, air mattresses, or other decompression devices), skin care protocols (specific instructions for cleaning, moisturizing, and massage), nutritional intervention recommendations (for patients with low tissue tolerance), and monitoring frequency recommendations (suggested intervals for nurses to check specific areas). The generated nursing intervention recommendations are in a structured format, including a detailed description of the intervention, its frequency of implementation, precautions, and expected results, facilitating understanding and execution by nurses. These recommendations are dynamically adjusted based on changes in the patient's condition and feedback on the intervention's effectiveness, ensuring that the interventions remain optimally targeted and effective. Through this machine learning-based personalized nursing intervention recommendation system, expert-level pressure ulcer prevention knowledge can be transformed into specific, actionable nursing guidance, improving the quality and efficiency of nursing care.

[0142] In step S5.6, the system implements proactive early warning and nursing intervention notification functions to ensure that high-risk situations can be handled promptly. It continuously monitors the risk status of all key nursing areas, triggering early warnings and intervention notifications when specific conditions are met. Triggering conditions include: a pressure ulcer risk score exceeding 85 points in any key nursing area (indicating extremely high risk); a risk score not exceeding 85 points but with a significantly rising predicted trend (expected to exceed 85 points within 2 hours); and a risk score remaining at a high-risk level (>75 points) for more than a preset safety time limit (usually 2 hours). When these conditions are triggered, the system automatically generates turning reminders and nursing intervention notifications. The notification content is based on the personalized nursing intervention suggestions generated in step S5.5, including specific risk areas, current risk scores, risk trends, suggested intervention measures, and operational instructions. These notifications are delivered through multiple methods: audio-visual prompts from the bedside control unit (providing intuitive risk level and location indications via LED indicators and buzzers); risk alarms from the nurse station monitoring terminal (displaying patient information requiring intervention and detailed risk data on the central monitoring system); and mobile push notifications (pushing notifications to nurses' handheld devices or smartwatches, ensuring they receive notifications even when not in the ward). A notification priority management and escalation mechanism is also implemented: for extremely high-risk situations (score > 90), a high-priority alarm is issued, and the notification level is automatically escalated if no response is received, ensuring timely handling of emergencies. After receiving a notification and implementing intervention measures, nurses are required to confirm the intervention and record its effects. Nurses can input the implementation status and observed effects of the intervention through mobile terminals or the bedside control unit. This feedback information is recorded in the patient's nursing log and used to evaluate the intervention's effectiveness and optimize future intervention recommendations. Risk status is continuously monitored after intervention; if the intervention fails to effectively reduce the risk, further intervention recommendations or suggestions for evaluation by the medical team will be generated. This closed-loop early warning-intervention-feedback mechanism ensures that high-risk situations are handled promptly and effectively, significantly improving the effectiveness and efficiency of pressure ulcer prevention.

[0143] like Figure 3 As shown, the present invention also provides an intelligent pressure ulcer prevention mattress, comprising: a mattress body having an array of multiple pressure sensors thereon; an airbag unit disposed within the mattress body; an airbag control system for controlling the inflation and deflation of the airbag unit; and a control unit including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the pressure ulcer risk prediction method described above.

[0144] The intelligent pressure-relief mattress features a multi-layered composite structure, consisting of a contact layer, sensor layer, airbag support layer, and base support layer from top to bottom. The contact layer is made of highly elastic medical-grade silicone, offering excellent biocompatibility and comfort. Its honeycomb microstructure design increases the contact area, reduces pressure per unit area, and provides good breathability and heat dissipation, preventing skin damage caused by excessive local temperature and humidity. The contact layer is 10-15mm thick, offering moderate firmness that provides good comfort without excessive envelopment. The sensor layer embeds an 8×16 matrix of high-precision flexible pressure sensors, covering all pressure-bearing areas of the body. Each sensor unit is 2cm×2cm in size and only 0.8mm thick, barely affecting the mattress's comfort. These sensors utilize the piezoresistive principle, measuring from 0-200mmHg with an accuracy of ±1mmHg and a response time of less than 50ms, accurately capturing minute changes in pressure distribution. The sensors are connected via flexible circuit boards, and the signal lines employ a redundant design to ensure that a single point of failure will not affect the overall functionality.

