Non-contact automated noise classification method and system

Abstract

This invention provides a non-contact automated agitation classification method and system, wherein the method includes the following steps: Step A: collecting historical data of multiple patients, including historical images; Step B: training a machine learning model using the historical images to generate an agitation classification model; Step C: obtaining an image of a patient to be analyzed, extracting at least one agitation feature from the image, and analyzing the agitation feature using the agitation classification model to determine whether the patient is in an agitated state; wherein the agitation feature is a facial movement feature, a trunk movement feature, a lower limb movement feature, and / or a restraint behavior feature. Accordingly, this invention can achieve continuous and accurate monitoring of patient agitation, assisting medical staff in adjusting sedation strategies in a timely manner, improving patient safety, and reducing the occurrence of adverse events.

Description

Technical Field

[0001] This invention relates to image detection technology, and more particularly to a non-contact automated noise classification method and system. Prior Technology

[0002] Patients in the intensive care unit often exhibit agitation or excessive sedation due to illness, endotracheal intubation, clinical procedures, and environmental factors, which may interfere with treatment outcomes and affect clinical safety.

[0003] Therefore, the Richmond Agitation-Sedation Scale (RASS), the Sedation-Agitation Scale (SAS), and the Motor Activity Assessment Scale (MAAS) are commonly used in clinical practice to assess patients' agitation levels. However, these scales all rely on the subjective judgment of nursing staff, making it difficult to ensure consistency. Furthermore, manual assessment not only increases the nursing burden but also prevents continuous monitoring, limiting the frequency of assessments and thus affecting the accuracy of sedation control. Insufficient or excessive sedation may even increase patient risk, impacting treatment safety and prognosis.

[0004] Therefore, developing more objective, timely, and continuously monitorable sedation assessment tools has become an important clinical issue. Summary of the Invention

[0005] The main objective of this invention is to provide a non-contact automated agitation classification method and system, which can achieve continuous and accurate monitoring of patient agitation, assist medical staff in adjusting sedation strategies in a timely manner, improve patient safety, and reduce the occurrence of adverse events.

[0006] Therefore, in order to achieve the above objectives, the present invention provides a non-contact automated agitation classification method, comprising the following steps: Step A: collecting historical data of multiple patients, the historical data including historical images; Step B: training a machine learning model with the historical images to generate an agitation classification model; Step C: obtaining an image of a patient to be analyzed, extracting at least one agitation feature from the image, and analyzing the agitation feature using the agitation classification model to determine whether the patient to be analyzed is in an agitated state; wherein, the agitation feature is a facial movement feature, a trunk movement feature, a lower limb movement feature, and / or a restraint behavior feature.

[0007] In one embodiment, step B further includes a data filtering procedure to remove those historical data that do not meet a filtering criterion, and to use the historical images of the historical data that meet the filtering criterion as training data.

[0008] In one embodiment, the screening criterion is to check whether the historical RASS score included in the patient's historical data is within a preset range, and use it as the judgment criterion. When the historical RASS score does not fall within the preset range, the historical data is deleted.

[0009] In one embodiment, the preset range is a set of values ​​including -2, -1, 0, 1, 2, 3, 4.

[0010] In one embodiment, step B further includes a preprocessing procedure, which includes an image segmentation procedure, which segments each historical image into multiple master segments, and then segments each master segment into multiple sub-segments, wherein the duration of each sub-segment is less than the duration of each master segment.

[0011] In one embodiment, the duration of each parent segment is 30 seconds, and the duration of each child segment is 2 seconds.

[0012] In one embodiment, the preprocessing procedure further includes a noise removal procedure to exclude all parent segments containing noise.

[0013] In one embodiment, the noise removal procedure determines whether the duration of noise in each parent segment is less than a preset time. If the duration of noise exceeds the preset time, the parent segment is deleted.

[0014] In one embodiment, the preset time is 10 seconds, and the noise refers to camera shake or interference from medical personnel.

[0015] In one embodiment, the preprocessing procedure further includes a labeling procedure for identifying and labeling at least one agitation target in each of the parent segments and each of the child segments, wherein the agitation target is the head, torso, lower limbs, or restraint behavior.

