Patient care risk early warning system based on machine learning

Abstract

The present invention relates to the field of health data processing technology, and specifically to a patient care risk warning system based on machine learning, comprising: an acquisition end, a first processing end, and a second processing end, wherein the acquisition end collects image data of at least one target in a care environment area, and the first processing end analyzes the image data to determine the regional activity factor of each superpixel area in the image data, so as to determine the potential behavior risk coefficient of the target active area through the regional activity factor, and then the second processing end uses the potential behavior sub-risk coefficient to determine the adaptive edge retention scale coefficient of the target active area, so as to obtain the behavior feature data of the target using the adaptive edge retention scale coefficient to determine the behavior risk value of the target. That is, the invention implements effective nursing risk warning by carrying out risk warning of the target based on the behavior risk value.

Description

Technical Field

[0001] The present invention relates to the technical field of health data processing, and in particular to a patient care risk early warning system based on machine learning. Background Art

[0002] With the rise of personalized medicine, timely identification of high-risk patients during the care process can significantly reduce complications and medical costs. The patient care risk early warning system based on machine learning is a tool that uses artificial intelligence technology to predict and identify risk factors that may occur during patient care. The system helps medical staff identify potential risks in advance by analyzing a large amount of patient data, so that they can take corresponding measures to reduce risks and ultimately improve patient safety and care quality.

[0003] However, the existing patient care risk warning system based on machine learning mainly relies on real-time acquisition of patient behavior images to extract their behavior features and monitor their status. However, the patient's behavior pattern is complex, and in the image, the patient may be blocked by medical equipment in the environment or in close contact with people around him. This makes it difficult for traditional feature extraction methods to accurately capture the patient's behavior characteristics, which may lead to misjudgment of monitoring results. Therefore, a more efficient and reliable patient care risk warning system is urgently needed. Summary of the invention

[0004] In order to solve the technical problem that the feature extraction method in the existing patient care risk warning system based on machine learning is difficult to accurately capture the patient's behavioral characteristics, which may lead to misjudgment of the monitoring results, the purpose of the present invention is to provide a patient care risk warning system based on machine learning, by collecting image data of at least one target in the care environment area and analyzing the image data to determine the regional activity factor of each superpixel area in the image data, so as to determine the potential behavioral risk coefficient of the target active area through the regional activity factor, and then use the potential behavioral risk coefficient to determine the adaptive edge preservation scale coefficient of the target active area, so as to use the adaptive edge preservation scale coefficient to obtain the target's behavioral feature data to determine the target's behavioral risk value, so as to perform risk warning of the target based on the behavioral risk value, and realize effective nursing risk warning.

[0005] The technical solution adopted is specifically as follows: providing a patient care risk warning system based on machine learning, the patient care risk warning system based on machine learning comprising: an acquisition end, acquiring image data of at least one target in the acquisition area; a first processing end, acquiring at least one superpixel area in the image data, and determining the regional activity factor of each superpixel area; and acquiring at least one target active area in the image data, using the regional activity factor to determine the potential conflict coefficient of each target active area, and then using the potential conflict coefficient to determine the potential behavior risk coefficient of each target active area; a second processing end, using the potential behavior risk coefficient to determine the adaptive edge retention scale coefficient of each target active area, and using the adaptive edge retention scale coefficient to acquire the behavior feature data of the target, and then determining the behavior risk value of the target, and using the behavior risk value to perform risk warning for the target.

[0006] In one embodiment of the present invention, the acquisition end includes: a data acquisition module, which acquires multiple behavioral images of at least one of the targets in the acquisition area to form the image data; a data preprocessing module, which preprocesses the image data and stores the preprocessed image data in a local database, wherein the preprocessing includes at least denoising processing, brightness adjustment processing, and contrast adjustment processing.

[0007] In one embodiment of the present invention, the first processing end includes: a first processing module, which obtains at least one superpixel area in the image data, and obtains a displacement scale coefficient and a pixel value set of each superpixel area, and then uses the displacement scale coefficient and the pixel value set to determine the regional activity factor of each superpixel area; a second processing module, which obtains at least one target active area in the image data, and obtains regional parameters of the target active area, uses the regional activity factor to determine a potential conflict coefficient of each target active area, and uses the regional parameters to determine a regional integrity coefficient of each target active area; and then uses the potential conflict coefficient and the regional integrity coefficient to determine a potential behavior risk coefficient of each target active area.

[0008] In one embodiment of the present invention, obtaining at least one target active area in the image data includes: using the regional activity factor and a preset activity threshold to judge the superpixel area, and taking the superpixel area where the regional activity factor is greater than the activity threshold as the target active area.

