Hip joint postoperative delirium patient behavior pattern constraint risk identification method

Abstract

The invention relates to the technical field of medical monitoring, in particular to a hip joint postoperative delirium patient behavior pattern constraint risk identification method, which comprises the following steps: step 1, acquiring multi-parameter data of a patient; 2, multi-parameter data preprocessing: filtering the physiological parameters acquired in the step 1, denoising and segmenting behavior data, and standardizing all data to form a structured data set; step 3, behavior pattern feature extraction; 4, applying a constraint risk identification model, and inputting the key features extracted in the step 3 into a pre-trained risk identification model; and step 5, outputting a risk result and giving an alarm, generating an audio-visual alarm and a risk report according to the constraint risk level in the step 4, and recording the response time of the medical personnel. Through multi-parameter data acquisition and scene specific feature extraction, delirium behaviors and normal postoperative behaviors are effectively distinguished, and false alarms and missing alarms are reduced.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and in particular to a method for identifying behavioral pattern constraints in patients with postoperative delirium after hip surgery. Background Technology

[0002] Hip surgery is a common orthopedic procedure for elderly patients, and postoperative delirium is a frequent complication, manifesting as acute altered consciousness, inattention, cognitive fluctuations, and sensory abnormalities. Delirious patients often exhibit behavioral disturbances, such as agitation, attempts to remove drainage tubes, aimless movement of the affected limb, and even attempts to leave the bed. These behaviors not only increase the risk of falls, wound dehiscence, or infection but may also lead to the dislodgement of medical equipment, seriously threatening patient safety. To prevent such events, physical restraint methods, such as using restraint straps to immobilize the patient's limbs, are often used in clinical practice. However, restraint itself brings a series of negative effects, including pressure sores, muscle damage, psychological trauma, and loss of patient dignity. Furthermore, unnecessary restraint can prolong hospital stays and increase medical costs.

[0003] In current technologies, the identification of behavioral risks in patients with postoperative delirium after hip arthroplasty mainly relies on the subjective observation and experience of healthcare professionals. Nurses conduct regular rounds and assess patient behavior, screening based on clinical scales (such as CAM-ICU). However, this method has significant limitations: First, subjective assessments are easily influenced by personal experience and fatigue, leading to low accuracy. Second, delirium is characterized by its sudden onset and fluctuations, making real-time monitoring impossible during intermittent rounds, resulting in missed risk reports. Third, general behavioral monitoring technologies (such as video surveillance or wearable motion sensors) lack optimization for the specific scenario of postoperative delirium after hip arthroplasty, failing to distinguish between movement caused by normal postoperative pain and delirium-related agitation. For example, a patient adjusting their position due to pain may be confused with unconscious struggling caused by delirium, leading to false alarms.

[0004] Some advanced technologies attempt to identify risks through automated devices, such as using cameras to capture patient movements or wearable devices to monitor physiological parameters. However, these methods often analyze single types of data in isolation, failing to integrate multimodal information and neglecting specific behavioral patterns of patients after hip surgery, such as abnormal flexion, adduction, or attempts to flex the hip. Furthermore, existing technologies lack in-depth analysis of the correlation between behavioral patterns and restraint risk, leading to delayed or inaccurate warnings. Therefore, there is an urgent need in the field for a behavioral pattern restraint risk identification method specifically for patients with post-hip surgery delirium, capable of identifying high-risk behaviors in real time and accurately, thereby optimizing restraint use and improving patient safety. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides a method for identifying behavioral pattern constraints and risks in patients with postoperative delirium after hip surgery, the method comprising the following steps:

[0006] Step 1: Multi-parameter patient data collection. Physiological parameters of patients with post-hip surgery delirium are collected using medical monitoring equipment. These physiological parameters include heart rate data, blood pressure data, and blood oxygen saturation data. At the same time, behavioral data of patients are collected using behavioral monitoring equipment. These behavioral data include body movement data, gesture data, and facial expression data. All data are synchronized via timestamps.

[0007] Step 2: Multi-parameter data preprocessing. The physiological parameters collected in Step 1 are filtered, the behavioral data are denoised and segmented, and all data are standardized to form a structured dataset.