[0145] The airbag support layer is the core of the smart mattress, containing 64 independently controlled airbag units, each corresponding to a pressure sensor matrix. Each airbag is made of medical-grade TPU material, boasting high strength, elasticity, and durability, capable of withstanding repeated inflation and deflation without deformation. The airbags are 4cm x 4cm in size, with an adjustable height ranging from 2-8cm, allowing for different levels of support firmness and height by changing the internal air pressure. The airbags are arranged in a honeycomb pattern, with appropriate gaps between adjacent airbags, ensuring overall support while allowing for independent deformation and improving local adaptability. The bottom of each airbag connects to the air circulation system, with each airbag equipped with an independent air inlet and outlet valve for precise individual control. The base support layer is made of high-density sponge material, 5cm thick, providing basic support and stability while absorbing vibrations to enhance overall comfort. Reinforced edges around the mattress prevent edge sagging from prolonged use, extending the mattress's lifespan.

[0146] The airbag control system is located in a control box on the side of the mattress, and includes a miniature air pump assembly, a solenoid valve array, pressure sensors, and control circuitry. The miniature air pump assembly consists of four independent, silent, oil-free air pumps, each controlling 16 airbags. It has a maximum output pressure of 250 mmHg and a flow rate of 3 L / min, capable of fully deflating and refilling a single airbag within 10 seconds. The solenoid valve array contains 128 miniature solenoid valves (one inlet valve and one outlet valve per airbag), employing a low-power design with static power consumption less than 0.1W and dynamic power consumption less than 0.5W. It supports PWM control for precise airflow adjustment. Each airbag also contains a miniature pressure sensor that monitors the internal pressure in real time, forming a closed-loop control system to ensure the airbag pressure accurately reaches the set value. The control circuitry uses a modular design, including a main control board and multiple slave control boards, communicating via a CAN bus, offering high reliability and scalability. The power consumption of the entire airbag control system is less than 5W in standby mode and does not exceed 50W in maximum working mode. It can be connected to the hospital power supply system through a medical-grade power adapter. It is also equipped with a built-in lithium battery, which can maintain normal operation of the system for more than 4 hours in the event of a power outage.

[0147] The control unit is the brain of the entire intelligent pressure ulcer prevention mattress, responsible for data processing, risk assessment, and control decisions. The control unit employs an embedded system design, including a high-performance processor, large-capacity memory, and multiple communication interfaces. The processor uses a quad-core ARM Cortex-A72 architecture with a clock speed of 1.5GHz, equipped with a dedicated AI acceleration unit to support accelerated neural network inference, enabling efficient execution of pressure ulcer risk prediction algorithms. Memory includes 4GB LPDDR4 RAM and 64GB eMMC flash memory, providing sufficient operating space and data storage capacity to store over one year of historical pressure data and risk assessment records. Communication interfaces include Wi-Fi, Bluetooth, Ethernet, and USB, supporting seamless connectivity with hospital information systems, nurse station monitoring terminals, and mobile devices. The control unit also features a 7-inch touchscreen display with a resolution of 1024×600, displaying real-time pressure distribution heatmaps, risk assessment results, and system status, allowing caregivers to operate and view data directly at the bedside. The system software is developed based on the Linux operating system and adopts a modular architecture, including a data acquisition module, a feature extraction module, a risk prediction module, an airbag control module, and a user interface module. The software implements the pressure ulcer risk prediction method of the present invention, including core functions such as pressure data preprocessing, feature selection, neural network model inference, differential evolution algorithm optimization, and intelligent decision-making algorithm.

[0148] The workflow of the intelligent pressure ulcer prevention mattress is as follows: First, a distributed pressure sensor array collects pressure distribution data on the mattress surface in real time, with a sampling frequency of 10Hz, and transmits the data to the control unit via Bluetooth Low Energy. The control unit filters and standardizes the raw pressure data to generate a standardized pressure distribution matrix. Then, the system uses a binary random forward algorithm to extract key features from the pressure data and fuses these features with patient vital signs (obtained from the hospital information system or manually input). Next, an enhanced convolutional neural network model processes this fused data to generate a pressure load distribution map, visually displaying the risk index of each area. Based on this map, the system uses a differential evolution algorithm with neighborhood mutation to calculate risk assessment parameters for each body area, outputting a pressure ulcer risk score and a prediction of the development trend over the next 4 hours. Finally, an intelligent decision-making algorithm calculates the optimal airbag inflation and deflation parameters based on the risk assessment results, adjusting the pressure of each airbag through the airbag control system to achieve precise redistribution of pressure on the mattress surface and reduce pressure load in high-risk areas. Simultaneously, the system generates personalized nursing intervention suggestions and, when necessary, sends turning reminders to caregivers via audio-visual prompts and mobile terminal push notifications.