[0009] In one embodiment of the present invention, the first processing module includes: a first acquisition component, which acquires at least one superpixel area in the image data, and acquires the pixel value set of each superpixel area, and acquires the displacement scale coefficient of each superpixel area in the adjacent image data, and also acquires the adjacent moment image set of each superpixel area; a first calculation component, which determines the regional activity factor of each superpixel area using the displacement scale coefficient, the pixel value set and the adjacent moment image set.

[0010] In one embodiment of the present invention, the operation of the first acquisition component is as follows: the image data is segmented into regions using a superpixel region segmentation method to obtain a plurality of local regions with similar pixel features, and the local regions are used as the superpixel regions; the centroid coordinate position of the superpixel region is used to obtain the first Euclidean distance between the centroid coordinates of the same superpixel region in adjacent frame images, and the first Euclidean distance is used as the displacement scale coefficient.

[0011] In one embodiment of the present invention, the second processing module includes: a second acquisition component, which acquires at least one target active area in the image data and acquires the optical flow vector of the target active area; a second calculation component, which uses the optical flow vector to determine the angle between the motion direction vector and the horizontal direction vector of each of the target active areas, and uses the angle and the area activity factor to determine the potential conflict coefficient of each of the target active areas; a third acquisition component, which acquires the area parameters of the target active area, wherein the area parameters at least include the edge curve perimeter and the area contour area; a third calculation component, which uses the edge curve perimeter and the area contour area to determine the area integrity coefficient of each of the target active areas; and a fourth calculation component, which uses the potential conflict coefficient and the area integrity coefficient to determine the potential behavior risk coefficient of each of the target active areas.

[0012] In one embodiment of the present invention, the fourth calculation component also includes: a first sub-acquisition component, which acquires the second Euclidean distance between the center of mass coordinates of different target active areas; a correction component, which uses the potential conflict coefficient and the area integrity coefficient to determine the corrected conflict coefficient of each target active area; and a first sub-calculation component, which uses the corrected conflict coefficient, the second Euclidean distance and the area activity factor to determine the potential behavior risk coefficient of each target active area.

[0013] In one embodiment of the present invention, the second processing end includes: a fourth acquisition component, which acquires the initial edge retention scale coefficient of the target active area; a fifth calculation component, which uses the initial edge retention scale coefficient and the potential behavior risk coefficient to determine the adaptive edge retention scale coefficient of each target active area; a sixth calculation component, which uses the adaptive edge retention scale coefficient to acquire the target edge curves of all the target active areas in the image data, and uses the target edge curves to acquire the behavior feature data of the target; an evaluation component, which uses the behavior feature data to determine the behavior risk value of the target in the real-time behavior, and uses the behavior risk value to issue a risk warning for the target.

[0014] In one embodiment of the present invention, the evaluation component also includes: a neural network training subcomponent, which uses the behavioral feature data for training to optimize the behavioral risk value of the target in real-time behavior; and a determination subcomponent, which uses the optimized behavioral risk value to issue a risk warning for the target.

[0015] The beneficial effects of the present invention are as follows: The patient care risk warning system based on machine learning in the present invention includes: an acquisition end, which acquires image data of at least one target in the acquisition area; a first processing end, which acquires at least one superpixel area in the image data and determines the regional activity factor of each superpixel area; and acquires at least one target active area in the image data, and determines the potential conflict coefficient of each target active area using the regional activity factor, and then determines the potential behavior risk coefficient of each target active area using the potential conflict coefficient; a second processing end, which determines the adaptive edge retention scale coefficient of each target active area using the potential behavior sub-risk coefficient, and acquires the behavior feature data of the target using the adaptive edge retention scale coefficient, and then determines the behavior risk value of the target, and uses the behavior risk value to perform risk warning for the target; that is, the present application acquires image data of at least one target in the care environment area through the acquisition end, and determines the regional activity factor of each superpixel area in the image data through the first processing end, so as to determine the potential behavior risk coefficient of the target active area through the regional activity factor, and then the second processing end determines the adaptive edge retention scale coefficient of the target active area using the potential behavior sub-risk coefficient, so as to acquire the behavior feature data of the target using the adaptive edge retention scale coefficient to determine the behavior risk value of the target. That is, the invention is to carry out risk warning for a target based on the behavior risk value, so as to realize effective nursing risk warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A schematic structural diagram of a patient care risk early warning system based on machine learning provided by the present invention. DETAILED DESCRIPTION

[0018] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation, structure, features and effects of a medical and nursing record electronic information management system proposed by the present invention in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0019] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0020] The specific scheme of the patient care risk early warning system based on machine learning provided by the present invention is described in detail below with reference to the accompanying drawings.

