Pressure damage dynamic risk assessment method based on multi-modal sensing

Abstract

According to the pressure damage dynamic risk assessment method based on multi-modal sensing, multi-physical field data is collected in real time through an MEMS piezoresistive array, a near infrared spectrum sensor and a flexible capacitance type humidity sensor, and early precise early warning and automatic grading intervention are achieved in combination with an LSTM neural network and dynamic threshold adjustment. According to the method, the early warning time of the stress injury is advanced by 4-6 hours, so that the clinical misjudgment rate is reduced to be less than 15% from 38%, and the problems that in the prior art, poor blood circulation of subcutaneous tissues cannot be accurately identified by traditional visual inspection, the hysteresis of the currently adopted preventive measures and the clinical accurate judgment error rate are relatively high, and the clinical accuracy is relatively low are solved. In order to solve the problem that the occurrence risk and the compression degree of the stress injury cannot be accurately judged due to the fact that the existing tissue ischemia assessment method is based on the three-layer architecture of'multi-modal sensing and dynamic modeling combined with intelligent closed loop ', the transformation from'passive discovery' to'active prevention 'is realized, and a more comprehensive tissue ischemia assessment system is constructed.

Description

Technical Field

[0001] The present application relates to the field of medical device technology, and in particular to a dynamic risk assessment method for pressure injuries based on multimodal sensing. Background Art

[0002] Pressure injury (PI), formerly known as pressure ulcers, is a localized injury caused by prolonged pressure, shear, or friction on the skin and subcutaneous tissue, leading to localized tissue ischemia, hypoxia, and cellular metabolic disorders. This complication is particularly common in patients who are bedridden, immobilized, or wheelchair-bound for extended periods. The core pathological mechanism is that sustained mechanical pressure exceeds the capillary closure pressure (approximately 32 mmHg, or 4.3 kPa), resulting in interruption of local tissue blood flow, which in turn triggers ischemia-reperfusion injury and an inflammatory cascade.

[0003] Statistics show that the global incidence of hospital-acquired pressure injuries ranges from 5% to 15%, with rates exceeding 30% in intensive care unit (ICU) patients and a lifetime incidence as high as 25% to 85% in spinal cord injury patients. According to the international staging criteria (NPUAP 2016), these injuries are classified as Stage I (epidermal erythema), Stage II (dermal damage), Stage III (full-thickness skin loss), Stage IV (tissue necrosis involving musculoskeletal tissue), and unstageable, deep tissue injuries. Pressure injuries are most common over bony prominences such as the sacrum, ischial tuberosity, heels, and scapulae. These areas have thin subcutaneous fat and poor buffering capacity, which can easily lead to stress concentration when the body is fixed. During the pathological process, deep muscle tissue is less tolerant of ischemia than the epidermis, resulting in the common phenomenon of "mild surface damage while deep tissue necrosis"—a key factor in the difficulty of early clinical detection.

[0004] In current clinical practice, the prevention and monitoring technology for pressure injuries primarily relies on physical intervention and manual observation. Specifically, periodic turning (Q2H) care, which periodically changes the patient's position to temporarily relieve local tissue pressure, is the most fundamental clinical preventive measure. Static pressure relief devices, such as air mattresses and foam dressings, disperse the pressure load through elastic deformation of the material, reducing contact surface pressure. Meanwhile, nurses assess skin condition based on visual observation of skin color changes (such as erythema reactions), palpation of skin temperature and firmness, and grading of pressure ulcers using pressure ulcer risk assessment scales (such as the Braden Scale). From a technical perspective, existing monitoring methods primarily rely on subjective judgment by caregivers and lack quantitative monitoring equipment. For example, traditional pressure distribution monitoring can only obtain static pressure data through disposable consumables such as pressure-sensing paper and cannot achieve dynamic and continuous monitoring. Regarding skin microenvironment management, although some humidity sensors are used in wound care, they are not integrated into daily care products, making it difficult to track the duration of epidermal moisture exposure in real time. Traditional air mattresses disperse pressure only through alternating inflation and lack the ability to actively regulate the local microenvironment, such as humidity and temperature. Furthermore, the design of existing pressure relief devices generally lacks precise ergonomic adaptation. For example, the pressure dispersion effect in high-risk areas such as the sacrum and greater trochanter is uneven, and the material's insufficient breathability can easily lead to skin immersion, further increasing the risk of injury. Therefore, how to accurately determine the risk and extent of pressure injuries while also preventing or improving them has become a technical challenge that urgently needs to be addressed in the current process of preventing and monitoring pressure injuries. Summary of the Invention

[0005] The present application provides a dynamic risk assessment method for pressure injuries based on multimodal sensing to address the problem in the prior art that traditional visual observation cannot accurately identify poor blood circulation in subcutaneous tissue, the lag in the preventive measures currently taken and the high error rate in clinical precision judgment, making it impossible to accurately judge the risk of pressure injuries and the degree of pressure.

[0006] The present application provides a method for dynamic risk assessment of pressure injuries based on multimodal sensing, the method comprising:

[0007] MEMS piezoresistive arrays, near-infrared spectral sensors, and flexible capacitive humidity sensors are used to collect real-time pressure data, tissue oxygenation status, and skin surface temperature and humidity data;

[0008] The collected data is cleaned and calculated to obtain the pressure gradient, oxygenation drop rate and humidity time integral characteristics;

[0009] A dynamic risk scoring model is constructed based on an LSTM neural network, the features are input and the threshold is adjusted in combination with the patient's physiological parameters to calculate the dynamic risk score;

[0010] A risk level classification warning is performed according to the dynamic risk score, and a body position adjustment prompt message is sent when the risk level is low risk, or an alarm message is sent when the risk level is high risk.

[0011] In some possible implementations, performing graded warnings based on the dynamic risk score includes:

[0012] If the dynamic risk score is higher than 80 points, the risk level is confirmed as high risk and an alarm message is sent to prompt interventional treatment;

[0013] If the dynamic risk score is less than 80 points, determine the number of influencing factors; wherein the influencing factors are determined based on temperature data, humidity data, and pressure data;

[0014] If the number of influencing factors is greater than 2, the risk level is confirmed to be high risk, and an alarm message is sent to prompt interventional treatment.

[0015] In some possible implementations, determining the number of impact factors includes:

[0016] Calculate the temperature change value based on the current temperature and the previous temperature in the temperature data;

[0017] Determine the humidity value of the current humidity data and the corresponding duration;

[0018] Constructing a pressure-time integral to calculate a first integral value;

[0019] The temperature change value, the humidity value, and the first integral value are respectively compared with their corresponding preset thresholds to obtain the number of influencing factors; wherein the number exceeding the preset threshold is the number of influencing factors.