[0008] Step 3: Behavioral pattern feature extraction. Key features are extracted from the preprocessed data. These key features include heart rate variability index, blood pressure fluctuation coefficient, frequency of decrease in blood oxygen saturation, body movement amplitude, gesture frequency, facial expression change rate, degree of abnormal movement of affected limb, and frequency of attempted hip flexion.

[0009] Step 4: Application of the constraint risk identification model. The key features extracted in Step 3 are input into the pre-trained risk identification model. The risk identification model is based on a machine learning algorithm and outputs the constraint risk level, which includes low risk, medium risk and high risk.

[0010] Step 5: Risk Result Output and Alarm. Based on the risk level constraints in Step 4, generate audiovisual alarms and risk reports, and record the response time of medical staff.

[0011] Preferably, the specific process of acquiring multi-parameter patient data in step 1 includes:

[0012] Heart rate data is continuously monitored using an electrocardiogram monitor, with the sampling frequency set according to clinical standards to ensure real-time data transmission.

[0013] Blood pressure data are measured using a non-invasive blood pressure monitor at preset intervals, which are adjusted based on the patient's postoperative condition.

[0014] The pulse oximeter continuously acquires blood oxygen saturation data, which is output as a digital signal.

[0015] Body motion data is captured using depth cameras installed in the bedside area. The depth cameras record the movement trajectory of the patient's torso and limbs in the form of three-dimensional point clouds.

[0016] Gesture data is acquired using an inertial measurement unit (IMU) sensor worn on the patient's wrist. The IMU sensor includes an accelerometer and a gyroscope to measure hand acceleration and angular velocity.

[0017] High-definition cameras are used to capture facial expression data. The high-definition cameras focus on the patient's facial area and record expression changes in the form of video streams.

[0018] The process of synchronizing timestamps across all devices is achieved through a central clock module, ensuring that physiological parameters and behavioral data are aligned in time, and the collected data is temporarily stored in a local buffer.

[0019] Preferably, the specific process of multi-parameter data preprocessing in step 2 includes:

[0020] A low-pass filter is used to remove high-frequency noise from the heart rate data. The cutoff frequency of the low-pass filter is adaptively adjusted based on the clinical fluctuation range of the heart rate data.

[0021] The blood pressure data is smoothed using a moving average method, and the size of the moving window is dynamically set according to the coefficient of variation of the blood pressure data.

[0022] Median filtering is used to eliminate transient outliers in blood oxygen saturation data. The window length of the median filter is determined based on the sampling rate of the blood oxygen saturation data.

[0023] Environmental interference was reduced by using a point cloud denoising algorithm on body motion data. The point cloud denoising algorithm was implemented based on neighborhood statistical analysis, and a skeletal key point extraction algorithm was used to segment the motion sequences of the patient's trunk and limbs.

[0024] The zero-point drift of the inertial measurement unit sensor is calibrated for the gesture data. The calibration process uses static reference measurement and a threshold method is used to identify valid gesture actions. The threshold is set based on the historical distribution of acceleration and angular velocity of the gesture data.

[0025] The facial expression data is cropped using a face detection algorithm based on the Haar feature classifier, and the lighting conditions are normalized. The normalization process uses a histogram equalization method.

[0026] Data standardization employs the min-max normalization method to convert physiological parameters and behavioral data into a uniform numerical range. The standardized data is then organized according to time series to form a structured dataset.

[0027] Preferably, the specific process of extracting behavioral pattern features in step 3 includes:

[0028] Heart rate variability is obtained by analyzing the time-domain standard deviation of heart rate data. The length of the time window for calculating the time-domain standard deviation is set based on the sampling frequency of the heart rate data.

[0029] The blood pressure variability coefficient is calculated using the coefficient of variation of blood pressure data, which is the ratio of the standard deviation to the mean of the blood pressure data.

[0030] The frequency of decreased blood oxygen saturation is obtained by counting the number of times blood oxygen saturation falls below a preset threshold. The preset threshold is dynamically adjusted based on the patient's baseline blood oxygen saturation value.

[0031] The range of body movement is obtained by the displacement vector magnitude of key points on the torso in the depth camera data. The displacement vector magnitude is calculated based on the three-dimensional coordinate changes.