[0021] See also Figure 1 , which shows a structural schematic diagram of a patient care risk early warning system based on machine learning provided by the present invention.

[0022] like Figure 1 As shown, a patient care risk warning system 10 based on machine learning includes: an acquisition end 100, which acquires image data of at least one target in an acquisition area; a first processing end 200, which acquires at least one superpixel area in the image data and determines the regional activity factor of each superpixel area; and acquires at least one target active area in the image data, determines the potential conflict coefficient of each target active area using the regional activity factor, and then determines the potential behavior risk coefficient of each target active area using the potential conflict coefficient; a second processing end 300, determines the adaptive edge retention scale coefficient of each target active area using the potential behavior risk coefficient, and acquires the behavior feature data of the target using the adaptive edge retention scale coefficient, and then determines the behavior risk value of the target, and uses the behavior risk value to perform risk warning of the target.

[0023] Among them, the patient care risk warning system based on machine learning refers to a technology that uses algorithms to analyze patient data (such as images, behavioral patterns, etc.) to identify potential risks. The system automatically detects possible risk factors by monitoring and evaluating the patient's status in real time, so that medical personnel can issue risk warnings for the target based on the corresponding behavioral risk values. The acquisition end refers to the device or module responsible for acquiring image data, which can be a camera or sensor, and is mainly used to capture visual information of the target area.

[0024] Superpixel regions refer to small areas divided from an image using superpixel segmentation technology. These regions are determined based on the similarity of features such as color, texture, and brightness in the image, so that the pixels within each superpixel have high consistency, and there are clear boundaries between adjacent superpixels.

[0025] The regional activity factor is used to measure the activity of certain features in the superpixel region, reflecting the dynamic characteristics of the region under specific circumstances. The target active region refers to the region identified as the object of interest in the image, which is usually related to the patient's behavior or state.

[0026] The potential conflict coefficient is used to assess the degree of possible conflict between target active areas and helps identify potential risks.

[0027] The potential behavior risk coefficient is used to indicate the risk of adverse behavior that may be caused by the target active area. The adaptive edge preservation scale coefficient is an image processing parameter that aims to retain the edge information of the image when extracting features to ensure the accuracy of recognition. The behavior feature data is the data extracted by analyzing the target active area, describing the patient's behavior pattern and state. The behavior risk value is a quantitative indicator obtained through comprehensive determination, indicating the risk level of a specific target, and is used to trigger the early warning mechanism.

[0028] In this embodiment, the patient care risk warning system based on machine learning realizes risk assessment through multiple processing terminals. First, the acquisition terminal 100 is responsible for acquiring the image data of the target. Then, the first processing terminal 200 analyzes these images, identifies the superpixel area, determines the regional activity factor, and evaluates the potential conflict coefficient and potential behavior risk coefficient of each target. Next, the second processing terminal 300 uses the potential behavior risk coefficient to determine the adaptive edge retention scale coefficient to extract the behavior feature data of the target, and finally determines the behavior risk value. That is, the system can accurately identify the potential risks of patients, extract the behavior feature data of the target, and determine the behavior risk value, so as to achieve early warning, improve the quality of care, reduce the possibility of risk occurrence, and enhance the safety and effectiveness of patient care.

[0029] In some embodiments, the acquisition end may include a data acquisition module and a data preprocessing module.

[0030] The acquisition end includes: a data acquisition module, which acquires multiple behavioral images of at least one target in the acquisition area to form image data; a data preprocessing module, which preprocesses the image data and stores the preprocessed image data in a local database, wherein the preprocessing includes at least denoising processing, brightness adjustment processing, and contrast adjustment processing.

[0031] The data acquisition module may refer to the installation of a high-resolution camera in the target patient's care environment, which is capable of real-time recording or timed shooting, for example, once every 1 second, to ensure coverage of the patient's key areas and capture their diverse behavioral images. The arrangement of these cameras ensures the monitoring of important behaviors, such as the patient's activities, behavioral changes, etc., thereby providing a rich data basis for subsequent analysis.

[0032] The data preprocessing module refers to the preprocessing that needs to be performed before the captured patient behavior images enter the local database to improve the image quality and ensure the accuracy of the analysis.

[0033] The preprocessing steps may include denoising, brightness adjustment, and contrast adjustment. Denoising refers to improving image clarity by removing interfering noise from the image. Brightness adjustment refers to adjusting the brightness of the image so that images under different lighting conditions are easier to analyze. Contrast adjustment refers to enhancing the contrast of the image so that important features are more prominent, facilitating subsequent feature extraction and analysis. The preprocessed image data can be transmitted to a local database in real time through wireless network technology for real-time monitoring and analysis.