[0020] In some possible implementations, the method further includes:

[0021] Determining whether the temperature change value exceeds a second temperature threshold;

[0022] Determining whether the first integral value exceeds a second integral threshold;

[0023] If the temperature change value does not exceed the second temperature threshold, the first integral value does not exceed the second integral threshold, and the number of impact factors is 1, then the risk level is determined to be low risk;

[0024] If the temperature change value does not exceed the second temperature threshold, the first integral value does not exceed the second integral threshold, and the number of influencing factors is 2, then the risk level is determined to be medium risk;

[0025] If the temperature change value exceeds a second temperature threshold, and / or the first integral value exceeds a second integral threshold, the risk level is determined to be high risk.

[0026] In some possible implementations, the method further includes:

[0027] When the risk level is low risk or medium risk, if the first integral value exceeds a first integral threshold, an automatic decompression function is activated to adjust the mattress to a 30° tilt;

[0028] Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, run the automatic decompression function for the first duration; otherwise, send an alarm message to prompt interventional treatment.

[0029] In some possible implementations, the method further includes:

[0030] When the risk level is low risk or medium risk, if the temperature change value exceeds a first temperature threshold, a local cooling function is activated to apply a semiconductor cooling patch;

[0031] Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, apply the semiconductor cooling patch for the second time period; otherwise, send an alarm message to prompt interventional treatment.

[0032] In some possible implementations, the method further includes:

[0033] When the risk level is low risk or medium risk, if the humidity value exceeds a first humidity threshold and lasts for more than a preset time, start the fan;

[0034] Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, run the fan for a third time period; otherwise, send an alarm message to prompt interventional treatment.

[0035] In some possible implementations, a dynamic risk scoring model is constructed based on an LSTM neural network, and the features are input and the thresholds are adjusted in combination with the patient's physiological parameters, including:

[0036] Z-normalization was performed on the pressure gradient, oxygenation decrease rate, and humidity time integral;

[0037] The patient's physiological parameters are converted into normalized feature vectors and mapped to the same dimensional space as the time series features through a fully connected layer;

[0038] A time sliding window is used to segment the multi-physics field data to generate a time series feature matrix; the multi-physics field data includes pressure data, tissue oxygenation status, and skin surface temperature and humidity data;

[0039] The input layer of the LSTM neural network receives a time series feature matrix, wherein the pressure gradient, oxygenation drop rate, and humidity time integral are processed by a time series input branch, and the physiological parameters are processed by a parallel static feature branch;

[0040] Thresholds are set based on the patient's physiological parameters, and correlation analysis is performed on the pressure gradient, oxygenation decrease rate, humidity time integral and historical pressure injury data to obtain the corresponding first correlation weight, second correlation weight and third correlation weight. The dynamic risk scoring model is reconstructed based on the first correlation weight, second correlation weight and third correlation weight.

[0041] In some possible implementations, the method further includes:

[0042] A near-infrared spectral sensor with a wavelength of 650-950nm is used to continuously monitor tissue oxygen saturation StO2 and oxygen metabolism rate. An early warning is activated when StO2 is less than 40% for 2 hours.

[0043] In some possible implementations, the method further includes:

[0044] When an early warning is triggered, the alarm time, pressure location and corresponding treatment measures are automatically extracted to generate a structured nursing record;

[0045] A federated learning mechanism is used to update risk model parameters and upload them to the server.

[0046] From the above content, it can be seen that the present application provides a dynamic risk assessment method for pressure injuries based on multimodal sensing, which includes using MEMS piezoresistive arrays, near-infrared spectral sensors and flexible capacitive humidity sensors to collect pressure data, tissue oxygenation status, skin surface temperature and humidity data in real time; performing signal cleaning and calculation on the collected data to obtain pressure gradient, oxygenation decrease rate and humidity time integral characteristics; constructing a dynamic risk scoring model based on an LSTM neural network, inputting the characteristics and adjusting the threshold in combination with the patient's physiological parameters to calculate the dynamic risk score; executing risk level classification warning according to the dynamic risk score, sending a posture adjustment prompt when the risk level is low risk, or sending an alarm message when the risk level is high risk. The present application collects multi-physical field data in real time through a MEMS piezoresistive array, a near-infrared spectral sensor and a flexible capacitive humidity sensor, and combines the LSTM neural network with dynamic threshold adjustment to achieve early accurate warning and automated graded intervention. This method advances the warning time of pressure injuries by 4-6 hours, reducing the clinical misjudgment rate from 38% to <15%. It solves the problem in existing technologies that traditional naked eye observation cannot accurately identify poor blood circulation in subcutaneous tissue, the lag in taking preventive measures and the high error rate in clinical accurate judgment, making it impossible to accurately judge the risk and extent of pressure injuries. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0048] Figure 1 Flowchart of the dynamic risk assessment method for pressure injuries based on multimodal sensing provided in this application;

[0049] Figure 2 This is an impact factor analysis diagram in the examples provided in this application;

[0050] Figure 3 This is a flowchart of the hierarchical intervention in the examples provided in this application. DETAILED DESCRIPTION

[0051] The embodiments described in the following examples do not represent all embodiments consistent with the present application, but are merely examples of systems and methods consistent with some aspects of the present application as detailed in the claims.

[0052] Pressure injury (PI), formerly known as pressure ulcers, is a localized injury caused by prolonged pressure, shear, or friction on the skin and subcutaneous tissue, leading to localized tissue ischemia, hypoxia, and cellular metabolic disorders. This complication is particularly common in patients who are bedridden, immobilized, or wheelchair-bound for extended periods. The core pathological mechanism is that sustained mechanical pressure exceeds the capillary closure pressure (approximately 32 mmHg, or 4.3 kPa), resulting in interruption of local tissue blood flow, which in turn triggers ischemia-reperfusion injury and an inflammatory cascade.

[0053] Statistics show that the global incidence of hospital-acquired pressure injuries ranges from 5% to 15%, with rates exceeding 30% in intensive care unit (ICU) patients and a lifetime incidence as high as 25% to 85% in spinal cord injury patients. According to the international staging criteria (NPUAP 2016), these injuries are classified as Stage I (epidermal erythema), Stage II (dermal damage), Stage III (full-thickness skin loss), Stage IV (tissue necrosis involving musculoskeletal tissue), and unstageable, deep tissue injuries. Pressure injuries are most common over bony prominences such as the sacrum, ischial tuberosity, heels, and scapulae. These areas have thin subcutaneous fat and poor buffering capacity, which can easily lead to stress concentration when the body is fixed. During the pathological process, deep muscle tissue is less tolerant of ischemia than the epidermis, resulting in the common phenomenon of "mild surface damage while deep tissue necrosis"—a key factor in the difficulty of early clinical detection.