[0016] In one embodiment, the preprocessing procedure further includes a background separation procedure for removing background from each of the parent segments and each of the child segments.

[0017] In one embodiment, the preprocessing procedure further includes a data quantization procedure for analyzing the agitation targets in each of the parent segments and each of the child segments, and recording and quantifying the degree of movement of the head, torso, or lower limbs when the agitation target is the head, torso, or lower limbs.

[0018] In one embodiment, step B further includes a data classification program that analyzes the parent segments or the sub-segments and marks each parent segment or sub-segment as requiring attention or not requiring attention based on the analysis results.

[0019] In one embodiment, the data classification procedure includes a threshold classification procedure, which classifies the degree of movement into three levels: no movement (1 point), in-bed movement (2 points), and obvious out-of-bed movement (3 points). The degree of movement of the head, torso, and lower limbs in these sub-segments is scored accordingly and summed to calculate a total score, which ranges from 3 to 9 points. When the total score exceeds a preset threshold, the sub-segment is marked as requiring attention; when the total score does not exceed the preset threshold, the sub-segment is marked as not requiring attention.

[0020] In one embodiment, the preset threshold is 5.

[0021] In one embodiment, the data classification procedure includes a random forest classification procedure, which obtains the degree of movement of the head, torso, and lower limbs in the sub-segments as training samples, and uses the Random Forest (RF) algorithm to construct the agitation classification model as a random forest classification model, and can perform judgment analysis to label each sub-segment as needing attention or not needing attention.

[0022] In one embodiment, the data classification procedure includes a long short-term memory network classification procedure, which obtains the degree of movement of the head, torso and lower limbs in the parent segments as training samples, and constructs the agitation classification model as a long short-term memory network classification model based on long short-term memory (LSTM), and can perform judgment analysis, and label each parent segment as needing attention or not needing attention.

[0023] Furthermore, the present invention also provides a non-contact automated noise classification system, including a computing device for performing the aforementioned operations.

[0024] Therefore, the present invention has at least the following characteristics:

[0025] 1. Automated assessment of patient agitation: It can automatically and continuously assess whether a patient is agitated without relying on manual observation, and can also reduce errors and biases in human judgment, and improve the accuracy and efficiency of assessment.

[0026] 2. Non-contact monitoring: Features are extracted through image analysis, avoiding the limitations of traditional methods that require contact with patients, reducing patient discomfort and potential infection risks, and is suitable for critically ill patients who require long-term monitoring.

[0027] 3. Multi-feature comprehensive judgment: This invention does not rely on a single movement feature, but integrates multiple features such as face, trunk, lower limbs and restraint behavior for classification and analysis, which can more comprehensively capture the multi-dimensional information of the patient's agitation and improve the accuracy of agitation state judgment.

[0028] 4. Historical data as training samples: By using historical image data of multiple patients to train the machine learning model, the generalization ability and accuracy of the model are ensured, enabling it to be effectively applied in real clinical settings.

[0029] 5. Assisting clinical decision-making: By automatically analyzing patients' agitation status, physicians can obtain relevant information in a timely manner, thereby assisting clinical decision-making, identifying potential agitation risks in advance, reducing the workload of medical staff, and improving the quality of patient care. Simple Explanation of the Diagram

[0030] Figure 1 is a flowchart of the non-contact automated noise classification method of the present invention. Figure 2 is a schematic diagram of background separation. Figure 3 is a schematic diagram of the average movement values ​​of the head, trunk, or lower limbs at different time points. Figure 4 is a box plot showing the degree of head movement. Figure 5 is a box plot showing the degree of movement of the torso. Figure 6 is a box plot showing the degree of movement of the lower limbs. Figure 7 shows the box plots for those who need to pay attention and those who do not after the summation calculation. Figure 8 shows the ROC curves for different models. Figure 9 is a schematic diagram of the confusion matrix for the lower limbs. Figure 10 is a schematic diagram of the confusion matrix of the torso. Figure 11 is a schematic diagram of the confusion matrix of the head. Figure 12 shows the trend of continuous monitoring of various parts of an actual patient every 2 seconds. Figure 13 shows the trend of continuous monitoring of various parts of another actual patient every 2 seconds. Figure 14 shows the trend of various parts of the body when the patient's unrestrained behavior is continuously monitored every 2 seconds without attention. Figure 15 shows the trend of various body parts when patients are continuously monitored for restrained behavior every 2 seconds without needing to pay attention. Figure 16 shows the trend of various parts of the body when patients are continuously monitored for unrestrained behavior every 2 seconds. Figure 17 shows the trend of various parts of the body that require continuous monitoring of the patient's restrained behavior every 2 seconds. Figure 18 is a system block diagram of the non-contact automated noise classification system of the present invention. Implementation