[0034] Specifically, the acquisition end includes a data acquisition module and a data preprocessing module. The data acquisition module is used to obtain multiple behavioral images of the target patient's area, which provide basic data for subsequent analysis. The data preprocessing module performs denoising, brightness adjustment, and contrast adjustment on the acquired images to improve image quality and ensure the accuracy and availability of information. Using modern image processing technology and wireless communication technology, an efficient and accurate patient monitoring environment is created, thereby supporting medical staff to identify and respond to potential risks more quickly during the care process.

[0035] In some embodiments, the first processing end may include a first processing module and a second processing module.

[0036] The first processing module is used to obtain at least one superpixel region in the image data, and to obtain a displacement scale coefficient and a pixel value set of each superpixel region, and then determine the regional activity factor of each superpixel region using the displacement scale coefficient and the pixel value set.

[0037] The pixel value set generally refers to a set of data representing the color information of each pixel in the image.

[0038] Specifically, the first processing module may include: a first acquisition component and a first calculation component.

[0039] The working principle of the first processing module is as follows:

[0040] First, the first acquisition component acquires at least one superpixel region in the image data, acquires a set of pixel values ​​of each superpixel region, acquires a displacement scale coefficient of each superpixel region in adjacent image data, and acquires a set of adjacent moment images of each superpixel region.

[0041] The operation of the first acquisition component is as follows:

[0042] First, the image data is segmented using the superpixel region segmentation method to obtain multiple local regions with similar pixel features, and the local regions are used as superpixel regions. Among them, pixel features refer to the color, brightness, texture and other attributes of each pixel. The purpose of superpixel segmentation is to divide the image data into multiple local regions, and the pixels in these local regions have high similarity in color and texture. After this processing, the computational complexity of subsequent processing can be reduced while retaining the structural information of the image.

[0043] Then, the centroid coordinate position of the superpixel area is used to obtain the first Euclidean distance between the centroid coordinates of the same superpixel area in adjacent frame images, and the first Euclidean distance is used as the displacement scale coefficient, wherein the displacement scale coefficient is obtained by determining the Euclidean distance between the centroid coordinates of the superpixel area in adjacent frame images and the centroid coordinates of the superpixel area in adjacent frame images, reflecting the temporal activity of the superpixel area.

[0044] Secondly, based on the first calculation component, the regional activity factor of each superpixel region is determined using the displacement scale coefficient, the pixel value set and the adjacent time image set, wherein the regional activity factor of the superpixel region reflects the motion activity and potential risk of the regional activity factor of the superpixel region.

[0045] The following is a calculation formula for determining the regional activity factor of each superpixel region based on the displacement scale coefficient, the pixel value set and the adjacent time image set:

[0046]

[0047] In the formula, Represents the image data Superpixel area, Represents the The number of images in all adjacent moments of a superpixel region, Represents The collection A set of images at adjacent moments, Represents the The ath moment is located on the right. represents the left moment adjacent to a, represents the image at the ath moment The set of pixel values ​​in the superpixel region, represents the image at the bth moment The set of pixel values ​​in the superpixel region, represents the mean square error expression, Represents the The image at the adjacent time The displacement scale coefficient of the superpixel area is Represents the The regional activity factor of the superpixel area; the mean square error is a commonly used indicator for evaluating the performance of the regression model. The mean square error is used to measure the difference between two values. The smaller the MSE value, the smaller the difference between the two values. The displacement scale coefficient reflects the change amplitude of the centroid of the superpixel area in adjacent frames.

[0048] Specifically, the first processing module first extracts at least one superpixel region from the image data. Then, the first processing module determines the displacement scale coefficient of the superpixel region, and at the same time, the first processing module also obtains the pixel value set of the superpixel region. Using the two data of the displacement scale coefficient and the pixel value set of the superpixel region, the first processing module can determine the regional activity factor of each superpixel region, and then evaluate the motion activity and potential risk of the region. This process not only improves the monitoring ability of the patient care environment, but also provides a scientific basis for subsequent risk intervention.

[0049] In addition, the second processing module is used to obtain at least one target active area in the image data, and obtain the regional parameters of the target active area, use the regional activity factor to determine the potential conflict coefficient of each target active area, and use the regional parameters to determine the regional integrity coefficient of each target active area; and then use the potential conflict coefficient and the regional integrity coefficient to determine the potential behavior risk coefficient of each target active area.

[0050] The second processing module may include: a second acquisition component, a second calculation component, a third acquisition component, a third calculation component and a fourth calculation component.

[0051] The working principle of the second processing module is as follows:

[0052] First, the second acquisition component acquires at least one target active area in the image data, and acquires an optical flow vector of the target active area.