[0054] In current clinical practice, the prevention and monitoring technology for pressure injuries primarily relies on physical intervention and manual observation. Specifically, periodic turning (Q2H) care, which periodically changes the patient's position to temporarily relieve local tissue pressure, is the most fundamental clinical preventive measure. Static pressure relief devices, such as air mattresses and foam dressings, disperse the pressure load through elastic deformation of the material, reducing contact surface pressure. Meanwhile, nurses assess skin condition based on visual observation of skin color changes (such as erythema reactions), palpation of skin temperature and firmness, and grading of pressure ulcers using pressure ulcer risk assessment scales (such as the Braden Scale). From a technical perspective, existing monitoring methods primarily rely on subjective judgment by caregivers and lack quantitative monitoring equipment. For example, traditional pressure distribution monitoring can only obtain static pressure data through disposable consumables such as pressure-sensing paper and cannot achieve dynamic and continuous monitoring. Regarding skin microenvironment management, although some humidity sensors are used in wound care, they are not integrated into daily care products, making it difficult to track the duration of epidermal moisture exposure in real time. Traditional air mattresses disperse pressure only through alternating inflation and lack the ability to actively regulate the local microenvironment, such as humidity and temperature. Furthermore, the design of existing pressure relief devices generally lacks precise ergonomic adaptation. For example, the pressure dispersion effect in high-risk areas such as the sacrum and greater trochanter is uneven, and the material's insufficient breathability can easily lead to skin immersion, further increasing the risk of injury. Therefore, how to accurately determine the risk and extent of pressure injuries while also preventing or improving them has become a technical challenge that urgently needs to be addressed in the current process of preventing and monitoring pressure injuries.

[0055] To address the existing challenges of traditional visual observation, which cannot accurately identify poor subcutaneous blood circulation, resulting in a delay in implementing preventive measures and a high error rate in clinically accurate judgments, and thus preventing the accurate assessment of the risk and extent of pressure injuries, this application provides a dynamic risk assessment method for pressure injuries based on multimodal sensing. This method utilizes a MEMS piezoresistive array, a near-infrared spectral sensor, and a flexible capacitive humidity sensor to collect multi-physics field data in real time. Combined with an LSTM neural network and dynamic threshold adjustment, this method achieves early, accurate warning and automated, graded intervention. This method advances pressure injury warning time by 4-6 hours and reduces the clinical misjudgment rate from 38% to less than 15%. This method addresses the existing challenges of traditional visual observation, which cannot accurately identify poor subcutaneous blood circulation, resulting in a delay in implementing preventive measures and a high error rate in clinically accurate judgments, thus preventing the accurate assessment of the risk and extent of pressure injuries. Through a three-layer architecture of "multimodal sensing + dynamic modeling + intelligent closed loop," this method achieves a shift from passive detection to active prevention, building a more comprehensive tissue ischemia assessment system.

[0056] In some embodiments, as Figure 1As shown, the present application provides a dynamic risk assessment method for pressure injuries based on multimodal sensing, the method comprising:

[0057] MEMS piezoresistive arrays, near-infrared spectral sensors, and flexible capacitive humidity sensors are used to collect real-time pressure data, tissue oxygenation status, and skin surface temperature and humidity data;

[0058] The collected data is cleaned and calculated to obtain the pressure gradient, oxygenation drop rate and humidity time integral characteristics;

[0059] A dynamic risk scoring model is constructed based on an LSTM neural network, the features are input and the threshold is adjusted in combination with the patient's physiological parameters to calculate the dynamic risk score;

[0060] A risk level classification warning is performed according to the dynamic risk score, and a body position adjustment prompt message is sent when the risk level is low risk, or an alarm message is sent when the risk level is high risk.

[0061] This application uses MEMS piezoresistive arrays, near-infrared spectral sensors, and flexible capacitive humidity sensors to collect multi-physical field data in real time, and combines LSTM neural networks with dynamic threshold adjustment to achieve early and accurate warning and automated graded intervention.

[0062] This application uses a MEMS piezoresistive array (high-frequency sampling) to monitor pressure distribution and gradient changes in real time, and can accurately locate pressure concentration areas in high-risk areas such as bony protuberances (such as the sacral pressure peak). Compared with static monitoring methods such as traditional pressure sensing paper, it achieves dynamic and continuous pressure load tracking.

[0063] The near-infrared spectral sensor (wavelength 650-950nm) penetrates the skin to a depth of 3mm and directly monitors tissue oxygen saturation (StO2) and oxygen metabolism rate. When StO2 is less than 40% for 2 hours, an early warning can be activated. Compared with traditional naked eye observation of epidermal erythema (Stage I), deep tissue ischemia can be detected 4-6 hours earlier, solving the clinical misjudgment problem of "normal surface but deep necrosis".

[0064] The flexible capacitive humidity sensor tracks the skin surface humidity (RH) and duration in real time.

[0065] An immersion warning is triggered when 80% RH persists for 20 minutes, preventing skin barrier damage caused by delayed subjective judgment. By simultaneously collecting multi-dimensional data such as pressure, oxygenation, temperature, and humidity, a full-chain risk assessment of "mechanical pressure-tissue hypoxia-microenvironment deterioration" is formed. Compared with single-parameter monitoring (such as measuring only pressure or humidity), the misjudgment rate is reduced from 38% to <15%. Clinical data show that the incidence of pressure injuries in ICU patients has been reduced by 58%.

[0066] After signal cleaning of the collected data, quantitative features such as pressure gradient (rate of change of pressure per unit space), oxygenation decline rate (dStO2 / dt), and humidity time integral (∫RH·dt) are calculated to convert abstract physiological risks into calculable indicators.

[0067] A dynamic risk scoring model was constructed based on the LSTM neural network. Dimensional differences were eliminated through Z-normalization, and a time sliding window was used to capture dynamic temporal correlations (such as the cumulative effect of pressure-oxygenation). The model's AUC (area under the curve) was improved from 0.62 of the traditional method to 0.85, and the ability to focus on risks in vulnerable areas was improved by 45%.

[0068] By mapping physiological parameters such as patient BMI and HbA1c into a time-series feature space, a dual-branch architecture (dynamic time series combined with static physiology) enables joint modeling of "dynamic risk changes and individual tolerance differences." For example, the ischemic sensitivity recognition weight for diabetic patients (HbA1c ≥ 8) is automatically increased by 27%, avoiding misjudgments caused by traditional fixed thresholds (such as 32 mmHg), and improving the model's adaptability to individual differences by 60%.

[0069] For low- and medium-risk scenarios, the system dynamically adjusts interventions based on real-time risk scores. For example, after initiating automatic decompression, if the risk level decreases, intervention is maintained (e.g., the mattress tilts for two hours); otherwise, an alarm is immediately escalated. Clinical data shows that this closed-loop mechanism has increased the effectiveness of interventions to 78% and improved nursing resource utilization by 40%, avoiding the waste of resources caused by "one-size-fits-all" interventions.

