Anti-falling infrared sensing monitoring device and method for acute and critical patients

Abstract

The invention discloses an anti-falling infrared sensing monitoring device and method for an acute and critical patient, and belongs to the technical field of medical monitoring. The anti-falling infrared sensing monitoring device comprises a multi-mode sensor array, an edge calculation module and a grading alarm module; the multi-mode sensor array is used for collecting data, real-time processing of the edge calculation module and machine learning model analysis are combined, graded early warning is achieved, and the method has the advantages that monitoring accuracy is improved, active early warning is achieved, response delay is reduced, and the false alarm rate is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical monitoring, and particularly relates to an infrared sensing monitoring device and method for preventing falls in critically ill patients. Background Art

[0002] The current monitoring of falls and physiological abnormalities in critically ill patients mainly relies on the following technologies: single-sensor solutions, such as infrared or camera-based behavior monitoring systems, are vulnerable to environmental occlusion and light changes, and cannot detect minute movements (such as precursors of convulsions) or quantify gait parameters; passive alarms, most systems only trigger alarms after a fall occurs, lacking the ability to actively warn of precursors such as rapid breathing and gait imbalance; strong dependence on the cloud, traditional solutions need to transmit raw data to the cloud for processing, resulting in excessive delays (>2 seconds) and unable to meet the real-time requirements of emergency scenarios; high false alarm rate, due to the lack of fusion of multi-modal data (such as pressure distribution and kinematic correlation analysis), nursing operations (such as turning over and patting the back) are easily misjudged as fall events. In view of the above problems, the existing technologies urgently need to be improved. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technologies, the present invention provides an infrared sensing monitoring device and method for preventing falls in critically ill patients, which solves the above problems.

[0004] To achieve the above objectives, the present invention is realized through the following technical solutions: an infrared sensing monitoring device for preventing falls in critically ill patients, comprising:

[0005] A multi-modal sensor array, including a pyroelectric infrared sensor array arranged in the patient's activity area for detecting the spatio-temporal changes of human infrared radiation, a wearable inertial measurement unit configured on the patient's waist or wrist to collect three-axis acceleration, angular velocity and attitude angle data in real time, and a piezoelectric pressure sensor integrated on the hospital bed or the floor of the rehabilitation corridor to monitor the patient's body position distribution and the trajectory of the plantar pressure center;

[0006] An edge computing module, including an embedded AI chip and a signal processing circuit for real-time filtering, feature extraction and time synchronization of multi-modal sensor data, running a lightweight machine learning model, fusing infrared trajectory, IMU kinematic features and pressure distribution data, and generating a fall risk level;

[0007] A hierarchical alarm module, which triggers the following actions according to the risk level:

[0008] First-level warning: Prompt the patient of gait imbalance or rapid breathing through voice prompts or local lighting warnings;

[0009] Second-level alarm: Push the patient's location, abnormal physiological parameter values and recommended intervention measures to the medical staff terminal.

[0010] Based on the above technical solutions, the present invention also provides the following alternative technical solutions:

[0011] Further technical solution: The arrangement method of the pyroelectric infrared sensor array is as follows:

[0012] Directional infrared sensors are arranged at the four corners of the hospital bed to cover the out-of-bed detection area;

[0013] Transmissive infrared fences are installed on both sides of the rehabilitation corridor to form a continuous motion trajectory monitoring network.

[0014] Further technical solution: The machine learning model of the edge computing module includes:

[0015] A time series classifier based on 1D-CNN, with the input being the standard deviation of the acceleration of the IMU, the change rate of the attitude angle, and the stationary duration after impact;

[0016] A dynamic threshold adjustment unit that adaptively corrects the alarm trigger condition according to the patient's historical activity data.

[0017] Further technical solution: It further includes a tachypnea quantification module:

[0018] By reusing the thoracic motion data of the IMU or an independent respiratory belt sensor, the respiratory rate and tidal volume are calculated;

[0019] When RR > 25 times / minute and TV < 300 ml, it is marked as a tachypnea event and associated with a secondary alarm.