[0053] The target active area is usually the part of the image that shows dynamic behavior or activity. The optical flow vector represents the movement direction and speed of the pixels in the target active area between consecutive image frames. By obtaining these optical flow vectors, more complex motion analysis can be performed, such as behavior recognition, anomaly detection, trajectory prediction, etc. For example, in the patient care risk warning system, by analyzing the patient's movement pattern, potential risk behaviors such as falling, rapid movement, etc. can be identified and timely warnings can be issued.

[0054] Secondly, the second calculation component determines the angle between the motion direction vector and the horizontal direction vector of each target active region using the optical flow vector, and determines the potential conflict coefficient of each target active region using the angle and the region activity factor.

[0055] Among them, the mean of all optical flow vectors of the active area is the motion direction vector of the target active area. The angle between the motion direction vector of the target active area and the horizontal direction vector can reflect the motion trend of the target. The potential conflict coefficient can indicate the degree of conflict that may exist between the target active area and other active areas. The horizontal direction vector refers to a unit vector along the horizontal axis in a two-dimensional or three-dimensional coordinate system, where the horizontal axis is usually the x-axis. The horizontal direction vector represents a component only in the horizontal direction in space and no component in other directions.

[0056] The following is the calculation process for determining the potential conflict coefficient of each target active area based on the angle and the regional active factor:

[0057]

[0058] In the formula, Represents the target image data Target active areas, Represents the The regional activity factor of the target active area, Represents the target image data except The set of all target active areas outside the target active areas, Represents The collection Target active areas, Represents the The regional activity factor of the target active area, Represents the The angle between the moving direction vector of the target active area and the horizontal direction, Represents the The angle between the moving direction vector of the target active area and the horizontal direction, Represents the The potential conflict coefficient of the target active area.

[0059] Next, the third acquisition component acquires the area parameters of the target active area.

[0060] The region parameters at least include the perimeter of the edge curve and the contour area of ​​the region. The perimeter of the edge curve can reflect the shape complexity of the region, while the contour area can represent the size of the region.

[0061] Specifically, the third acquisition component can use the Canny edge detection algorithm to extract the edge curve of each target active area in the current moment image, and obtain the edge curve perimeter B of the target active area in the current moment and the previous moment image, and the area contour area X.

[0062] Next, the third calculation component is used to determine the region integrity coefficient of each target active region by using the edge curve perimeter and the region contour area.

[0063] Among them, the regional integrity coefficient is usually used to measure the shape integrity of the region. The more regular the shape, the higher the coefficient.

[0064] Specifically, the calculation formula of the regional integrity coefficient of the target active area is as follows:

[0065]

[0066] In the formula, Represents the current moment of patient care risk early warning by the machine learning-based patient care risk early warning system, Represents the moment before the current moment. Represents the The target active area is The perimeter of the edge curve of the region in the image at the moment, Represents the The target active area is The perimeter of the edge curve of the region in the image at the moment, Represents the The target active area is The contour area of ​​the region in the image at the moment, Represents the The target active area is The contour area at the moment, Represents the The regional integrity coefficient of the target active area.

[0067] Finally, the fourth calculation component determines the potential behavior risk coefficient of each target active area using the potential conflict coefficient, the area integrity coefficient and the area activity factor.

[0068] Among them, the potential behavior risk coefficient of the target active area can be used to determine whether the behavior in the area may lead to potential problems, such as conflicts, collisions, etc.

[0069] In general, when there are many patients with high activity levels around the target active patient, due to the highly active and unpredictable behavior of these patients, this environment may lead to accidental collisions, conflicts or other safety hazards, and the potential risk factor faced by the target active patient will increase significantly. To this end, the fourth computing component may also include a first sub-acquisition component, a correction component and a first sub-computing component.

[0070] The specific working principle of the fourth computing component is as follows:

[0071] First, the first sub-acquisition component acquires the second Euclidean distance between the centroid coordinates of different target active regions.

[0072] The first sub-acquisition component is used to acquire the second Euclidean distance between different target active regions. The second Euclidean distance helps to identify the relative positions between regions and provide information on regional distribution.

[0073] Next, the correction component determines a corrected conflict coefficient for each target active region using the potential conflict coefficient and the region integrity coefficient.

[0074] Among them, the modified conflict coefficient is used to evaluate the possible interference or competition relationship between different target active areas, ensuring that the dynamics of the area can be more accurately reflected in subsequent calculations.

[0075] Specifically, based on the potential conflict coefficient and the area integrity coefficient, the calculation formula for determining the modified conflict coefficient of each target active area is as follows:

[0076]

[0077] In the formula, Represents the The potential conflict coefficient of the target active area, Represents the The regional integrity coefficient of the target active area, Represents the Corrected conflict coefficient for target active areas, Represents the target image data target active area.