[0032] The gesture frequency is obtained by counting the peak acceleration data from the inertial measurement unit sensor data, and the peak acceleration count is obtained by the sliding window detection method.

[0033] The rate of change of facial expression was obtained by the average motion optical flow of facial feature points in high-definition camera data. The motion optical flow was calculated based on the Lucas-Kanade algorithm.

[0034] The degree of abnormal movement of the affected limb is obtained by comparing the ratio of the movement amplitude of the affected limb to that of the healthy limb. The ratio of movement amplitude is calculated based on the displacement of key points of the limb in the depth camera data.

[0035] The frequency of attempted hip flexion movements was identified by the hip joint angle change sequence in depth camera data, and the angle difference was calculated based on the skeletal key points.

[0036] Preferably, the specific process of applying the constraint risk identification model in step 4 includes:

[0037] The risk identification model is a three-class classification model based on the support vector machine algorithm. The kernel function type of the support vector machine algorithm is selected based on the feature distribution, and the kernel function type includes linear kernel or radial basis function kernel.

[0038] The model was trained using historical data, which included multi-parameter features of patients with post-hip surgery delirium and corresponding constraint event annotations. The constraint event annotations were manually added by medical staff based on clinical records.

[0039] The model input consists of the key features extracted in step 3, including heart rate variability index, blood pressure fluctuation coefficient, frequency of decrease in blood oxygen saturation, body movement amplitude, gesture frequency, facial expression change rate, degree of abnormal movement of affected limb, and frequency of attempted hip flexion.

[0040] The process of model outputting risk levels includes calculating classification scores, which are obtained through support vector machine decision functions, and classifying risk levels based on score thresholds, which are determined through cross-validation during model training.

[0041] The model is updated in real time, with the update frequency set based on the sampling rate of the data collection, to ensure the timeliness of risk identification.

[0042] Preferably, the specific process of risk result output and alarm in step 5 includes:

[0043] An audiovisual alarm is generated. When the constraint risk level is medium or high, the audiovisual alarm is automatically triggered. The audiovisual alarm signal is sent to the nurse station monitoring terminal. The alarm information includes the patient's identifier, risk level and timestamp.

[0044] Risk reports are generated periodically and display feature changes and risk history in the form of visual charts, including line charts and bar charts. The reports are transmitted to the electronic medical record database through the hospital information system.

[0045] An output feedback mechanism records the response time of medical staff. The response time is calculated from the time interval between the alarm being triggered and the medical staff's confirmation. The recorded data is used for model optimization and data retraining.

[0046] Alarm priority settings dynamically adjust alarm priorities based on the constraint risk level, with high-risk levels corresponding to high-priority alarms and medium-risk levels corresponding to medium-priority alarms.

[0047] Preferably, the extraction process of the degree of abnormal movement of the affected limb in step 3 further includes:

[0048] The degree of abnormal movement of the affected limb is calculated by the ratio of the amplitude of movement of the affected limb to the amplitude of movement of the healthy limb. The amplitude of movement is obtained based on the magnitude of the displacement vector of the key point of the limb in the depth camera data.

[0049] The ratio calculation uses a sliding time window, the length of which is adaptively adjusted based on historical data of patient behavior patterns.

[0050] The threshold for abnormal movement of the affected limb is set based on clinical data statistics. The threshold is determined by the distinguishing point between delirious behavior and normal behavior in historical data.

[0051] The process of extracting the frequency of attempted hip flexion movements further includes: identifying attempted hip flexion movements through changes in hip joint angles in depth camera data; calculating hip joint angles based on skeletal key points; detecting angle change sequences using a differential method; and frequency statistics based on the number of movements per unit time.

[0052] Preferably, the training process of the risk identification model in step 4 includes:

[0053] Historical data collection included multi-parameter data from multiple patients with postoperative delirium after hip arthroplasty. The multi-parameter data included physiological parameters and behavioral data, and corresponding constraint events were labeled. The constraint event labels were independently verified by multiple medical staff.

[0054] Feature engineering extracts key features from historical data, including heart rate variability indicators, blood pressure fluctuation coefficient, frequency of decrease in blood oxygen saturation, body movement amplitude, gesture frequency, facial expression change rate, degree of abnormal movement of affected limb, and frequency of attempted hip flexion.