[0020] Further technical solution: It further includes: The pre-seizure detection module uses surface electromyography sensors attached to the patient's limbs to capture the root mean square and high-frequency energy ratio of the muscle discharge signal, and when the RMS of the surface electromyography signal suddenly increases by more than 2 standard deviations from the baseline within a 10-second window and the high-frequency energy > 40%, a seizure warning is triggered.

[0021] An infrared sensing monitoring method for preventing falls in critically ill patients, using the above-mentioned infrared sensing monitoring device for preventing falls in critically ill patients, includes the following steps:

[0022] S1. Synchronously collect infrared radiation, kinematic, and pressure data through a multi-modal sensor array;

[0023] S2. Perform the following operations in the edge computing module:

[0024] Perform peak detection on the IMU data, segment the gait cycle, and calculate the step length, walking speed, and standard deviation of the trunk inclination;

[0025] Perform clustering analysis on the infrared sensor data to identify sudden changes in the patient's movement direction or abnormal stillness;

[0026] S3, perform risk assessment on the fusion features based on the random forest model and output a risk level of 0-10 points;

[0027] S4. When the risk level is ≥ 7 points, activate the second-level alarm and retrieve the video clips of the related ward for review by medical staff.

[0028] Further technical solution: The risk assessment model is optimized through transfer learning based on a pre-trained model skeleton based on a public fall dataset and using patient gait and breathing data from a local hospital to fine-tune the classification boundaries.

[0029] The present invention provides an infrared sensing monitoring device and method for preventing critically ill patients from falling, which has the following beneficial effects compared with the prior art:

[0030] 1. This application collects data through a multimodal sensor array, combines real-time processing of the edge computing module and machine learning model analysis to achieve graded early warning, which has the advantages of improving monitoring accuracy, achieving active early warning, reducing response delays, and reducing false alarm rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0033] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0034] Example 1

[0035] In the existing technology, the monitoring of falls and physiological abnormalities in critically ill patients mainly adopts a single sensor solution, such as an infrared or camera-based behavioral monitoring system. Such systems are easily affected by environmental occlusion and light changes, and cannot detect small movements or quantify gait parameters. Traditional monitoring methods are mostly passive alarms, which are only triggered after a fall occurs, and lack the ability to actively warn of precursors such as shortness of breath and gait imbalance. In addition, existing solutions rely on the cloud to process raw data, resulting in excessive latency and unable to meet the real-time requirements of emergency scenarios. The defect of not integrating multimodal data leads to a high false alarm rate. For example, nursing operations can easily be misjudged as fall events.

[0036] To solve the above problems, it is first necessary to overcome the limitations of single-sensor monitoring and improve the detection accuracy through multi-dimensional data fusion. For the problem of early warning lag, a real-time localization processing mechanism needs to be designed. To reduce the false alarm rate, an association analysis model between multi-modal data should be established. Finally, a hierarchical response mechanism needs to be formed to balance the timeliness of early warning and the allocation efficiency of medical resources.

[0037] Please refer to Figure 1 , a fall prevention infrared sensing monitoring device for critically ill patients provided by an embodiment of the present invention, includes a multi-modal sensor array, an edge computing module, and a hierarchical alarm module. The multi-modal sensor array includes a pyroelectric infrared sensor array arranged in the patient's activity area, a wearable inertial measurement unit, and a piezoelectric pressure sensor integrated on the ground of the hospital bed or the rehabilitation corridor. The edge computing module includes an embedded AI chip and a signal processing circuit for real-time processing of sensor data and running a lightweight machine learning model. The hierarchical alarm module triggers different levels of early warning actions according to the risk level.

[0038] Among them, the pyroelectric infrared sensor array realizes non-contact monitoring by detecting the spatio-temporal changes of human infrared radiation. Specifically, a sensor network with multi-node distribution can be used to cover the patient's activity area to capture sudden changes in the motion trajectory. The wearable inertial measurement unit quantifies kinematic characteristics by collecting triaxial acceleration and attitude angle data. Specifically, a micro-sensor with low-power Bluetooth transmission can be used. The piezoelectric pressure sensor verifies gait abnormalities by monitoring the plantar pressure distribution. Specifically, a flexible thin-film sensor array can be used. The lightweight machine learning model of the edge computing module realizes real-time risk assessment by fusing multi-source features. Specifically, a classification algorithm based on a convolutional neural network can be used. The differential response mechanism of the hierarchical alarm module is divided by risk level thresholds. Specifically, a preset weight allocation strategy can be used to trigger corresponding early warning actions.