[0078] Finally, the first sub-computation component determines the potential behavior risk coefficient of each target active area using the modified conflict coefficient, the second Euclidean distance and the area activity factor.

[0079] Among them, the potential behavior risk coefficient is a key indicator for assessing the behavioral impact that may occur in a region under specific circumstances.

[0080] Based on the modified conflict coefficient, the second Euclidean distance and the regional activity factor, the calculation formula for determining the potential behavior risk coefficient of each target active area is as follows:

[0081]

[0082] In the formula, Represents the The target active area corrects the conflict coefficient, Represents the The set of m target active areas closest to the target active area, Represents The first in the collection Target active areas, Represents the The regional activity factor of the target active area, Represents the The target active area and The second Euclidean distance between the centroid coordinates of the target active area, Represents the The potential behavioral risk coefficient of the target active area, where the value of m can be set to 8.

[0083] In actual operation, according to the potential behavioral risk coefficient of the patient area in the overall image of the nursing environment, when performing patient behavior feature extraction based on edge segmentation, an adaptive non-maximum suppression edge preservation parameter is set to help obtain more edge details of the local area under a complex background, thereby better extracting and distinguishing the differences between individual areas of the patient.

[0084] In the existing Canny edge detection image segmentation method, the image is first smoothed by a Gaussian filter to reduce noise. Next, the Sobel operator is used to calculate the gradient strength and direction of each point in the image. Then, local maximum suppression is performed in the gradient direction to obtain the precise location of the edge. Finally, the non-maximum suppression method is applied to determine the strength of the edge and mark the edge point as a strong edge or a weak edge.

[0085] However, in actual situations, there are usually a large number of overlapping edges in the occluded area of ​​the patient, which makes the traditional non-maximum suppression method likely to ignore the detailed edges of some occluded areas, resulting in the loss of detailed information in the local area.

[0086] Furthermore, the second processing end may also include a fourth acquisition component, a fifth calculation component, a sixth calculation component and an evaluation component, and the fourth acquisition component, the fifth calculation component, the sixth calculation component and the evaluation component each assume a specific function.

[0087] The specific working principle is as follows:

[0088] First, the fourth acquisition component acquires an initial edge-preserving scale factor of the target active region.

[0089] The edge preservation scale factor is an adjustment parameter used to balance edge preservation and smoothing effects in image processing; the initial edge preservation scale factor can be calculated by performing statistical analysis on the edge strength of the target active area.

[0090] The specific process of the fourth acquisition component acquiring the initial edge preservation scale coefficient of the target active area is as follows:

[0091] First, a Gaussian filter is used to smooth the target active area.

[0092] The smoothing process is performed on the target active area to reduce the noise in the target active area image so that the subsequent edge detection is more accurate.

[0093] Next, the Sobel operator is used to calculate the gradient intensity and direction of each pixel in the target active area, and the gradient intensity map is obtained.

[0094] Among them, gradient strength is an indicator used in image processing to measure the degree of change of a part of an image. Specifically, gradient strength indicates the speed and magnitude of the change in brightness or color around a pixel in the image. Areas with high gradient strength usually correspond to edges or details, while areas with low gradient strength indicate small changes in brightness, usually flat or uniform areas. The gradient strength map is used to represent the degree of change of each pixel in the target active area.

[0095] Next, in the gradient intensity map, a threshold is set to extract the edge curve.

[0096] Among them, edge curves are formed by connecting pixels with high intensity and usually represent important structures in the image.

[0097] Specifically, the process of extracting edge curves in the gradient intensity map is as follows: First, non-maximum suppression is applied to refine the edges, retaining the pixels with the largest gradient intensity and setting the rest of the pixels to 0. This step is done by checking the gradient direction of each pixel and setting it to 0 if it is not a local maximum. Next, a double threshold method is used to determine strong and weak edges: a strong edge refers to a gradient intensity above a high threshold, a weak edge is a gradient intensity between a high threshold and a low threshold, and a non-edge is a gradient intensity below a low threshold. Finally, the final edge curve is formed by connecting the strong edge with the connected weak edges, and only the weak edges connected to the strong edges will be retained.

[0098] Next, for each extracted edge curve, its overall gradient strength is calculated. The strength value of the edge can be obtained by averaging or weighted summing the gradient strengths of all pixels on the edge curve.

[0099] Then, according to the strength value of each edge, all edges are sorted to form an edge strength sequence.