[0031] The following will explain the terminology used in this invention. If a term is not listed in the following description, it will be interpreted in accordance with references recognized by those skilled in the art, such as dictionaries, encyclopedias, literature, or common knowledge.

[0032] The term "Richmond Agitation-Sedation Scale (RASS)" is a scale used to measure the degree of agitation or sedation in patients. RASS scores range from -4 to +4, each representing a different behavioral state.

[0033] The term "MediaPipe Pose technology" is a computer vision framework developed by Google to achieve human pose estimation. It uses deep learning models to identify and track key points of the human body in real time, and can mark the positions of 33 pose nodes of the human body (such as: head, shoulders, elbows, knees, ankles, etc.).

[0034] The term "OpenCV Background Subtractor MOG2 technology" refers to a background modeling and foreground separation algorithm in the OpenCV library. This algorithm primarily learns the background information of each pixel in a video to automatically identify and separate moving objects. It is widely used in video surveillance, behavior analysis, motion detection, and other fields, especially for extracting moving foreground objects from video streams. In computer vision, background subtraction is a commonly used technique for extracting foreground objects from video sequences. The core idea of ​​background subtraction is to model the background and then compare the current frame with the background model to separate the foreground objects. OpenCV (Open Source Computer Vision Library) is a cross-platform computer vision library initiated and co-developed by Intel, available free of charge in commercial and research fields. OpenCV can be used to develop real-time image processing, computer vision, and image recognition programs. For example, OpenCV provides various background subtraction algorithms, with the MOG algorithm and MOG2 algorithm being common. The MOG algorithm is a background subtraction method based on a Gaussian mixture model (MOG). Its core idea is to use multiple Gaussian distributions to model the pixel values ​​in the background. Each pixel value is treated as a random variable, and its distribution consists of multiple Gaussian distributions. In this way, the MOG algorithm can handle complex variations in the background, such as changes in lighting and shadows. Furthermore, the MOG2 algorithm is an improved version of the MOG algorithm. The main difference is that it can automatically select the number of Gaussian distributions and better adapt to background variations. Moreover, the MOG2 algorithm can more accurately model the background by dynamically adjusting the number and parameters of the Gaussian distributions, thereby improving the accuracy of foreground detection.

[0035] Specifically, the non-contact automated noise classification method of the present invention includes the following steps:

[0036] Step A: Collect historical data from multiple patients, including historical video footage;

[0037] Step B: Use these historical images to train a machine learning model to generate a disturbance classification model;

[0038] Step C: Obtain an image of a patient to be analyzed, extract at least one agitation feature from the image, and analyze the agitation feature using the agitation classification model to determine whether the patient to be analyzed is in an agitated state; wherein, the agitation feature is a facial movement feature, a trunk movement feature, a lower limb movement feature, and / or a restraint behavior feature.

[0039] Therefore, this method enables continuous and accurate monitoring of patient agitation, assisting medical staff in adjusting sedation strategies in a timely manner, improving patient safety, and reducing the occurrence of adverse events.

[0040] The following further describes the detailed technical features of the present invention, and to facilitate a better understanding of the invention, descriptions will focus on data collection, data processing, data classification, model evaluation, and real-time disturbance classification. Please also refer to the flowchart in Figure 1.