[0039] Specifically, the system synchronously collects infrared radiation, kinematic, and pressure data through the multi-modal sensor array. The infrared sensor network continuously monitors the patient's motion trajectory and triggers data collection when an abnormal displacement is detected. The wearable inertial measurement unit records the changes in acceleration and attitude angle in real time and aligns them with the pressure sensor data through timestamps. The edge computing module conducts joint analysis on the multi-source data and extracts characteristic parameters such as the step length variation coefficient and the standard deviation of the trunk inclination angle. The machine learning model maps the fused feature vector to a fall risk level and updates the risk assessment result when abnormal breathing frequency or center of pressure shift is detected. The hierarchical alarm module selects a response method according to the dynamic assessment result. Low-risk events trigger local prompts, while high-risk situations activate the medical staff linkage mechanism.

[0040] Compared with the prior art, the proposed solution effectively distinguishes real falls from nursing operations through multi-modal data fusion. For example, the pressure sensor can eliminate false alarms caused by simple limb movements. The localization processing of the edge computing module avoids cloud transmission delays, reducing the warning response time to the sub-second level. The hierarchical alarm mechanism not only preserves the patient's opportunity for self-adjustment but also ensures timely intervention in high-risk events, meeting clinical needs better than traditional single-alarm modes. Additionally, the dynamically updated risk assessment model can optimize the judgment threshold according to the patient's individual characteristics, showing higher adaptability compared to monitoring systems with fixed thresholds.

[0041] Through the above technical solution, this application realizes real-time multi-dimensional monitoring of the physiological state of critically ill patients, effectively identifying the precursors of falls and reducing the false alarm rate. The system ensures the timeliness requirements of the first-aid scenario through localized data processing, and the hierarchical response mechanism optimizes the efficiency of medical resource allocation. The collaborative work of multi-modal sensors enhances the monitoring reliability in complex environments, providing accurate data support for early intervention.

[0042] Preferably, the arrangement of the pyroelectric infrared sensor array is as follows: directional infrared sensors are set at the four corners of the hospital bed to cover the out-of-bed detection area; opposed infrared barriers are installed on both sides of the rehabilitation corridor to form a continuous motion trajectory monitoring network.

[0043] Among them, the directional infrared sensor refers to an infrared sensing device with a specific angular detection range, which can be specifically realized by using a Fresnel lens in combination with a pyroelectric element. Its function is to eliminate the detection blind area of a single sensor through cross-coverage by multiple sensors. The opposed infrared barrier refers to an array of infrared light beams arranged in pairs with a transmitter and a receiver, which can be specifically realized by combining a pulse-modulated infrared diode and a phototransistor. Its function is to capture the continuous change of the motion trajectory through a two-way light beam network.

[0044] Specifically, the four groups of directional infrared sensors set at the four corners of the hospital bed can detect the limb movements of the patient when getting out of bed from different directions through the cross-coverage principle. The detection area of each sensor is angle-calibrated to form a three-dimensional monitoring network covering the edge area of the bed. When a part of the patient's body exceeds the bed boundary, at least two sensors will trigger signals simultaneously, thus eliminating the visual dead angle existing in the traditional single-sensor arrangement. The opposed infrared barriers deployed on both sides of the rehabilitation corridor construct a parallel light beam array extending along the corridor through multiple groups of infrared transmitting-receiving devices arranged at intervals. When the patient moves, the occlusion sequence of the body to the light beams can be converted into spatio-temporal distribution data of the motion trajectory, continuously recording the change characteristics of the moving direction. The collaborative work of the two arrangement methods not only realizes the precise monitoring of key areas around the hospital bed but also ensures the continuous tracking of the motion state on the rehabilitation path.