[0100] Finally, by analyzing the edge strength sequence, the initial edge preservation scale factor can be set. The edge curve of is regarded as the final edge curve of the image, and the initial edge preservation scale coefficient is calculated by using the edge intensity mean plus a certain multiple of the standard deviation.

[0101] Secondly, the fifth calculation component determines an adaptive edge preservation scale factor for each target active region using the initial edge preservation scale factor and the potential behavior risk factor.

[0102] Among them, the target edge curve is a curve that describes the contour of the target active area, which can contain the shape and motion information of the target area. By analyzing the curve of the target active area contour, the target's behavioral feature data, such as movement direction, speed and acceleration, can be extracted.

[0103] Specifically, based on the initial edge preservation scale coefficient and the potential behavior risk coefficient, the calculation formula for determining the adaptive edge preservation scale coefficient of each target active area is as follows:

[0104]

[0105] In the formula, Represents the The potential behavioral risk coefficient of the target active area, Represents the The initial edge preservation scale factor of the target active area, Represents the The adaptive edge-preserving scaling factor of the target active region, Represents the target image data target active area.

[0106] Next, the sixth computing component uses the adaptive edge preservation scale coefficient to obtain target edge curves of all target active areas in the image data, and uses the target edge curves to obtain target behavior feature data.

[0107] Among them, the target's behavioral characteristic data reflects the target's behavior pattern in a specific environment.

[0108] Finally, the evaluation component uses the behavioral feature data to determine the behavioral risk value of the target in real-time behavior, and uses the behavioral risk value to issue a risk warning for the target.

[0109] Among them, the behavioral risk value of the target in real-time behavior is used to determine whether the target's behavior has potential risks, and issue corresponding risk warnings accordingly.

[0110] Furthermore, the evaluation component also includes: a neural network training subcomponent, which uses behavioral feature data for training to optimize the behavioral risk value of the target in real-time behavior; and a determination subcomponent, which uses the optimized behavioral risk value to provide risk warning for the target.

[0111] Among them, the main task of the neural network training subcomponent is to train the neural network model using the extracted behavioral feature data. First, the system matches the historical behavioral feature data with the known risk values ​​(usually from previous clinical data or expert evaluations) to build a training data set. During the training process, the neural network adjusts internal weights and biases to minimize the difference between the predicted risk value and the actual risk value. This process uses the backpropagation algorithm and, after multiple iterations, optimizes the parameters of the model so that it can accurately identify and evaluate new behavioral feature data.

[0112] The determination subcomponent uses the optimized neural network model to input new behavioral feature data and predict the corresponding risk value. This risk value is quantified between 1 and 10, where 1 represents low risk and 10 represents high risk. Based on this risk value, the system will trigger the corresponding risk warning and notify medical staff to take timely intervention measures. In this way, medical staff can respond in the first time to ensure the safety and quality of care of patients.

[0113] That is, the evaluation component constructs an adaptive real-time risk assessment mechanism through the training of neural networks and the determination of the behavioral risk value of the target in real-time behavior, which can effectively improve the safety and efficiency of patient care.