[0070] In some embodiments, performing graded warning according to the dynamic risk score includes:

[0071] If the dynamic risk score is higher than 80 points, the risk level is confirmed as high risk and an alarm message is sent to prompt interventional treatment;

[0072] If the dynamic risk score is less than 80 points, determine the number of influencing factors; wherein the influencing factors are determined based on temperature data, humidity data, and pressure data;

[0073] If the number of influencing factors is greater than 2, the risk level is confirmed to be high risk, and an alarm message is sent to prompt interventional treatment.

[0074] In this embodiment, 80 points is set as the scoring threshold of the dynamic risk score. The dynamic risk score is obtained by combining the dynamic weighted calculation of multiple physical field characteristics such as pressure gradient and oxygenation decrease rate. In this application, the dynamic risk is calculated by multimodal data fusion, so that the sensitivity of high-risk judgment is increased to 92%, avoiding misjudgment caused by fluctuations in a single parameter (such as a short-term increase in pressure but normal oxygenation to avoid false alarms). At the same time, when the dynamic score does not reach the threshold, by counting the number of influencing factors that exceed the preset threshold in the temperature, humidity, and pressure data (such as temperature change > 1.5℃ / h, humidity > 80% RH for 30min, PTI > 1500kPa·min, etc.), it can effectively cover potential risk scenarios of multi-parameter coordinated abnormalities (such as slight pressure exceeding the limit, continuous high humidity or local temperature rise). Clinical data demonstrate that this dynamic risk score, combined with a dual-judgment mechanism based on multiple influencing factors, has reduced the missed diagnosis rate for high-risk events from 9% to less than 3%, shortening the high-risk response time for ICU patients from an average of 47 minutes to 8 minutes. Furthermore, with interventional treatment by medical staff, the rate of pressure injuries progressing to Stage II or above has decreased by 68%, significantly improving the timeliness and safety of care for critically ill patients. Furthermore, this mechanism replaces subjective judgment with quantitative indicators, reducing the consistency error of risk assessment from ±30% to ±8%, providing a more objective basis for clinical decision-making.

[0075] In some embodiments, determining the number of impact factors includes:

[0076] Calculate the temperature change value based on the current temperature and the previous temperature in the temperature data;

[0077] Determine the humidity value of the current humidity data and the corresponding duration;

[0078] Constructing a pressure-time integral to calculate a first integral value;

[0079] The temperature change value, the humidity value, and the first integral value are respectively compared with their corresponding preset thresholds to obtain the number of influencing factors; wherein the number exceeding the preset threshold is the number of influencing factors.

[0080] In this application, the number of influencing factors is determined by calculating the temperature change value (the difference between the current temperature and the previous temperature), the humidity value and the duration, and constructing the pressure-time integral (PTI). By converting the abstract physiological risk into a quantifiable influencing factor, when any parameter exceeds the preset threshold, it is counted as one influencing factor. This quantitative analysis breaks through the subjectivity of traditional visual observation and can significantly reduce the error in risk assessment. For example, in this embodiment, the temperature change value, the humidity value, and the first integral value correspond to their preset thresholds: the first temperature threshold, the second temperature threshold, the first humidity threshold, the second humidity threshold, the first integral value, and the second integral value, respectively. In this embodiment, the first temperature threshold is 1.5°C, the second temperature threshold is 2°C, the first integral value is 1500kPa·min, the second integral value is 3000kPa·min, the first humidity threshold is 80%RH, and the second humidity threshold is 85%RH. When the temperature change value is greater than 1.5℃ / h, it prompts an early inflammatory response, and can detect tissue abnormalities 2-3 hours earlier than traditional palpation; when the humidity is greater than 80% RH for 30 minutes, it triggers an immersion warning, avoiding skin damage caused by delayed subjective judgment; when the PTI is greater than 1500kPa·min, it quantifies the cumulative effect of pressure, which is more in line with individual tolerance differences than a fixed pressure threshold.

[0081] The pressure-time integral (PTI) formula in this application is as follows:

[0082]

[0083] This application can quickly identify high-risk scenarios with multi-parameter coordinated abnormalities by counting the number of influencing factors that exceed the threshold. When the number of influencing factors is greater than 2, even if a single parameter does not reach the severe threshold, an early warning can be triggered through multi-factor coupling analysis, reducing the missed diagnosis rate from 12% to below 5%. Clinical data show that this mechanism increases the early intervention rate of pressure injuries by 73%, switching from passive treatment to active prevention, and combined with subsequent graded early warning measures, it can reduce the incidence of pressure injuries in ICU patients by 58%, significantly improving the quality of care and patient prognosis.

[0084] like Figure 2 As shown, in some embodiments, the method further includes:

[0085] Determining whether the temperature change value exceeds a second temperature threshold;

[0086] Determining whether the first integral value exceeds a second integral threshold;

[0087] If the temperature change value does not exceed the second temperature threshold, the first integral value does not exceed the second integral threshold, and the number of impact factors is 1, then the risk level is determined to be low risk;

[0088] If the temperature change value does not exceed the second temperature threshold, the first integral value does not exceed the second integral threshold, and the number of influencing factors is 2, then the risk level is determined to be medium risk;

[0089] If the temperature change value exceeds a second temperature threshold, and / or the first integral value exceeds a second integral threshold, the risk level is determined to be high risk.

[0090] In this application, the temperature change value and the first integral value (PTI) are used as the core risk anchor points, and a three-level risk division is performed in combination with the number of influencing factors to construct a more accurate risk stratification system. When the temperature change value does not exceed the second temperature threshold (such as 2°C / h) and the PTI does not exceed the second integral threshold (such as 3000kPa·min), low risk and medium risk are distinguished according to the number of influencing factors (1 or 2), avoiding the traditional "either high or low" extensive classification. For example, when a single humidity exceeds the limit (influence factor = 1), it is judged as a low risk and only requires position adjustment; when the humidity + pressure dual factors exceed the limit (influence factor = 2), it is judged as a medium risk and dynamic decompression and local drying need to be initiated, so that the intervention measures are accurately matched to the risk level, and the utilization rate of nursing resources is increased by 40%.

[0091] When the temperature change value or any core indicator of PTI exceeds the limit, it is directly judged as high risk. This mechanism effectively captures critical scenarios such as deep tissue ischemia (PTI is too high) or acute inflammation (temperature rise). Clinical data show that this hierarchical judgment increases the accuracy of the correspondence between risk level and injury progression to 89% (traditional methods are only 65%), advancing the intervention time for medium-risk patients by 2-3 hours and shortening the response time for high-risk patients to within 10 minutes. After combined with graded warnings, the incidence of pressure injuries Stage I is reduced by 52%, and the incidence of severe injuries of Stage III-IV is reduced by 79%. In addition, this mechanism effectively reduces the misdiagnosis rate of risk assessment through cross-validation of quantitative indicators, providing a more scientific basis for clinical graded nursing decision-making.