[0045] Compared with the prior art, the traditional single infrared sensor can only cover one-sided area in the hospital bed monitoring, which is prone to missed detection due to the change of the patient's body position. While in the corridor monitoring, the single-point sensor is usually adopted, which cannot continuously capture the movement trajectory. Through the layout of the multi-angle sensor array, this solution forms a redundant detection mechanism in the same space, significantly improving the detection reliability of the out-of-bed action. The corridor monitoring adopts a two-way beam network to replace the traditional single-point trigger mode, which can continuously record the change of the movement direction and speed, providing more complete spatio-temporal data for the analysis of the precursor of falling.

[0046] Through the above technical solution, this application can effectively identify the limb overstep behavior when the patient gets out of bed, avoiding the false alarm phenomenon caused by the sensor blind area. At the same time, during the rehabilitation walking process, it continuously monitors the mutation characteristics of the movement trajectory, providing more accurate spatio-temporal information support for early warning, and significantly reducing the misjudgment risk caused by discontinuous monitoring.

[0047] Preferably, the machine learning model of the edge computing module includes a temporal classifier based on a one-dimensional convolutional neural network and a dynamic threshold adjustment unit.

[0048] Among them, the temporal classifier of the one-dimensional convolutional neural network refers to a deep learning architecture for extracting time series features from the data of the inertial measurement unit. Specifically, it can be implemented by using a neural network structure with three convolutional kernels. The width of the convolutional kernel is set to 5 time steps to capture the temporal correlation of the standard deviation of acceleration, the change rate of the attitude angle, and the stationary duration after impact. This architecture processes continuous time series data through a sliding window method, and can effectively extract the mutation characteristics unique to the falling action.

[0049] Among them, the dynamic threshold adjustment unit refers to an adaptive module for optimizing the alarm condition according to the individual activity pattern of the patient. Specifically, it can be realized by statistically calculating the mean and standard deviation of the patient's historical activity parameters through a sliding time window, and automatically updating the decision boundary threshold of the classification model every day. This unit continuously records the gait stability data of the patient at different rehabilitation stages, and establishes personalized alarm reference parameters.

[0050] Specifically, during the continuous acquisition of the patient's movement data by the inertial measurement unit, the temporal classifier of the one-dimensional convolutional neural network performs waveform analysis on the standard deviation of acceleration to identify sudden movement events. The change rate of the attitude angle parameter judges whether the body center of gravity deviation exceeds the normal gait range by calculating the change rate of the trunk tilt angle. The stationary duration feature after impact distinguishes real falls from normal actions such as quickly sitting down by detecting the duration of movement stagnation after a sudden action. The dynamic threshold adjustment unit automatically adjusts the alarm threshold range of the change rate of the attitude angle by analyzing the distribution of the patient's activity parameters in the past 24 hours. For example, a looser threshold is set for patients in the rehabilitation period to avoid over-alarm, and a more sensitive trigger condition is set for patients with muscle strength decline.

[0051] Compared with existing technologies, traditional fall monitoring systems use fixed threshold judgment rules, such as setting a uniform setting of an alarm when the posture angle change rate exceeds 30 degrees / second, which cannot distinguish individual differences among patients. This solution extracts multi-dimensional time series features through a machine learning model and establishes a dynamic alarm threshold mechanism. For example, it optimizes the classification model parameters based on the tremor characteristics of Parkinson's patients, effectively avoiding misjudging pathological tremors as fall events.

[0052] Through the above-mentioned technical scheme, the present application can reduce the problem of false alarms caused by differences in individual activity patterns of patients, such as avoiding misjudging wheelchair transfer movements as falls, and dynamically adapt to the gradual changes in gait characteristics of patients during rehabilitation, such as gradually relaxing the posture angle alarm threshold during the muscle strength recovery period, thereby achieving continuous matching of the alarm logic with the patient's actual physiological state.

[0053] Preferably, a shortness of breath quantification module is also included, which calculates the respiratory rate and tidal volume by reusing the chest motion data of the inertial measurement unit or an independent respiratory belt sensor. When the respiratory rate exceeds 25 times per minute and the tidal volume is less than 300 ml, it is marked as a shortness of breath event and associated with a secondary alarm.