[0149] The intelligent pressure ulcer prevention mattress features multiple safety safeguards: an overpressure protection mechanism ensures the pressure inside the airbags does not exceed a safe threshold; a power outage protection function automatically locks the current airbag status in the event of a power failure, preventing sudden deflation from causing patient discomfort; a fault self-diagnosis system detects the operating status of sensors, air pumps, and solenoid valves, promptly reporting any abnormalities; and a data backup mechanism regularly synchronizes important data to the cloud or local backup devices to prevent data loss. Furthermore, the system supports remote monitoring and maintenance, allowing technicians to securely connect and remotely diagnose system problems, update software, and adjust parameters to ensure the system is always in optimal condition.

[0150] Compared to traditional pressure ulcer mattresses, the intelligent pressure ulcer mattress of this invention offers significant advantages: it shifts from passive prevention to proactive prediction, enabling the identification of high-risk areas and intervention before pressure ulcers develop; personalized pressure adjustment programs consider individual patient differences and real-time changes in their condition, providing precise pressure management; an intelligent nursing reminder system reduces the workload of caregivers and improves nursing efficiency; and comprehensive data recording and analysis functions support long-term risk management and nursing quality assessment. These advantages enable the intelligent pressure ulcer mattress of this invention to significantly reduce the incidence of pressure ulcers, alleviate patient suffering, lower medical costs, and improve the quality of care.

[0151] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0152] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the patent protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting pressure ulcer risk, characterized in that, The pressure ulcer risk prediction method includes the following steps: Acquire pressure distribution data collected by a distributed pressure sensor array, and perform signal filtering and standardization processing on the pressure distribution data to obtain a standardized pressure distribution matrix; Based on the standardized pressure distribution matrix, a binary random forward algorithm is applied to select features and output the optimal feature subset. The patient's vital signs parameters are obtained, and the patient's vital signs parameters are fused with the optimal feature subset. The system is then trained using an enhanced convolutional neural network with an activation function to generate a pressure load distribution map. Based on the pressure load distribution map, the differential evolution algorithm with neighborhood mutation is applied to calculate the risk assessment parameters of each body region, and output the pressure ulcer risk score and development trend prediction. Based on the pressure ulcer risk score and the development trend prediction, the optimal airbag inflation and deflation parameters are calculated by an intelligent decision-making algorithm, and execution instructions are generated to the airbag control system. At the same time, nursing intervention suggestions are generated. Specifically, the step of calculating risk assessment parameters for each body region based on the pressure load distribution map and using a differential evolution algorithm with neighborhood mutations to output a pressure ulcer risk score and development trend prediction includes: Based on the pressure load distribution map and combined with the preset body area pressure ulcer susceptibility weights, a multidimensional risk assessment parameter vector including pressure index, duration, and tissue tolerance is constructed to form an initial parameter space. Latin hypercube sampling was used to initialize the initial parameter space, generating a diverse combination of risk assessment parameter vectors containing 30 individuals; When the differential evolution algorithm gets stuck in a local optimum, it performs a small-range search in the neighborhood space near the current optimal solution based on the diverse risk assessment parameter vector combination to generate an optimized parameter search strategy. Based on the optimized parameter search strategy, an adaptive crossover and selection strategy is implemented on the diversified risk assessment parameter vector combination. The crossover probability is dynamically adjusted according to the population diversity to optimize the parameter combination and generate a new generation of risk assessment parameter vector combination. The pressure ulcer risk score is calculated for each body region using the new generation of risk assessment parameter vector combination, and the risk change trend is estimated in the next 4 hours using a time series prediction model to obtain the pressure ulcer risk score and the development trend prediction.