[0055] The model training uses the support vector machine algorithm. The training process includes data partitioning, parameter tuning, and validation. Data partitioning adopts random stratified sampling, parameter tuning uses the grid search method, and validation uses k-fold cross-validation.

[0056] Model evaluation is based on accuracy, recall, and F1 score, and the evaluation results are used for model selection and improvement.

[0057] The model update mechanism retrains the model periodically with new data, and the retraining frequency is set based on clinical needs.

[0058] Preferably, the data standardization process in step 2 further includes:

[0059] The specific application of the min-max normalization method is to scale each feature value to the range of zero to one, with the scaling formula based on the minimum and maximum values ​​of the feature;

[0060] The determination of the minimum and maximum values ​​of the features is based on historical data statistics, which cover a variety of clinical scenarios;

[0061] The standardized data is stored in a structured dataset, which is indexed by time series to facilitate subsequent feature extraction and model application.

[0062] Data quality checks involve verifying data integrity before standardization. Verification methods include missing value detection and outlier removal. Missing values ​​are handled using interpolation methods, while outliers are removed using statistical methods.

[0063] Preferably, the method is integrated into a hospital monitoring system, and specific implementation methods include:

[0064] The data acquisition equipment is integrated with the hospital’s existing monitoring system, and the medical monitoring equipment and behavioral monitoring equipment are connected to the central processor through a standard interface;

[0065] The preprocessing and feature extraction algorithms run on an embedded processor configured to process the data stream in real time.

[0066] The risk identification model is deployed on the server side, and the server side communicates with the embedded processor via the network to achieve cloud analysis;

[0067] The entire process is implemented in software, with seamless data flow between steps. Step 1 outputs raw multi-parameter data to Step 2, Step 2 outputs preprocessed data to Step 3, Step 3 outputs feature vectors to Step 4, and Step 4 outputs risk levels to Step 5.

[0068] The implementation follows medical data privacy guidelines, and all data is anonymized using an identifier substitution method.

[0069] The beneficial effects of this invention are:

[0070] 1. By collecting multi-parameter data and extracting scene-specific features, delirium behavior can be effectively distinguished from normal postoperative behavior, reducing false alarms and missed alarms.

[0071] 2. Based on continuous data collection and model application, high-risk behaviors can be identified in a timely manner, allowing for early intervention.

[0072] 3. By automating risk identification, the burden on medical staff can be reduced, unnecessary physical constraints can be minimized, and the patient experience can be improved. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0074] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0075] Figure 2 This is a flowchart of the steps for collecting multi-parameter patient data in step 1 of the method of the present invention. Detailed Implementation

[0076] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0077] Please see Figures 1-2This invention provides a method for identifying behavioral pattern constraints and risks in patients with postoperative delirium after hip surgery. In step 1, the patient's physiological parameters, including heart rate, blood pressure, and blood oxygen saturation, are collected in real time using medical monitoring equipment. These physiological parameters reflect the patient's physiological state and can help monitor postoperative recovery progress. Simultaneously, behavioral monitoring equipment monitors the patient's behavior, collecting data on body movements, gestures, and facial expressions. This data provides crucial evidence for identifying delirium-related behaviors. All data is synchronized using timestamps to ensure accuracy and real-time performance.

[0078] In step 2, the collected physiological parameters are filtered to remove noise and ensure data clarity. Behavioral data undergoes denoising algorithms and segmentation to more accurately reflect the patient's behavioral state through changes in movement trajectories, gestures, and facial expressions. Simultaneously, all data is standardized to unify its dimensions, making subsequent data analysis more efficient and accurate.

[0079] In step 3, after data preprocessing, key features related to constraint risk are extracted. These features include changes in physiological parameters such as heart rate variability, blood pressure fluctuations, and the frequency of decreases in blood oxygen saturation, reflecting the patient's physiological fluctuations. Simultaneously, behavioral characteristics such as body movement amplitude, gesture frequency, facial expression changes, as well as abnormal limb movements and hip flexion frequency, can reflect whether the patient exhibits delirium-related agitation, struggling, or other behaviors. These key features provide necessary evidence for subsequent risk identification.