[0054] Among them, the chest motion data of the multiplexing inertial measurement unit refers to the use of motion sensors that have been configured on the patient's body surface to capture the chest rise and fall signal. Specifically, it can be achieved by using the periodic fluctuation data of the axial direction of the accelerometer to avoid adding dedicated respiratory monitoring equipment. The independent respiratory belt sensor refers to the use of elastic fabric with a built-in piezoresistive sensor unit to surround the chest. Specifically, it can be achieved by using the linear relationship between the piezoresistive signal and the chest circumference change, providing redundant measurements when the inertial measurement unit signal is disturbed by motion. The calculation of respiratory frequency refers to the statistical number of periodic chest rises and falls per unit time. Specifically, the peak detection algorithm can be used to automatically count the respiratory waveform. The calculation of tidal volume refers to the estimation of the gas exchange volume in a single respiratory cycle. Specifically, it can be mapped and converted by the chest circumference expansion amplitude and the preset vital capacity calibration curve. The respiratory rate threshold of 25 times per minute is set based on the clinical diagnostic standard for shortness of breath, and the tidal volume threshold of 300 ml refers to the typical characteristic value of shallow and fast breathing.

[0055] Specifically, the module collects respiratory function parameters synchronously by reusing the chest movement signals of existing motion sensors or adding a dedicated respiratory belt. The respiratory rate is obtained by detecting the number of chest rise and fall cycles, and the tidal volume is converted according to the chest circumference expansion amplitude. When the respiratory rate exceeds the clinical threshold and is accompanied by a significant decrease in tidal volume, the system determines it as a shortness of breath event. This dual-parameter joint judgment avoids the misjudgment of a single indicator caused by motion interference. For example, when a patient turns over, the chest movement frequency may temporarily increase, but the tidal volume remains normal. After the abnormal event is marked, the system automatically raises the alarm level to level 2, triggering a directional push to the medical terminal.

[0056] Compared with the prior art, traditional solutions rely on single infrared or camera monitoring for behavior changes, unable to quantify respiratory parameters and vulnerable to environmental interference. This solution realizes dynamic monitoring of respiratory function by reusing motion sensors or adding a respiratory belt. The prior art only alarms when breathing completely stops, while this solution can capture early abnormalities in respiratory frequency and depth and give an early warning at the stage of tidal volume decline. Due to the lack of multi-modal data association in traditional systems, it is impossible to distinguish normal activities from real respiratory abnormalities, resulting in a high false alarm rate. This solution significantly reduces false triggers caused by nursing operations through cross-verification of kinematics and respiratory parameters.

[0057] Through the above technical solutions, this application solves the problems of late warning and high false alarm rate of the existing system for tachypnea events. Dual-modal data acquisition ensures the reliability of respiratory parameter acquisition, and the combined judgment conditions effectively distinguish physiological wheezing from pathological tachypnea. The intelligent association mechanism between abnormal events and secondary alarms enables medical staff to obtain early warnings before respiratory failure occurs and implement oxygen therapy or airway management measures in a timely manner.

[0058] Preferably, the solution of the pre-seizure detection module includes: surface electromyography sensors are attached to the limbs of the patient to capture the root mean square and high-frequency energy ratio of muscle discharge signals; when the root mean square of the electromyography signal suddenly increases by more than 2 times the standard deviation of the baseline within a 10-second window and the high-frequency energy exceeds 40%, a seizure warning is triggered.

[0059] Among them, the surface electromyography sensor refers to capturing the electrophysiological activities generated during muscle contraction through a bioelectric signal acquisition device, which can be specifically implemented by a differential electrode and a signal amplification circuit, and is used to directly monitor abnormal discharge activities of limb muscles.

[0060] The root mean square is a mathematical index for quantifying the time-domain energy of the electromyography signal amplitude, which can be specifically implemented by a sliding window integration algorithm and is used to characterize the instantaneous change of muscle contraction intensity.

[0061] The high-frequency energy ratio refers to the energy ratio of the high-frequency components after spectrum analysis of the electromyography signal, which can be specifically implemented by combining fast Fourier transform with frequency band energy accumulation calculation and is used to identify abnormal increases in nerve impulse frequency.