[0041] Data collection

[0042] First, it must be stated that all data used in this case (such as historical data, historical images, and images to be identified) has been reviewed and approved by the Research Ethics Review Committee of Taichung Veterans General Hospital, and was collected in accordance with the principles of informed consent and anonymity from patients. Accordingly, in 2023, 4K network cameras were used to record 8-10 minutes of monitoring footage of 427 non-communicating critically ill patients in the Intensive Care Unit (ICU) at 1080p / 30fps. This footage served as historical images, and relevant physiological characteristics were recorded simultaneously, particularly the historical RASS scores used to measure patient agitation levels using the RASS sedation assessment scale. In other possible implementation scenarios, the duration of each historical image can be increased or decreased as needed.

[0043] Data processing

[0044] In constructing this agitation classification model, in addition to the method described in step B above, to make the model more accurate, step B further includes a data screening procedure and a preprocessing procedure. The data screening procedure removes historical data that does not meet a screening criterion, and uses the historical images of those that meet the screening criterion as training data. Furthermore, the screening criterion checks whether the historical RASS score included in each patient's historical data falls within a preset range, using this as the judgment standard. If the historical RASS score does not fall within the preset range, the historical data is deleted. The preset range is a set of values ​​including -2, -1, 0, 1, 2, 3, and 4. In one example, using the aforementioned screening method, 61 patients were selected from 150 patients, with an average age of approximately 60 years, and a majority of male patients (i.e., a male-to-female ratio of M / F = 46 / 15). In addition, the screening criteria can also exclude patients under 20 years old, pregnant women, and HIV patients.

[0045] The preprocessing procedure includes an image segmentation procedure, a noise removal procedure, a labeling procedure, a background separation procedure, and a data quantization procedure. The image segmentation procedure divides each historical image into multiple master segments, and then further divides each master segment into multiple sub-segments, with the duration of each sub-segment being shorter than the duration of each master segment. Specifically, each master segment has a duration of 30 seconds for continuous observation and classification. Each sub-segment has a duration of 2 seconds for single-action classification. In other possible implementations, the durations of each master segment and each sub-segment can be arbitrarily increased or decreased according to actual needs.

[0046] This noise removal procedure is used to exclude all parent segments containing noise. Specifically, the procedure determines whether the duration of noise in each parent segment is less than a preset time. If the duration of noise exceeds the preset time, the parent segment is deleted. The preset time is 10 seconds, and the noise refers to camera shake or interference from medical personnel. In one example, using the aforementioned exclusion method, a total of 427 parent segments for continuous observation and 6405 sub-segments for single-action classification were obtained.

[0047] Furthermore, after the noise removal process is completed, in order to avoid insufficient training data, the present invention can also utilize a data augmentation step to replace the deleted mother segment with other adjacent mother segments, so as to make the training data more complete.

[0048] This annotation procedure identifies and annotates at least one agitation target in each parent segment and each child segment. The agitation target can be the head, trunk, lower limbs, or restraint behavior. Specifically, the head refers to the area where the head shakes or lifts, using the shoulder joints as a reference. The trunk includes the upper limbs, and the areas where the upper limbs shake or lift, centered on the shoulder and hip joints, are observed. The lower limbs refer to the areas where the lower limbs shake or lift, using the hip joints as a reference. Restraint behavior refers to using any appliance, material, or device to immobilize the body, restricting an individual's freedom of movement in a specific environment.

[0049] To further improve the accuracy of image recognition, this invention employs MediaPipe Pose technology to mark the patient's limb skeleton and key points of the extremities. Using the shoulder and hip joints as references, image segmentation is performed to ensure consistency in the analysis results of images taken from different angles. Furthermore, this technology can automatically correct the patient's orientation and accurately distinguish the range of motion of the head, trunk, and lower limbs, thereby providing more standardized agitation assessment indicators. For example, this technology determines the patient's supine position (horizontal or vertical) by analyzing the y-coordinate distances between the left and right shoulders, and between the right shoulder and right hip. Furthermore, for supine patients, the head orientation is determined by comparing the x-coordinate between the left shoulder and hip.

[0050] This background separation procedure is used to remove the background from each parent segment and each child segment. Specifically, the background separation procedure uses OpenCV Background Subtractor MOG2 technology to separate the background, convert the video into a black and white image, and perform appropriate weight distribution adjustments to help identify the background. As shown in Figure 2, the left image is an illustration before background separation, and the right image is an illustration before background separation, where the black area represents the background and the white area represents the patient's movement.