[0092] like Figure 3 As shown, in some embodiments, the method further includes:

[0093] When the risk level is low risk or medium risk, if the first integral value exceeds a first integral threshold, an automatic decompression function is activated to adjust the mattress to a 30° tilt;

[0094] Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, run the automatic decompression function for the first duration; otherwise, send an alarm message to prompt interventional treatment.

[0095] In low-risk or medium-risk scenarios, when the first integral value (PTI) exceeds the first integral threshold (such as 1500kPa·min), the system automatically activates the decompression function of the mattress with a 30° tilt, reducing the local pressure accumulation effect by changing the distribution of pressure-bearing parts. This automated intervention can reduce the peak sacral pressure by 42%, which is more accurate and delay-free than traditional manual turning (the average manual operation takes 8 minutes, and the automatic decompression response time is less than 10 seconds). At the same time, the system continuously calculates the dynamic risk score and evaluates the level changes: if the risk is reduced, the automatic decompression function is maintained for the first duration (such as 2 hours), and the PTI accumulation rate is reduced by 63% through continuous decompression, effectively preventing the progression of pressure injuries; if the risk is not reduced, the alarm is immediately triggered to prompt interventional treatment to avoid intervention delays caused by the failure of a single decompression measure.

[0096] Clinical data show that this closed-loop mechanism has reduced the pressure injury conversion rate for low- and medium-risk patients from 27% to 9%. The effective intervention rate of automatic decompression reached 78%, and when combined with dynamic risk assessment, the timeliness of adjusting nursing measures increased by 90%. Furthermore, this mechanism reduces manual operations by 40% through "automatic intervention combined with intelligent assessment," while increasing the personalized adaptability of decompression strategies to 85% (traditional fixed turning only covers 50% of patient needs). This significantly reduces the waste of medical resources while improving nursing efficiency.

[0097] In some embodiments, the method further comprises:

[0098] When the risk level is low risk or medium risk, if the temperature change value exceeds a first temperature threshold, a local cooling function is activated to apply a semiconductor cooling patch;

[0099] Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, apply the semiconductor cooling patch for the second time period; otherwise, send an alarm message to prompt interventional treatment.

[0100] When the temperature change value in a low / medium risk scenario exceeds the first temperature threshold (such as 1.5°C / h), the system automatically activates the semiconductor cooling patch (-5°C) for local cooling, which can reduce the skin temperature of the inflamed area by 2.3°C within 10 minutes, effectively inhibiting tissue hypermetabolism and worsening ischemia. This measure is three times more efficient than traditional alcohol wiping (the traditional method takes 30 minutes to cool down 1°C), and avoids temperature fluctuations throughout the body through precise application. At the same time, the system continuously calculates the dynamic risk score. If the risk level is reduced, the cooling patch is maintained for the second period (such as 1 hour). Through continuous cooling, the local oxygen consumption rate is reduced by 41%, blocking the inflammatory cascade reaction; if the risk does not decrease, an alarm is immediately triggered to prompt interventional treatment to prevent local ischemia from progressing to irreversible damage.

[0101] Clinical data show that this mechanism reduces the incidence of temperature-related pressure injuries by 67%, with an 82% effectiveness rate for early inflammatory intervention and an alarm response time of less than 5 minutes when the risk level remains unchanged. Furthermore, this closed-loop control reduces manual cooling operations by 55%, increasing the personalized adaptability of cooling strategies to 89% (traditional methods only cover 40% of patient needs). This improves the accuracy of care while reducing the workload of medical staff, providing an intelligent solution for tissue protection in critically ill patients.

[0102] In some embodiments, the method further comprises:

[0103] When the risk level is low risk or medium risk, if the humidity value exceeds a first humidity threshold and lasts for more than a preset time, start the fan;

[0104] Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, run the fan for a third time period; otherwise, send an alarm message to prompt interventional treatment.

[0105] When the humidity value exceeds the first humidity threshold (such as 80% RH) in a low / medium risk scenario and lasts for more than a preset time (such as 20 minutes), the system automatically starts a micro fan (2m / s air supply) for local drying, which can reduce the humidity of the skin surface by 28% within 15 minutes, effectively improving the risk of immersion. Moisture loss in non-essential areas is avoided through directional airflow. At the same time, the system continuously calculates the dynamic risk score. If the risk level is reduced, the fan is maintained running for the third time (such as 30 minutes). Through continuous drying, the efficiency of skin barrier function recovery is increased by 63%; if the risk is not reduced, an alarm is immediately triggered to prompt interventional treatment to prevent the progression of skin barrier damage due to continued excessive humidity.

[0106] Clinical data demonstrates that this mechanism reduces the incidence of moisture-related pressure injuries by 71%, with early immersion intervention achieving an effectiveness rate of 85% and an alarm response time of less than 3 minutes when the risk level remains unchanged. Furthermore, this closed-loop control reduces manual drying operations by 60%, increasing the personalized adaptability of drying strategies to 91%. This not only enhances the level of care provided, but also reduces the repetitive workload of medical staff, providing technical support for intelligent care of the skin microenvironment in vulnerable areas.

[0107] In some embodiments, a dynamic risk scoring model is constructed based on an LSTM neural network, and inputting the features and adjusting the thresholds in combination with the patient's physiological parameters includes:

[0108] Z-normalization was performed on the pressure gradient, oxygenation decrease rate, and humidity time integral;

[0109] The patient's physiological parameters are converted into normalized feature vectors and mapped to the same dimensional space as the time series features through a fully connected layer;

[0110] A time sliding window is used to segment the multi-physics field data to generate a time series feature matrix; the multi-physics field data includes pressure data, tissue oxygenation status, and skin surface temperature and humidity data;

[0111] The input layer of the LSTM neural network receives a time series feature matrix, wherein the pressure gradient, oxygenation drop rate, and humidity time integral are processed by a time series input branch, and the physiological parameters are processed by a parallel static feature branch;

[0112] Thresholds are set based on the patient's physiological parameters, and correlation analysis is performed on the pressure gradient, oxygenation decrease rate, humidity time integral and historical pressure injury data to obtain the corresponding first correlation weight, second correlation weight and third correlation weight. The dynamic risk scoring model is reconstructed based on the first correlation weight, second correlation weight and third correlation weight.

[0113] This application performs Z-normalization on pressure gradients, oxygenation drop rates, and humidity time integrals, mapping patient physiological parameters to a temporal feature dimension space and using a time sliding window to generate a feature matrix. After processing using a dual-branch LSTM neural network architecture, the dynamic adjustment thresholds for physiological parameters and historical data correlation analysis are combined to determine feature weights and construct a dynamic risk scoring model. This process eliminates dimensional differences, improves the integrity of temporal features, and enables the joint modeling of physiological parameters and multi-physics field features. This improves the model's adaptability to individual tolerance differences by 60% and its ability to focus on vulnerable areas by 45%, providing precise technical support for early warning of pressure injuries.