[0062] The 10-second window refers to the time interval for dynamic baseline update, which can be specifically implemented by using a circular buffer to store historical data and calculate the moving average, and is used to balance signal stability and detection timeliness.

[0063] 2 times the standard deviation is an abnormal determination threshold based on statistical principles, which can be specifically implemented by calculating the standard deviation range of historical data distribution and is used to exclude the interference of fluctuations in normal electromyography activities of individuals.

[0064] The 40% threshold refers to the setting of the critical value of the high-frequency energy ratio. Specifically, the high-frequency energy distribution characteristics of typical convulsion events can be verified through clinical data to distinguish the myoelectric signals generated by pathological convulsions and regular movements.

[0065] Specifically, surface electromyography sensors are deployed at the limb muscle groups of patients to collect raw myoelectric signals in real time. After the signals are amplified and filtered, the root mean square value and the high-frequency energy ratio within the current time window are calculated respectively. The root mean square value is calculated using the sliding window integration method to reflect the mutation of muscle contraction intensity; the high-frequency energy ratio extracts the energy ratio in the 100 - 500 Hz frequency band through spectrum analysis to characterize the abnormality of nerve discharge frequency. The dynamic baseline system continuously updates the root mean square average value and standard deviation within the past 10 seconds. When the real-time root mean square value exceeds 2 times the standard deviation of the baseline mean and the high-frequency energy ratio is detected to exceed 40% synchronously, it is determined as a convulsion precursor and an early warning is triggered.

[0066] Compared with the prior art, the traditional solution relies on infrared or cameras to monitor limb movements, unable to capture the changes in muscle microcurrents and vulnerable to environmental light or occlusions. This solution directly obtains muscle bioelectric signals through surface electromyography sensors and combines time-domain and frequency-domain two-dimensional feature analysis to accurately identify abnormal nerve and muscle discharges before convulsions, overcoming the problem of insufficient sensitivity of optical monitoring methods. The dynamic baseline threshold mechanism further avoids false positives caused by individual differences in myoelectric activity and improves the early warning specificity.

[0067] Through the above technical solutions, this application realizes the early identification of convulsion precursors in critically ill patients. Through the collaborative analysis of bioelectric signals and dynamic threshold determination, it solves the technical defect that traditional motion monitoring methods cannot detect microcurrent changes, effectively reduces the false negative rate caused by environmental interference or limb occlusion, and secures a critical treatment time window for medical staff.

[0068] Example 2

[0069] An infrared sensing and monitoring method for preventing falls in critically ill patients, comprising the following steps:

[0070] Synchronously collect infrared radiation, kinematic, and pressure data through a multi-modal sensor array;

[0071] Perform peak detection on the inertial measurement unit data in the edge computing module, segment the gait cycle and calculate the step length, step speed, and standard deviation of the trunk inclination angle, and perform cluster analysis on the infrared sensor data to identify sudden changes in the patient's movement direction or abnormal stillness;

[0072] Conduct risk assessment on the fusion features based on a random forest model and output the risk level;

[0073] When the risk level reaches the set threshold, initiate a secondary alarm and retrieve the associated ward video clips for review.

[0074] Among them, the multi-modal sensor array refers to a collaborative detection system that includes infrared, kinematic, and pressure sensors. Specifically, it can be implemented by combining pyroelectric infrared sensors, inertial measurement units, and piezoelectric pressure sensors, and is used to collect the physiological and behavioral data of patients from different dimensions. The edge computing module refers to the local data processing unit deployed at the sensor end. Specifically, it can be implemented by an embedded AI chip and a signal processing circuit, and is used to perform real-time processing on the original data to reduce transmission latency. The random forest model refers to an ensemble learning algorithm based on multiple decision trees. Specifically, it can be implemented by fine-tuning a pre-trained model framework combined with local data, and is used to improve the robustness of risk assessment through multi-feature fusion. The video clip retrieval refers to the function of intercepting the ward surveillance video associated with the alarm event. Specifically, it can be implemented by timestamp matching and video stream caching technology, and is used to provide a visual basis for alarm review.