[0051] This data quantization procedure employs MediaPipe Pose technology to analyze agitation targets within each parent segment and each child segment. When the agitation target is the head, torso, or lower limbs, the degree of movement of the head, torso, or lower limbs is recorded and quantified. The average movement value of the head, torso, or lower limbs is calculated every 2 seconds, as shown in Figure 3. This avoids certain features with large numerical ranges dominating model training, helps optimize model parameters, accelerates model convergence, and improves model performance and generalization ability, thereby achieving the goal of data normalization. Furthermore, a larger degree of movement indicates a greater risk of agitation. For example, as shown in Figure 4, the degree of head movement is between 0.8 and 5; as shown in Figure 5, the degree of torso movement is between 0.8 and 14; and as shown in Figure 6, the degree of lower limb movement is between 0.8 and 11.

[0052] Accordingly, through the various processing steps in this preprocessing procedure, these historical images are transformed into numerical features suitable for training the machine learning model. At the same time, noise data that may interfere with learning is removed and the data is standardized, thereby improving the accuracy and reliability of the noisy classification model.

[0053] Data Classification

[0054] First, a data classification program is used to analyze the parent segments or sub-segments, and based on the analysis results, each parent segment or sub-segment is labeled as "requires attention" or "does not require attention." In this example, the data classification program includes a threshold classification program, a random forest classification program, and a long short-term memory network classification program, wherein:

[0055] The threshold classification program categorizes the degree of movement into three levels: no movement (1 point), in-bed movement (2 points), and significant out-of-bed movement (3 points). Based on these levels, the degree of movement of the head, trunk, and lower limbs within each sub-segment is scored, and the scores are summed to calculate a total score ranging from 3 to 9 points, as shown in Figure 7. When the total score exceeds a preset threshold, the sub-segment is marked as requiring attention; when the total score does not exceed the preset threshold, the sub-segment is marked as not requiring attention. As shown in Figure 7, the preset threshold is 5. Furthermore, this threshold classification program is used to construct a threshold classification model from the agitation classification model to perform the aforementioned labeling task.

[0056] This random forest classification program obtains the movement levels of the head, torso, and lower limbs from these sub-segments as training samples, constructing an agitation classification model that is a random forest model. This random forest model employs the Random Forest (RF) algorithm and includes multiple decision trees. It learns the correlations between sub-segments labeled as requiring or ignoring attention, thereby identifying movement patterns related to different levels of attention and using this as a basis for judging the patient's agitation state. Accordingly, this random forest classification model can label each sub-segment as requiring or ignoring attention, and even classify the patients corresponding to each sub-segment to determine whether the patient is in an agitated state. For example, based on the judgment result, it determines whether the patient belongs to "patients requiring attention" or "patients not requiring attention." The key parameters of this random forest classification program are set as follows: n_estimators = 100, max_features = auto, criterion = Gini.

[0057] This Long Short-Term Memory (LSTM) network classification program obtains the degree of movement of the head, trunk, and lower limbs from these parent segments as training samples. Based on Long Short-Term Memory (LSTM), the constructed agitation classification model is an LSTM network classification model to assist medical staff in real-time monitoring of patient status and providing appropriate clinical intervention. This LSTM network classification model can label each parent segment as requiring attention or not requiring attention, and even classify the patients corresponding to each parent segment to determine whether the patient is in an agitated state. For example, based on the judgment result, it determines whether the patient belongs to "patient requiring attention" or "patient not requiring attention".

[0058] To ensure the accuracy of classification results, this Long Short-Term Memory (LSTM) network classification model employs a two-layer stacked network structure with 20 hidden units, an input dimension of 4, and a time step of 15 to ensure the model can capture the time-series features of movement data. 80% of the training samples are used for training, 20% for testing, and 10% as a validation set to monitor the model's generalization ability. Furthermore, to improve classification performance, cross-entropy, a loss function, the Adam optimizer, and the Softmax activation function are used, enabling the LSM network classification model to effectively learn the degree or pattern of patient movement. Accordingly, this LSM network classification model automatically analyzes continuous patient movement data, accurately identifies time-series patterns, improves classification accuracy and stability, reduces the need for manual monitoring, and allows medical staff to obtain immediate assessments of patient agitation, enabling rapid clinical intervention and treatment, thus improving the efficiency of medical care.