[0114] Based on the LSTM neural network, multi-physical field data such as pressure gradient, oxygenation decrease rate, and humidity time integral are integrated. Dimensional differences are eliminated through Z normalization, and a time sliding window is combined to capture time series dynamic correlations (such as the cumulative effect of pressure-oxygenation). Compared with traditional single-point sampling or single parameter evaluation, the feature completeness is improved by 92%, the model AUC is increased from 0.62 to 0.85, and the accuracy is significantly optimized.

[0115] Individual physiological parameters drive dynamic adaptation: Physiological parameters such as BMI and HbA1c are mapped to the time series feature space, and a dual-branch architecture (dynamic time series combined with static physiology) is used to achieve joint modeling of "dynamic changes in risk and individual tolerance differences", which improves the adaptability of the pressure warning threshold to individual differences by 60% (such as a 27% increase in ischemia sensitivity recognition in diabetic patients), avoids misjudgments caused by traditional fixed thresholds (such as 32mmHg), and effectively improves the misjudgment rate.

[0116] Based on correlation analysis of historical pressure injury data, the system dynamically adjusts the weights of various features (e.g., a pressure gradient weight of ≥40% when the sacrum is compressed). This improves the risk focus on vulnerable areas by 45%, changing the traditional "one-size-fits-all" assessment model to enable differentiated and accurate assessments for different body parts and patient populations. Feature standardization improves model input stability by 70%, and a federated learning mechanism (to be updated) further optimizes cross-scenario adaptability, overcoming the limitations of traditional models in data compatibility and cross-group application.

[0117] In some embodiments, the method further comprises:

[0118] A near-infrared spectral sensor with a wavelength of 650-950nm is used to continuously monitor tissue oxygen saturation StO2 and oxygen metabolism rate. An early warning is activated when StO2 is less than 40% for 2 hours.

[0119] This application uses a NIRS sensor with a wavelength of 650-950nm, which can penetrate the skin to a depth of 3mm and directly monitor the oxygen saturation (StO2) and oxygen metabolism rate of deep tissues, breaking through the limitation of traditional naked eye observation that can only identify epidermal changes. When StO2 is less than 40% and lasts for 2 hours, the warning is activated. At this time, the tissue is in a reversible stage of ischemia and hypoxia, which can detect risks 4-6 hours earlier than traditional methods (such as epidermal redness Stage I), thus gaining a critical time window for clinical intervention. For example, when an ICU patient is under pressure on the sacrum and coccyx, this monitoring mechanism can trigger an early warning at the early stage of a decrease in oxygenation in deep muscle tissue, thereby avoiding necrosis of fat and muscle tissue due to continued ischemia.

[0120] At the same time, continuous oxygen metabolism monitoring can quantitatively assess the efficiency of tissue response to decompression measures. For example, after initiating mattress tilt, if StO2 rises above 50% within 30 minutes, it indicates that the intervention was effective. If it remains below 40%, an upgraded alarm is triggered, shifting the management of ischemic risk from "after-the-fact treatment" to "real-time regulation." Furthermore, this technology avoids the infection risks of invasive monitoring and can be worn long-term on vulnerable areas such as the hips and sacrum, providing continuous tissue oxygenation for bedridden patients.

[0121] In some embodiments, the method further comprises:

[0122] When the risk level is low, a yellow light warning is issued;

[0123] When the risk level is medium, an orange light warning will be issued;

[0124] When the risk level is high, a red light alarm will be issued.

[0125] This application can quickly confirm the risk level through different indicator lights, allowing medical staff and patients to quickly understand whether pressure injuries have occurred.

[0126] In some embodiments, the method further comprises:

[0127] When an early warning is triggered, the system automatically extracts the alarm time, pressure location and corresponding treatment measures, and generates a structured nursing record;

[0128] A federated learning mechanism is used to update risk model parameters and upload them to the server.

[0129] When the warning is triggered, the system automatically extracts the alarm time (accurate to seconds), the pressure position (positioned to 4×4cm 2 Structured nursing records are generated for each patient, including the patient's specific area (e.g., 30° tilt, local cooling, etc.), replacing traditional manual writing. This reduces recording time from an average of 15 minutes to less than 10 seconds and improves data integrity from 72% to 100%. Structured records are synchronized to electronic medical records via the HL7 protocol, eliminating human transcription errors, significantly improving the traceability and compliance of nursing data, and providing standardized data support for medical quality assessment.

[0130] When using a federated learning mechanism to update risk model parameters, each hospital node locally trains the model gradient using anonymized clinical data, which is then uploaded to the server via a secure aggregation protocol. Global parameter updates are completed monthly. This mechanism, while protecting patient privacy (without uploading original data), continuously improves the model's adaptability to different hospital patient populations. For example, for geriatric patients, the model can automatically increase the weight of the humidity integral (by 15%), and for ICU patients, it enhances sensitivity to the rate of oxygenation decline (by 20%).

[0131] Example

[0132] The pressure distribution time series data were acquired through the MEMS piezoresistive array (10 Hz sampling), the tissue oxygen saturation (StO2) was collected by the near-infrared spectroscopy sensor (NIRS, 5 Hz sampling), and the epidermal humidity (RH) was monitored by a flexible capacitive humidity sensor (5 Hz sampling).

[0133] FPGA is used to align the three types of data in real time to ensure that the spatiotemporal synchronization delay is less than 50ms, and Kalman filtering is used to eliminate motion artifacts such as turning over.

[0134] Pressure gradient (ΔP / Δx): Calculates the rate of change of pressure within a unit space and identifies areas of concentrated pressure.

[0135] Oxygenation decline rate (dStO2 / dt): Continuously monitors the decay rate of StO2 over time, reflecting the trend of tissue ischemia;

[0136] Humidity time integral (∫RH·dt): Accumulates the duration of exposure to high humidity (>80% RH) to assess the risk of skin immersion.

[0137] The pressure gradient, oxygenation decrease rate, and humidity time integral were Z-score standardized to eliminate the dimension effect;

[0138] The Spearman rank correlation coefficient is used to calculate the correlation between each feature and the historical pressure change data. The formula is:

[0139]

[0140] Among them, (d i ) is the rank difference between the feature and the pressure change, n is the sample size (based on ≥10 ^4 Sacrococcygeal and hip compression data of 100 patients);

[0141] Normalize the correlation degree to [0,1] to obtain the first correlation weight w1 (pressure gradient weight), the second correlation weight w2 (oxygenation rate weight), and the third correlation weight w3 (humidity integral weight), satisfying w 1+ w 2+ w 3＝1 , and when the sacrum and coccyx are compressed, w2 accounts for ≥40%.

[0142] Before the LSTM input layer, the three types of features are weighted and fused:

[0143] Weighted feature = w1×ΔP / Δx+w2×dStO2 / dt+w3×∫RH·dt;

[0144] The patient's BMI, HbA1c and other physiological parameters are converted into normalized feature vectors and mapped to the same dimensional space as the time series features through a fully connected layer;

[0145] A time sliding window (window length 2 hours, step length 15 minutes) is used to segment the multi-physics field data and generate a time series feature matrix.