[0114] It should be noted that the sequence of the above embodiments of the present invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0115] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A patient care risk early warning system based on machine learning, characterized in that: The patient care risk early warning system based on machine learning includes: The acquisition end obtains image data of at least one target in the acquisition area; The first processing end obtains at least one superpixel region in the image data and determines a regional activity factor of each superpixel region; and obtains at least one target active region in the image data, determines a potential conflict coefficient of each target active region using the regional activity factor, and further determines a potential behavior risk coefficient of each target active region using the potential conflict coefficient; The second processing end uses the potential behavior risk coefficient to determine an adaptive edge retention scale coefficient of each active area of ​​the target, and uses the adaptive edge retention scale coefficient to obtain the behavior feature data of the target, thereby determining the behavior risk value of the target, and using the behavior risk value to issue a risk warning for the target; The first processing end includes: A first processing module is configured to obtain at least one superpixel region in the image data, and obtain a displacement scale coefficient and a pixel value set of each superpixel region, and then determine a regional activity factor of each superpixel region by using the displacement scale coefficient and the pixel value set; The second processing module obtains at least one target active area in the image data, obtains regional parameters of the target active area, determines a potential conflict coefficient of each target active area using the regional activity factor, and determines a regional integrity coefficient of each target active area using the regional parameters; and then determines a potential behavior risk coefficient of each target active area using the potential conflict coefficient and the regional integrity coefficient; The acquiring at least one target active area in the image data comprises: The superpixel region is judged by using the regional activity factor and a preset activity threshold, and the superpixel region whose regional activity factor is greater than the activity threshold is used as the target active region; The first processing module includes a first acquisition component, and the operation of the first acquisition component is as follows: Performing region segmentation on the image data by using a superpixel region segmentation method to obtain a plurality of local regions with similar pixel features, and using the local regions as the superpixel regions; Using the centroid coordinate position of the superpixel region, obtaining a first Euclidean distance between the centroid coordinates of the same superpixel region in adjacent frame images, and using the first Euclidean distance as the displacement scale coefficient; The second processing module comprises: A second acquisition component is used to acquire at least one target active region in the image data and acquire an optical flow vector of the target active region; A second calculation component determines the angle between the motion direction vector and the horizontal direction vector of each target active region using the optical flow vector, and determines the potential conflict coefficient of each target active region using the angle and the region activity factor; A third acquisition component is used to acquire the region parameters of the target active region, wherein the region parameters at least include the perimeter of the edge curve and the region contour area; A third calculation component determines a region integrity coefficient of each of the target active regions using the edge curve perimeter and the region contour area; A fourth calculation component determines a potential behavior risk coefficient of each target active area by using the potential conflict coefficient, the area integrity coefficient and the area activity factor; The calculation formula of the regional activity factor is: In the formula, Represents the image data Superpixel area, Represents the The number of images in all adjacent moments of a superpixel region, Represents The collection A set of images at adjacent moments, Represents the The ath moment is located on the right. represents the left moment adjacent to a, represents the image at the ath moment The set of pixel values ​​in the superpixel region, represents the image at the bth moment The set of pixel values ​​in the superpixel region, represents the mean square error expression, Represents the The image at the adjacent time The displacement scale coefficient of the superpixel area is Represents the The regional activity factor of the superpixel region; The calculation formula of the potential conflict coefficient is: In the formula, Represents the target image data Target active areas, Represents the The regional activity factor of the target active area, Represents the target image data except The set of all target active areas outside the target active areas, Represents The collection Target active areas, Represents the The regional activity factor of the target active area, Represents the The angle between the moving direction vector of the target active area and the horizontal direction vector, Represents the The angle between the moving direction vector of the target active area and the horizontal direction vector, Represents the Potential conflict coefficient of target active areas; The calculation formula of the regional integrity coefficient is: In the formula, represents the current moment of patient care risk early warning by the patient care risk early warning system based on machine learning, Represents the moment before the current moment. Represents the The target active area is The perimeter of the edge curve of the region in the image at the moment, Represents the The target active area is The perimeter of the edge curve of the region in the image at the moment, Represents the The target active area is The contour area of ​​the region in the image at the moment, Represents the The target active area is The contour area at the moment, Represents the The regional integrity coefficient of the target active area; The calculation formula of the adaptive edge preservation scale coefficient is: In the formula, Represents the The potential behavioral risk coefficient of the target active area, Represents the The initial edge preservation scale factor of the target active area, Represents the The adaptive edge-preserving scaling factor of the target active region, Represents the target image data target active area.

2. The patient care risk early warning system based on machine learning according to claim 1 is characterized in that: The acquisition end includes: A data acquisition module, which acquires a plurality of behavior images of at least one of the targets in the acquisition area to form the image data; The data preprocessing module preprocesses the image data and stores the preprocessed image data in a local database, wherein the preprocessing at least includes denoising, brightness adjustment, and contrast adjustment.

3. The patient care risk early warning system based on machine learning according to claim 1 is characterized in that: The fourth computing component further includes: A first sub-acquisition component acquires a second Euclidean distance between the centroid coordinates of different target active areas; A correction component, using the potential conflict coefficient and the region integrity coefficient, determines a corrected conflict coefficient for each of the target active regions; The first sub-computing component determines the potential behavior risk coefficient of each target active area by using the modified conflict coefficient, the second Euclidean distance and the area activity factor.

4. The patient care risk early warning system based on machine learning according to claim 1 is characterized in that: The second processing end includes: A fourth acquisition component is used to acquire an initial edge preservation scale coefficient of the target active area; A fifth calculation component, using the initial edge preservation scale coefficient and the potential behavior risk coefficient, determines an adaptive edge preservation scale coefficient for each of the target active regions; A sixth computing component, using the adaptive edge preservation scale coefficient to obtain target edge curves of all target active areas in the image data, and using the target edge curves to obtain behavioral feature data of the target; The evaluation component determines the behavior risk value of the target in real-time behavior using the behavior feature data, and issues a risk warning to the target based on the behavior risk value.

5. The patient care risk early warning system based on machine learning according to claim 4 is characterized in that: The evaluation component further includes: A neural network training subcomponent is used to train the behavior feature data to optimize the behavior risk value of the target in real-time behavior; Determine the subcomponents and use the optimized behavior risk value to issue a risk warning for the target.

Images