[0059] Model Evaluation

[0060] As shown in the table below, the model evaluation method uses ROC curve, area under the ROC curve (AUC) value (as shown in Figure 8), accuracy (ACC), precision (P), recall (R), and F1 score as evaluation indicators. All of these models have good performance, especially the Long Short-Term Memory Network classification model.

[0061] Table 1. Model Evaluation Results Threshold classification model Random forest classification model Long Short-Term Memory Network Classification Model Accuracy: 0.80 Accuracy: 0.92 Accuracy: 0.92 Accuracy Recall rate F1 score Accuracy Recall rate F1 score Accuracy Recall rate F1 score Note 1 0.78 0.88 0.95 0.96 0.96 0.96 0.92 0.94 No need to pay attention 0.34 0.98 0.50 0.67 0.60 0.63 0.83 0.92 0.84

[0062] In addition, the present invention randomly selects 50 segments from 427 parent segments for model performance verification, as shown in the confusion matrix in Figures 9 to 11. The value of the preset threshold can also be adjusted according to the results to make the classification results of the model more in line with actual needs.

[0063] Instantaneous Disturbance Classification

[0064] As can be seen from step C, the non-contact automated agitation classification method provided by this invention can achieve real-time monitoring of patient agitation status and determine whether the patient to be analyzed is in an agitated state, thereby assisting medical staff in adjusting sedation strategies in a timely manner, improving patient safety, and reducing the occurrence of adverse events. In addition, the image to be identified can undergo preprocessing procedures before extraction, such as image segmentation, noise removal, annotation, background separation, and data quantization, to improve feature extraction efficiency and classification accuracy. Furthermore, the definitions of facial movement features, trunk movement features, lower limb movement features, and restraint behavior features are the same as those described in the aforementioned annotation procedure, and therefore will not be repeated.

[0065] To this end, the present invention also provides a non-contact automated noise classification system, including a database 10, an image capturing unit 20, a computing module 30 and a display unit 40, as shown in FIG18.

[0066] The database 10 is a storage device pre-configured by a computing device, such as phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), flash memory disk, read-only memory (ROM), random access memory (RAM), used to store historical data, historical images, the noise classification model, the image to be identified, the noise features, the threshold classification model, the random forest classification model, the long short-term memory network classification model, and other related data. The computing device can be, but is not limited to, a personal computer (PC), a laptop computer, a tablet, a smartphone, or other devices with computing or computational capabilities, and the computing device can execute the various programs, steps, and operations contained in the aforementioned method.

[0067] The image capturing unit 20 is a camera, camera, device containing a charge-coupled device (CCD), or device containing a complementary metal-oxide-semiconductor (CMOS) device preset by the computing device, used to record the image of the patient to be analyzed.

[0068] The computing module 30 is a pre-installed device of the computing apparatus, connected to the database 10 and the image acquisition unit 20 respectively. The computing module 30 extracts agitation features from the image to be identified and inputs them into the agitation classification model for analysis, thereby determining whether the patient to be analyzed is in an agitated state. The computing module 30 is a hardware such as a processor that can run programs, software, or train models, or perform operations such as "calculation," "analysis," "judgment," "comparison," "matching," and "operation" on the received data.

[0069] The display unit 40 is a device preset by the computing device or other connected devices, such as a liquid crystal display (LCD), an organic light-emitting diode display (OLED), an electronic whiteboard, a medical instrument panel, or other devices that can be recognized by human senses. The display unit 40 is connected to the image capturing unit 20 and the computing module 30 respectively, and is used to display relevant data or information.