[0080] In step 4, the extracted key features are input into a pre-trained machine learning model for analysis. Based on extensive historical data and clinical experience, this model can determine a patient's constraint risk level according to feature combinations. The model outputs risk levels categorized as low, medium, and high risk, helping clinicians assess a patient's current safety risk level.

[0081] In step 5, based on the model's output, the system will automatically generate audiovisual alerts and risk reports to remind medical staff to pay attention to abnormal patient behavior and avoid overuse of constraints. Simultaneously, the system will record the response time of medical staff, providing a basis for subsequent work efficiency analysis and improvement.

[0082] By combining real-time monitoring of various physiological and behavioral data, precise identification of patient behavior patterns and constraint risks is achieved. This not only avoids the biases caused by subjective assessments in traditional methods but also enables the detection of high-risk behaviors in a very short time, issuing timely alerts and ensuring patient safety. Furthermore, it avoids unnecessary physical constraints, reducing negative impacts such as pressure sores and muscle injuries, thus improving patient comfort and recovery efficiency.

[0083] In one possible implementation, the patient multi-parameter data acquisition process described in step 1 achieves comprehensive monitoring of patients with postoperative delirium after hip arthroplasty through the collaborative work of different devices. The specific process includes:

[0084] The patient's heart rate is continuously monitored using an electrocardiogram (ECG) monitor, with the sampling frequency set according to clinical standards. This ensures the real-time nature of the heart rate data, reflecting the patient's cardiac status and physiological changes, especially cardiovascular responses related to postoperative recovery.

[0085] Blood pressure data is measured at preset intervals using a non-invasive blood pressure monitor, with the intervals dynamically adjusted based on the patient's postoperative condition. For example, shorter measurement intervals may be needed in the early postoperative period, while these intervals can be appropriately extended once the patient's condition stabilizes. This flexibility ensures that blood pressure monitoring is highly consistent with the patient's actual condition.

[0086] A pulse oximeter continuously monitors a patient's blood oxygen saturation and outputs the data in real time as a digital signal. Blood oxygen saturation is an important indicator for assessing a patient's respiratory status and can promptly detect hypoxemia or other respiratory abnormalities.

[0087] Depth cameras installed in the bedside area record the patient's trunk and limb movement trajectories. By capturing three-dimensional point cloud data, the patient's movement status can be accurately tracked, especially abnormal movements and changes in posture, which helps to identify whether there is excessive agitation or abnormal limb activity caused by delirium.

[0088] The inertial measurement unit (IMU) sensor worn on the patient's wrist includes an accelerometer and a gyroscope, capable of measuring the acceleration and angular velocity of the hand. This device is used to capture changes in the patient's hand gestures and can promptly identify behaviors triggered by delirium, such as agitation and struggling.

[0089] A high-definition camera focuses on the patient's facial area, recording changes in facial expressions via video stream. Facial expressions are important signals reflecting a patient's emotions and psychological state, especially for recognizing emotional fluctuations and changes in consciousness caused by delirium.

[0090] Data collected from all devices is timestamped and synchronized via a central clock module, ensuring time alignment across different data types. This guarantees accurate matching of physiological parameters and behavioral data, providing high-quality input data for subsequent risk assessment.

[0091] Through the synchronized collaboration of the aforementioned devices, the patient's physiological and behavioral states can be comprehensively captured, data transmitted in real time, and analyzed. This data not only provides an accurate foundation for subsequent behavioral pattern recognition and risk assessment but also allows for real-time responses to changes in the patient, effectively preventing and reducing the risk of delirium, thereby improving patient safety and comfort. This invention improves the comprehensiveness, accuracy, and real-time nature of monitoring, ensuring maximum safety during the patient's recovery process.

[0092] In one possible implementation, the multi-parameter data preprocessing process described in step 2 aims to remove noise, smooth data fluctuations, and extract useful information, thereby improving the accuracy of subsequent data analysis. The specific process is as follows:

[0093] A low-pass filter is used to remove noise from heart rate data, especially high-frequency noise. The cutoff frequency of the low-pass filter is adaptively adjusted based on the clinical fluctuation range of the heart rate data. This method can effectively remove high-frequency noise caused by equipment errors or external interference, thereby ensuring the accuracy of heart rate data, especially in dynamic monitoring.