[0075] Specifically, the infrared sensor of the multi-modal sensor array captures the patient's movement trajectory, the inertial measurement unit collects acceleration and attitude angle data, and the pressure sensor monitors the plantar pressure distribution. In the edge computing module, the data of the inertial measurement unit divides the gait cycle through peak detection, and calculates the dynamic change indexes of step length, walking speed, and trunk inclination angle. For example, a sudden drop in walking speed may indicate gait imbalance; the infrared data identifies sudden changes in movement direction or abnormal stationary states through clustering analysis. For example, the direction change before a sudden fall. The fused features are input into the random forest model for risk assessment, and the risk level is comprehensively judged through the voting mechanism of multiple decision trees. When the score exceeds the preset threshold, a secondary alarm is triggered and the ward surveillance video 10 seconds before the event is automatically retrieved for medical staff to verify the effectiveness of the alarm.

[0076] Compared with the existing technology, the traditional solution relies on a single sensor and cannot detect tiny motion changes. For example, when only using an infrared sensor, it is easily affected by environmental occlusion; the passive alarm lacks the ability to recognize precursors such as gait imbalance; cloud processing results in a response delay of more than 2 seconds; the failure to fuse multi-modal data easily misjudges nursing operations as falls. This method can complete the entire process from data collection to alarm triggering within 500 milliseconds through multi-sensor data fusion and edge real-time processing. At the same time, by combining the correlation analysis of pressure data and kinematic features, it can effectively distinguish normal nursing actions from real fall events.

[0077] Through the above technical solution, the present application realizes the active early warning of the precursors of falls in critically ill patients. The multi-modal data collaborative verification reduces the false alarms caused by environmental interference. The local processing of the edge computing module shortens the response time to meet the first aid requirements. The video review mechanism improves the credibility of the alarm information, and solves the technical defects of high latency, many false alarms, and insufficient early warning ability in the traditional solution.

[0078] Preferably, the risk assessment model is optimized through transfer learning. Based on the publicly available fall dataset, the model skeleton is pre-trained, and the gait and respiration data of the patients in the local hospital are used to fine-tune the classification boundary.

[0079] Among them, transfer learning refers to the technical process of applying the knowledge obtained from one task to another related task. Specifically, it can be achieved by using the mechanisms of parameter sharing and feature reuse. By retaining some layer parameters of the pre-trained model and re-training the output layer, the problem of insufficient data in the target domain is solved. The pre-trained model skeleton of the publicly available fall dataset refers to the basic neural network structure constructed using the standardized dataset. Specifically, the labeled videos and inertial sensor data in the UR Fall Detection dataset can be used for model initialization, and a general feature extraction ability is formed through large-scale sample training. Fine-tuning the classification boundary with the local hospital data refers to the adaptive adjustment of the model decision layer. Specifically, the gradient descent algorithm can be used to optimize the classifier weights, so that the model can adapt to the differences in sensor layout and the characteristics of the patient group distribution in the target hospital.

[0080] Specifically, in the pre-training stage, the diverse fall scenario data in the publicly available dataset are used to establish a spatio-temporal feature extraction network with generalization ability, and a basic discrimination ability is formed by analyzing the fall patterns of different body postures and speeds. In the fine-tuning stage, the real patient gait data under the ward layout of the target hospital are collected, including the step length variability in the rehabilitation corridor and the getting-out-of-bed action characteristics at different bed heights. The model can identify the abnormal behavior patterns in a specific environment through parameter fine-tuning. In the process of optimizing the classification boundary, the decision hyperplane is recalibrated with the local data to eliminate the data distribution shift caused by differences in patient age distribution and rehabilitation stage, so that the risk level assessment result matches the actual needs of the target hospital.

[0081] Compared with the prior art, traditional fall detection models usually directly use a single dataset for training, and the performance drops significantly due to differences in sensor types or changes in patient behavior characteristics when applied across institutions. However, this solution uses a two-stage transfer learning framework to effectively adapt to the data characteristics of the target hospital while maintaining the basic feature recognition ability, avoiding the cost burden caused by re-labeling a large amount of local data, and at the same time overcoming the problem of misalignment of the classification boundary when directly transferring the pre-trained model.