[0146] A three-layer bidirectional LSTM network with 128 hidden neurons in each layer is used to capture the long-term and short-term dependencies between features through a gating mechanism.

[0147] The input layer receives a multimodal feature matrix, in which the pressure gradient, oxygenation drop rate, and humidity time integral are processed by the sequential input branch, and the physiological parameters are processed by the parallel static feature branch.

[0148] In the LSTM hidden layer, the attention mechanism is introduced to calculate the importance weight of the features of each time step. The formula is:

[0149]

[0150] Among them, h t is the hidden state at time t, st-1 is the context vector of the previous moment, and the attention score function adopts the dot product form.

[0151] Based on the patient's physiological parameters, an adaptive threshold vector θ = [θ1, θ2, θ3] is generated by a multi-layer perceptron (MLP), where:

[0152]

[0153] is the correction coefficient, which triggers nonlinear correction when BMI ≥ 30 or HbA1c ≥ 8;

[0154] Compare the feature vector output by LSTM with the threshold vector element by element to calculate the normalized risk score:

[0155]

[0156] Among them, x i is the LSTM output value of each feature, σ is the Sigmoid function;

[0157] The final risk score is calculated by combining the associated weights w1, w2, and w3:

[0158]

[0159] In this embodiment, if the risk level is less than 80 points, the response factor is judged. When PTI>3000kPa·min or the temperature change value ΔT>2°C and the humidity fluctuation time exceeds 20 minutes, a red light warning is triggered to prompt medical staff to intervene in treatment.

[0160] If the risk level is greater than 80 points, a red light warning will be triggered, prompting medical staff to intervene in treatment.

[0161] If the risk level is less than 80 points, the response factor is judged. When PTI>1500kPa·min and the temperature change value ΔT>0.8℃, the orange light warning is triggered and the dynamic decompression function is executed.

[0162] If the risk level is less than 80 points, the response factor is determined. When the humidity is greater than 80% RH, the temperature change value ΔT is greater than 1.5°C, and the pressure is greater than 1500 kPa·min, a yellow light warning is triggered and the body position is adjusted.

[0163] In some embodiments, the multimodal sensing-based pressure injury dynamic risk assessment method provided herein can be configured in an integrated chip in some pressure injury protection gear, wherein the pressure injury protection gear comprises:

[0164] Protective gear body; an anti-pressure pad sewn onto the protective gear body, wherein the coverage area of ​​the anti-pressure pad satisfies the following anatomical positioning relationship: the upper boundary of the longitudinal range of the anti-pressure pad is located 5 to 6 cm above the superior sacral spine, and the lower boundary is located 5 to 6 cm below the sacrum; the left and right boundaries of the transverse range of the anti-pressure pad are located 10 to 11 cm lateral to the left anterior superior iliac spine and the right anterior superior iliac spine, respectively; the length of the anti-pressure pad is 50 to 60 cm, the width of the anti-pressure pad is 40 to 50 cm, and the thickness of the anti-pressure pad is 0.8 to 1.2 mm;

[0165] The anti-pressure pad includes, from the outside to the inside, a flexible breathable layer, an optical silicone layer, a multimodal resistance sensing layer and a skin contact layer, and the flexible breathable layer, the optical silicone layer, the multimodal resistance sensing layer and the skin contact layer are bonded together by an adhesive; the thickness of the flexible breathable layer is 0.1 to 0.2 mm, the thickness of the optical silicone layer is 0.5 to 0.65 mm, the thickness of the multimodal resistance sensing layer is 0.1 to 0.15 mm, and the thickness of the skin contact layer is 0.1 to 0.2 mm.

[0166] The flexible breathable layer includes a TPU film and a dielectric elastomer, and the TPU film and the dielectric elastomer are connected by hot pressing or UV adhesive bonding, wherein the thickness of the TPU film is 0.05-0.1 mm, and the thickness of the dielectric elastomer is 0.1-0.15 mm.

[0167] The optical silicone layer is made of PDMS, and a triboelectric nanogenerator, a flexible microcircuit board, a battery, a temperature sensor, a humidity sensor and an infrared sensor are integrated inside the optical silicone layer; an integrated chip is provided on the flexible microcircuit board, and the integrated chip is electrically connected to the triboelectric nanogenerator, the temperature sensor, the humidity sensor and the infrared sensor. The triboelectric nanogenerator, the temperature sensor, the humidity sensor, the infrared sensor and the battery are connected to the ports on the flexible microcircuit board through wires.

[0168] The pressure sensor layer is mainly composed of a graphene-PDMS piezoresistive composite material and is connected to a multi-level sensor array through a silver nanowire embedded circuit; wherein the sensor array includes:

[0169] The first array, located in the sacral and coccyx regions, has an area of ​​4 × 4 cm and contains a 3 × 3 grid;

[0170] The second array, located in the femoral region, has an area of ​​4 × 4 cm and contains a 2 × 2 grid;

[0171] Each grid is equipped with a pressure sensor.

[0172] The skin contact layer is a modified hydrocolloid, and the surface of the skin contact layer is provided with a plurality of small holes with a diameter of 50 to 60 μm, and the hole density is ≥300 holes / cm 2 , evenly arranged in hexagons with a spacing of 150μm, a total breathable area of ​​1%±0.2%, and a moisture permeability of ≥1200g / m 2 ·day.

[0173] The surface of the anti-pressure pad is provided with indicator lights electrically connected to the integrated chip, and the indicator lights include a first indicator light, a second indicator light and a third indicator light. The first indicator light, the second indicator light and the third indicator light are used to display the pressure injury risk level in a graded manner.

[0174] The anti-pressure pad is designed with a length of 50 to 60 cm (adjusted according to individual conditions) and a width of 40 to 50 cm (adjusted according to individual conditions) to completely cover the main pressure-bearing surface of the human body when lying on the side, supine, or semi-recumbent. It has a certain degree of stretchability and can adapt to the stretching of the skin during movement. The ultra-thin structure with a thickness of 0.8 to 1.2 mm not only ensures the skin care and pressure relief effect, but also avoids the patient's movement being restricted due to the thickness of the pad. The flexible breathable layer combines TPU film and dielectric elastomer. While ensuring moisture permeability, it senses micro-deformation through the dielectric elastomer to assist in pressure distribution monitoring; the optical silicone layer has built-in temperature / humidity / infrared sensors and triboelectric nanogenerators, which can collect skin microenvironment data in real time and maintain the device's battery life through a self-powered system; the multimodal resistive sensing layer uses a graphene-PDMS piezoresistive composite material to construct a 5×5 grid sensor array with a pressure detection resolution of ±1 mmHg. It can capture continuous pressure >32 mmHg and trigger an early warning within 2 hours. The surface of the modified hydrocolloid layer is equipped with hexagonal uniform breathable holes with a pore size of 50±5μm, which not only ensures the skin's breathing but also absorbs exudate, reducing the risk of immersion. Through the humidity sensor and temperature compensation algorithm, combined with the porous flexible nanostructure of the breathable layer, the local humidity and heat environment can be adjusted in real time.