[0094] The moving average method smooths out fluctuations in blood pressure data, with the sliding window size dynamically set based on the coefficient of variation of the blood pressure data. This method helps reduce short-term fluctuations in blood pressure measurements, making the data more stable and facilitating accurate assessment of patients' blood pressure trends, especially in the postoperative period.

[0095] Median filtering was used to eliminate transient outliers in blood oxygen saturation data. The window length of the median filter was dynamically set based on the sampling rate of the blood oxygen saturation data, which effectively removed abnormal peaks and ensured data smoothness. Especially when the sampling frequency was high, it could effectively filter out irrelevant data fluctuations.

[0096] Point cloud denoising algorithms are used to reduce environmental interference. Based on neighborhood statistical analysis, this algorithm can identify and remove the impact of environmental noise on motion data. Simultaneously, a skeletal keypoint extraction algorithm is employed to segment the patient's trunk and limb movements, thereby accurately identifying the patient's movement trajectory and avoiding interference from environmental factors on the accuracy of motion data.

[0097] In gesture data processing, the accuracy of measurements is ensured by calibrating the zero-point drift of the inertial measurement unit (IMU) sensor. The calibration process uses static reference measurements, and a threshold method is used to identify valid gestures. The threshold value is set based on the historical distribution of acceleration and angular velocity in the gesture data, which can accurately capture the patient's gesture behavior and avoid misjudgment.

[0098] A face detection algorithm based on a Haar feature classifier was used to crop the facial region, and the lighting conditions were normalized. Histogram equalization was then employed to enhance image contrast. This process effectively eliminates the impact of lighting variations on facial expression data, improving the accuracy of expression recognition, especially in low-light environments.

[0099] All physiological parameters and behavioral data were standardized using the min-max normalization method, transforming different types of parameters to a uniform numerical range. The standardized data was then organized by time series to form a structured dataset, facilitating subsequent analysis and processing. Standardization reduces differences in data units, enhancing the stability and accuracy of the model.

[0100] This preprocessing procedure effectively improves data quality by reducing noise, smoothing data, and performing precise signal extraction. The optimized data enables more accurate subsequent behavioral pattern recognition, thereby effectively reducing the risk of postoperative delirium and improving patient care and safety.

[0101] In one possible implementation, the behavioral pattern feature extraction process in step 3 accurately captures postoperative behavioral changes by calculating multiple physiological and behavioral indicators, providing effective data support for risk identification. Specific implementation details are as follows:

[0102] The time-domain standard deviation of heart rate data is used to calculate heart rate variability, which reflects the degree of change in heart rhythm. The length of the time window is dynamically set according to the sampling frequency of the heart rate data, thereby ensuring the accuracy of data processing. This method can be used to determine the variability of cardiac health, especially during the postoperative recovery process of patients, where heart rate variability is an important indicator for assessing autonomic nervous system function.

[0103] The coefficient of variation (COP) of blood pressure data is calculated to further assess the degree of blood pressure fluctuation. The COP, which is the ratio of the standard deviation to the mean of blood pressure data, reflects the stability of blood pressure fluctuations. Excessive blood pressure fluctuations increase the risk of postoperative complications in patients; therefore, monitoring its volatility helps in timely adjustments to care plans.

[0104] This method assesses changes in a patient's blood oxygen level by counting the number of times their blood oxygen saturation falls below a preset threshold. The preset threshold is dynamically adjusted based on the patient's baseline blood oxygen saturation value, thus better adapting to individual patient differences. This method allows for real-time monitoring of abnormal blood oxygen levels and timely detection of potential respiratory problems.

[0105] The amplitude of body movement is assessed by using the displacement vector magnitude of key trunk points acquired by a depth camera. The displacement vector is calculated based on changes in three-dimensional coordinates. This method helps track the recovery of patients' postoperative motor abilities, especially changes in trunk movement, and can identify patients with limited mobility.

[0106] By using data from inertial measurement unit sensors, the number of acceleration peaks is calculated, and a sliding window detection method is used to statistically analyze frequent hand gestures. This method effectively captures the frequency of patient activity, especially during the postoperative recovery period, where frequent hand gestures may be an important indicator of a patient's emotions or behavior during recovery.