[0070] Furthermore, Figures 12 to 17 show the actual test results of the system, which prove that the system can indeed achieve continuous monitoring, improve patient safety, reduce the occurrence of adverse events, and reduce the burden on clinical manpower, thus having great clinical value.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

[0072] 10: Database 20: Image Capture Unit 30: Computing Module 40: Display Unit

Claims

1. A non-contact automated agitation classification method, comprising the following steps: Step A: collecting historical data of multiple patients, the historical data including historical images; Step B: training a machine learning model using the historical images to generate an agitation classification model; further comprising: a data screening procedure to remove those historical data that do not meet a screening criterion, and using the historical images of the historical data that meet the screening criterion as training data; a preprocessing procedure, including an image segmentation procedure, an annotation procedure, a background separation procedure, and a data quantization procedure, wherein: The image segmentation program segments each historical image into multiple parent segments, and then segments each parent segment into multiple sub-segments, with the duration of each sub-segment being shorter than the duration of each parent segment; the annotation program identifies and annotates at least one agitation marker in each parent segment and each sub-segment, where the agitation marker is the head, torso, lower limbs, or restraint behavior; the background separation program removes the background from each parent segment and each sub-segment; the data quantization program analyzes the agitation markers in each parent segment and each sub-segment, and records and quantifies the degree of movement of the head, torso, or lower limbs when the agitation marker is the head, torso, or lower limbs; a data classification program analyzes the parent segments or sub-segments and, based on the analysis results, labels each parent segment or sub-segment as requiring attention or not requiring attention, wherein the data classification program includes a random forest classification program or a long short-term memory network classification program; When the data classification program includes the random forest classification program, it acquires the movement levels of the head, torso, and lower limbs in the sub-segments as training samples, and uses the Random Forest (RF) algorithm to construct a disturbance classification model that is a random forest classification model capable of making judgments and analyses, labeling each sub-segment as requiring attention or not requiring attention. When the data classification program includes the Long Short-Term Memory (LSTM) network classification program, it acquires the movement levels of the head, torso, and lower limbs in the parent segments as training samples, and uses Long Short-Term Memory (LSTM) to construct a disturbance classification model that is a LSTM network classification model capable of making judgments and analyses, labeling each parent segment as requiring attention or not requiring attention. Step C: An image of a patient to be analyzed is obtained, and at least one agitation feature is extracted from the image. The agitation feature is then analyzed using the agitation classification model to determine whether the patient is in an agitated state. The agitation feature is a facial movement feature, a trunk movement feature, and / or a lower limb movement feature.

2. The non-contact automated noise classification method as described in claim 1, wherein, The screening criterion is to check whether the historical RASS score included in the patient's historical data is within a preset range, and use it as the judgment standard. If the historical RASS score does not fall within the preset range, the historical data is deleted.

3. The non-contact automated noise classification method as described in claim 2, wherein, The preset range is a set of numbers, including -2, -1, 0, 1, 2, 3, 4.

4. The non-contact automated noise classification method as described in claim 1, wherein, Each parent segment has a duration of 30 seconds, and each child segment has a duration of 2 seconds.

5. The non-contact automated noise classification method as described in claim 1, wherein, The preprocessing procedure also includes a noise removal procedure to exclude all parent segments containing noise.

6. The non-contact automated noise classification method as described in claim 5, wherein, The noise removal program determines whether the duration of noise in each parent segment is below a preset time. If the duration of noise exceeds the preset time, the parent segment is deleted.

7. The non-contact automated noise classification method as described in claim 6, wherein, The preset time is 10 seconds, and the noise refers to camera shake or interference from medical staff.

8. The non-contact automated noise classification method as described in claim 1, wherein, The data classification program includes a threshold classification program, which classifies the degree of movement into three levels: no movement (1 point), bed movement (2 points), and significant out-of-bed movement (3 points). The program scores the degree of movement of the head, trunk, and lower limbs in these sub-segments and sums them up to a total score, which ranges from 3 to 9 points. When the total score exceeds a preset threshold, the sub-segment is marked as requiring attention; when the total score does not exceed the preset threshold, the sub-segment is marked as not requiring attention.

9. The non-contact automated noise classification method as described in claim 8, wherein, The preset threshold is 5.

10. A non-contact automated noise classification system, comprising a computing device for performing the non-contact automated noise classification method as described in any one of claims 1 to 9.

Images