2. The method according to claim 1, characterized in that, The step of performing signal filtering and standardization on the pressure distribution data to obtain a standardized pressure distribution matrix includes: Based on the pressure distribution data, noise and outliers are removed by combining medium-value filtering and Kalman filtering to generate cleaned pressure data. The pressure data after cleaning is subjected to min-max normalization to map the values ​​to the [0,1] interval to obtain standardized data. Based on the sensor spatial arrangement information and the standardized data, the two-dimensional pressure distribution matrix is ​​reconstructed to obtain the standardized pressure distribution matrix.

3. The method according to claim 1, characterized in that, Based on the standardized pressure distribution matrix, a binary random forward algorithm is applied for feature selection to output the optimal feature subset, including: Based on the standardized pressure distribution matrix, a multi-dimensional feature pool containing statistical features, morphological features, and temporal features is constructed to form an initial feature set. The initial feature set is represented by binary encoding to construct a binary representation model for feature selection; The optimal feature subset is obtained by iteratively selecting the optimal feature combination from the binary representation model using a stochastic forward algorithm.

4. The method according to claim 1, characterized in that, The process of training an enhanced convolutional neural network with activation functions to generate a pressure load distribution map includes: The optimal feature subset is fused with the patient's vital signs parameters using multimodal data to construct an input tensor, forming the network input data; The design incorporates a deep neural network architecture with 5 convolutional layers. The first 3 layers use 3×3 convolutional kernels, and the last 2 layers use 5×5 convolutional kernels. The network performs multi-scale feature extraction on the input data and outputs multi-level feature representations. Based on the multi-level feature representation, an 8×16 feature map is generated through the last convolutional layer to obtain the pressure load distribution map.

5. The method according to claim 4, characterized in that, The activation function α and β are learnable parameters that are automatically adjusted through backpropagation during training.

6. The method according to claim 1, characterized in that, The calculation of the pressure ulcer risk score includes: Based on the pressure index, duration, tissue tolerance, pressure change rate, and cumulative risk parameters of each body region, a multi-factor risk model is used for weighted calculation, and the pressure ulcer risk score in the range of 0-100 is output. Based on the pressure ulcer risk score, risk levels are divided as follows: 0-25 points are low risk, 26-50 points are low to medium risk, 51-75 points are medium to high risk, and 76-100 points are high risk, generating risk level classification results.

7. The method according to claim 1, characterized in that, The step of calculating the optimal airbag inflation / deflation parameters using an intelligent decision-making algorithm based on the pressure ulcer risk score and the predicted development trend, and generating execution commands to the airbag control system, includes: Based on the pressure ulcer risk score and the development trend prediction, high-risk and medium-risk areas are identified and the target value of ideal pressure distribution for each area is calculated to determine the target body areas that need pressure redistribution. A transfer function model of airbag pressure and mattress surface pressure is established. The optimal inflation and deflation parameters of each airbag are calculated using a quadratic programming algorithm for the target body area and the target value of ideal pressure distribution, and a control command sequence is generated. According to the control command sequence, the pressure of each airbag is adjusted by controlling the micro air pump and solenoid valve through PWM signal, and the execution command is sent to the airbag control system.

8. The method according to claim 6, characterized in that, The simultaneous generation of nursing intervention recommendations includes: Based on the pressure ulcer risk score and the development trend prediction, high-risk areas that cannot be completely relieved by airbag adjustment are identified, and key care areas requiring manual intervention are determined. By combining patient historical data and the risk level classification results, a decision tree algorithm is used to determine the optimal intervention strategy for the key care areas, generating personalized nursing intervention recommendations. When the pressure ulcer risk score of any area in the key care area exceeds 85 points and the predicted trend of the development trend increases, a turning reminder is automatically generated based on the personalized nursing intervention suggestion and notified to the nursing staff through sound and light prompts and mobile terminal push notifications to complete the nursing intervention notification.

9. A smart anti-pressure sore mattress, characterized in that, include: The mattress body has multiple pressure sensors arrayed thereon; an airbag unit is disposed within the mattress body; an airbag control system is used to control the inflation and deflation of the airbag unit; and a control unit includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the pressure ulcer risk prediction method as described in any one of claims 1-8.

Images