[0175] As can be seen from the above embodiments, the present application provides a dynamic risk assessment method for pressure injuries based on multimodal sensing, which includes using a MEMS piezoresistive array, a near-infrared spectral sensor, and a flexible capacitive humidity sensor to collect pressure data, tissue oxygenation status, skin surface temperature, and humidity data in real time; performing signal cleaning and calculation on the collected data to obtain pressure gradient, oxygenation decrease rate, and humidity time integral characteristics; constructing a dynamic risk scoring model based on an LSTM neural network, inputting the characteristics and adjusting the threshold in combination with the patient's physiological parameters to calculate a dynamic risk score; executing a risk level classification warning based on the dynamic risk score, sending a posture adjustment prompt message when the risk level is low risk, or sending an alarm message when the risk level is high risk. The present application uses a MEMS piezoresistive array, a near-infrared spectral sensor, and a flexible capacitive humidity sensor to collect multi-physical field data in real time, and combines the LSTM neural network with dynamic threshold adjustment to achieve early accurate warning and automated graded intervention. This method advances the warning time of pressure injuries by 4-6 hours, reducing the clinical misjudgment rate from 38% to <15%. It solves the problem in existing technologies that traditional naked eye observation cannot accurately identify poor blood circulation in subcutaneous tissue, the lag in taking preventive measures and the high error rate in clinical accurate judgment, making it impossible to accurately judge the risk and extent of pressure injuries.

[0176] Similar parts between the embodiments provided in this application can be referenced to each other. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods expanded based on the scheme of this application without expending creative work shall fall within the scope of protection of this application.

Claims

1. A dynamic risk assessment method for pressure injuries based on multimodal sensing, characterized in that: The method comprises: MEMS piezoresistive arrays, near-infrared spectral sensors, and flexible capacitive humidity sensors are used to collect real-time pressure data, tissue oxygenation status, and skin surface temperature and humidity data; The collected data is cleaned and calculated to obtain the pressure gradient, oxygenation drop rate and humidity time integral characteristics; A dynamic risk scoring model is constructed based on an LSTM neural network, the features are input and the threshold is adjusted in combination with the patient's physiological parameters to calculate the dynamic risk score; A risk level classification warning is performed according to the dynamic risk score, and a body position adjustment prompt message is sent when the risk level is low risk, or an alarm message is sent when the risk level is high risk.

2. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 1, characterized in that: Executing graded warnings based on the dynamic risk score includes: If the dynamic risk score is higher than 80 points, the risk level is confirmed as high risk and an alarm message is sent to prompt interventional treatment; If the dynamic risk score is less than 80 points, determine the number of influencing factors; wherein the influencing factors are determined based on temperature data, humidity data, and pressure data; If the number of influencing factors is greater than 2, the risk level is confirmed to be high risk, and an alarm message is sent to prompt interventional treatment.

3. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 2, characterized in that: The number of impact factors to be determined includes: Calculate the temperature change value based on the current temperature and the previous temperature in the temperature data; Determine the humidity value of the current humidity data and the corresponding duration; Constructing a pressure-time integral to calculate a first integral value; The temperature change value, the humidity value, and the first integral value are respectively compared with their corresponding preset thresholds to obtain the number of influencing factors; wherein the number exceeding the preset threshold is the number of influencing factors.

4. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 3, characterized in that: The method further comprises: Determining whether the temperature change value exceeds a second temperature threshold; Determining whether the first integral value exceeds a second integral threshold; If the temperature change value does not exceed the second temperature threshold, the first integral value does not exceed the second integral threshold, and the number of impact factors is 1, then the risk level is determined to be low risk; If the temperature change value does not exceed the second temperature threshold, the first integral value does not exceed the second integral threshold, and the number of influencing factors is 2, then the risk level is determined to be medium risk; If the temperature change value exceeds a second temperature threshold, and / or the first integral value exceeds a second integral threshold, the risk level is determined to be high risk.

5. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 4, characterized in that: The method further comprises: When the risk level is low risk or medium risk, if the first integral value exceeds a first integral threshold, an automatic decompression function is activated to adjust the mattress to a 30° tilt; Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, run the automatic decompression function for the first duration; otherwise, send an alarm message to prompt interventional treatment.

6. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 4, characterized in that: The method further comprises: When the risk level is low risk or medium risk, if the temperature change value exceeds a first temperature threshold, a local cooling function is activated to apply a semiconductor cooling patch; Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, apply the semiconductor cooling patch for the second time period; otherwise, send an alarm message to prompt interventional treatment.

7. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 4, characterized in that: The method further comprises: When the risk level is low risk or medium risk, if the humidity value exceeds a first humidity threshold and lasts for more than a preset time, start the fan; Calculate the current dynamic risk score, and determine the risk level based on the dynamic risk score. If the risk level decreases, run the fan for a third time period; otherwise, send an alarm message to prompt interventional treatment.

8. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 1, characterized in that: A dynamic risk scoring model is constructed based on an LSTM neural network. The features are input and the thresholds are adjusted based on the patient's physiological parameters. Z-normalization was performed on the pressure gradient, oxygenation decrease rate, and humidity time integral; The patient's physiological parameters are converted into normalized feature vectors and mapped to the same dimensional space as the time series features through a fully connected layer; A time sliding window is used to segment the multi-physics field data to generate a time series feature matrix; the multi-physics field data includes pressure data, tissue oxygenation status, and skin surface temperature and humidity data; The input layer of the LSTM neural network receives a time series feature matrix, wherein the pressure gradient, oxygenation drop rate, and humidity time integral are processed by a time series input branch, and the physiological parameters are processed by a parallel static feature branch; Thresholds are set based on the patient's physiological parameters, and correlation analysis is performed on the pressure gradient, oxygenation decrease rate, humidity time integral and historical pressure injury data to obtain the corresponding first correlation weight, second correlation weight and third correlation weight. The dynamic risk scoring model is reconstructed based on the first correlation weight, second correlation weight and third correlation weight.

9. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 1, characterized in that: The method further comprises: A near-infrared spectral sensor with a wavelength of 650-950nm is used to continuously monitor tissue oxygen saturation StO2 and oxygen metabolism rate. An early warning is activated when StO2 is less than 40% for 2 hours.

10. The method for dynamic risk assessment of pressure injuries based on multimodal sensing according to claim 1, characterized in that: The method further comprises: When an early warning is triggered, the alarm time, pressure location and corresponding treatment measures are automatically extracted to generate a structured nursing record; A federated learning mechanism is used to update risk model parameters and upload them to the server.

Images