[0082] Through the above technical solutions, the present application can improve the adaptability of the risk assessment model among different medical institutions, and quickly realize model deployment and performance optimization without having to reconstruct a complete training set. Through the synergistic effect of feature extraction network sharing and dynamic adjustment of classification boundaries, the misjudgment rate caused by differences in hospital environments or characteristics of patient groups can be effectively reduced, ensuring the detection accuracy and reliability of the fall warning system in different application scenarios.

[0083] It should be noted that, in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusively, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0084] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An infrared sensing monitoring device for preventing falls of critically ill patients, characterized in that, Comprising: A multimodal sensor array, including a pyroelectric infrared sensor array arranged in the patient's activity area for detecting the spatio-temporal changes of human infrared radiation, a wearable inertial measurement unit configured on the patient's waist or wrist to collect triaxial acceleration, angular velocity and attitude angle data in real time, and a piezoelectric pressure sensor integrated on the hospital bed or the floor of the rehabilitation corridor to monitor the patient's body position distribution and the trajectory of the plantar pressure center; An edge computing module, including an embedded AI chip and a signal processing circuit for real-time filtering, feature extraction and time synchronization of multimodal sensor data, running a lightweight machine learning model, fusing infrared trajectories, IMU kinematic features and pressure distribution data, and generating a fall risk level; A hierarchical alarm module, which triggers the following actions according to the risk level: First-level warning: Prompt the patient of gait imbalance or shortness of breath through voice prompts or local lighting warnings; Second-level alarm: Push the patient's location, abnormal physiological parameter values and recommended intervention measures to the medical staff terminal.

2. The infrared sensing monitoring device for preventing falls of critically ill patients according to claim 1, wherein The arrangement method of the pyroelectric infrared sensor array is as follows: Directional infrared sensors are set at the four corners of the hospital bed to cover the off-bed detection area; Opposite infrared fences are installed on both sides of the rehabilitation corridor to form a continuous motion trajectory monitoring network.

3. The infrared sensing monitoring device for preventing falls of critically ill patients according to claim 2, wherein The machine learning model of the edge computing module includes: A time series classifier based on 1D-CNN, with the input being the acceleration standard deviation of the IMU, the change rate of the attitude angle and the stationary duration after impact; A dynamic threshold adjustment unit that adaptively corrects the alarm trigger condition according to the patient's historical activity data.

4. The infrared sensing and monitoring device for preventing falls of critically ill patients according to claim 1, wherein, Further includes a shortness of breath quantification module: By multiplexing the thoracic motion data of the IMU or an independent breathing belt sensor, calculate the respiratory rate and tidal volume; When RR>25 times / minute and TV<300ml, it is marked as a shortness of breath event and associated with the second-level alarm.

5. The infrared sensing monitoring device for preventing falls of critically ill patients according to claim 1, characterized in that, Further includes a premonitory seizure detection module that uses surface electromyography sensors attached to the patient's limbs to capture the root mean square and high-frequency energy ratio of muscle discharge signals, and when the RMS of the surface electromyography signal suddenly increases by more than 2 times the standard deviation of the baseline within a 10-second window and the high-frequency energy>40%, a seizure warning is triggered.

6. An infrared sensing monitoring method for preventing falls in critically ill patients, characterized in that, The infrared sensing monitoring for preventing falls in critically ill patients according to any one of claims 1-5 includes the following steps: S1. Synchronously collect infrared radiation, kinematic and pressure data through a multimodal sensor array; S2. Perform the following operations in the edge computing module: Perform peak detection on the IMU data, segment the gait cycle and calculate the step length, step speed and standard deviation of the trunk inclination angle; Perform clustering analysis on the infrared sensor data to identify sudden changes in the patient's movement direction or abnormal stillness; S3. Perform risk assessment on the fused features based on a random forest model and output a risk level of 0-10 points; S4. When the risk level≥7 points, start the second-level alarm and retrieve the associated ward video clip for medical staff to review.

7. The infrared sensing and monitoring method for preventing falls of critically ill patients according to claim 6, characterized in that, The risk assessment model based on the pre-trained model skeleton of the public fall dataset and fine-tuning the classification boundary with the patient's gait and breathing data of the local hospital is optimized by transfer learning.

Images