[0107] The process uses a high-definition camera to capture the motion optical flow of facial feature points and calculates their average value to assess changes in facial expressions. This process is based on the Lucas-Kanade optical flow algorithm, which can stably identify subtle changes in facial expressions under different lighting conditions, thereby reflecting changes in the patient's emotions. In particular, in patients with delirium, changes in facial expressions may be related to their psychological state.

[0108] Postoperative motor function of the affected limb is assessed by comparing the range of motion of the affected limb with that of the healthy limb. This process relies on displacement data of key points in the limb acquired by a depth camera. By comparing the differences in movement between the affected and healthy limbs, it helps determine whether the patient has experienced motor abnormalities, especially during the postoperative recovery period, when the recovery of the affected limb is a key indicator.

[0109] By capturing a sequence of hip joint angle changes using a depth camera, the system can identify whether a patient is making frequent hip flexion movements. This sequence of angle changes is calculated based on the angle differences at key skeletal points, enabling precise monitoring of whether the patient intends to flex their hip. Frequent hip flexion movements may be a response to pain, limited mobility, or other psychological factors.

[0110] This invention not only efficiently monitors the physiological and behavioral status of postoperative patients but also provides data support for timely intervention and care, helping to reduce the occurrence of postoperative delirium and improve the quality of postoperative recovery. Through multi-dimensional monitoring data, it enables a comprehensive assessment of the patient's physical condition and psychological response, greatly improving the accuracy and timeliness of risk identification.

[0111] In one possible implementation, a three-class classification model based on the Support Vector Machine (SVM) algorithm is used to identify behavioral pattern constraint risks in patients with post-hip surgery delirium. The specific implementation steps are as follows:

[0112] Support Vector Machine (SVM) is a classification model that maximizes the inter-class margin. In this application, the kernel function type of the SVM algorithm is selected based on the distribution of the feature data. Common kernel function types include linear kernels and radial basis function kernels. Linear kernels are suitable for data with linear relationships between features, while radial basis function kernels are better suited for handling non-linear feature relationships, ensuring that the model can effectively classify data with different distributions.

[0113] The training process utilized historical data, including multi-parameter feature data and corresponding constraint event annotations from patients with post-hip surgery delirium. The constraint event annotations were manually created by experienced healthcare professionals based on clinical records to ensure accuracy. This dataset was used to train the model, enabling it to learn the relationships between different features and constraint events, thus facilitating effective risk identification in the face of real-time data input.

[0114] The model's input data comes from key features extracted in step 3, including heart rate variability, blood pressure fluctuation coefficient, frequency of oxygen saturation decline, body movement amplitude, gesture frequency, facial expression change rate, degree of abnormal movement of the affected limb, and frequency of attempted hip flexion. These features comprehensively reflect the patient's physiological and behavioral state and form the basis for risk identification.

[0115] During model runtime, the input feature data is processed by the decision function of a support vector machine to calculate a classification score. By setting a score threshold, risk can be divided into multiple levels. The determination of the score threshold depends on the cross-validation process during model training to ensure that the classification of different risk levels has good discriminative power and accuracy.

[0116] During model application, real-time data is continuously input into the system for analysis to ensure the timeliness of risk identification. The update frequency is adjusted according to the data collection sampling rate to adapt to the update speed of different data sources, ensuring that the model can reflect the patient's latest status in real time.

[0117] Through the above steps, the designed risk identification model can assess the patient's constraint risk in real time based on their physiological and behavioral characteristics and issue timely alerts. This method can effectively improve the accuracy and response speed of risk identification for patients with postoperative delirium, thereby providing medical staff with more precise decision support, helping to reduce the occurrence of postoperative complications, improve the quality of care, and accelerate patient recovery.

[0118] In one possible implementation, when the model assesses a constraint risk level as medium or high risk, the system will automatically trigger an audiovisual alarm. This alarm signal is transmitted via network to the nurse station monitoring terminal, and the alarm information includes the patient's unique identifier, the current risk level, and a timestamp. This measure ensures that healthcare staff receive real-time alerts immediately, allowing for timely intervention and reducing the potential risks caused by delirium.

[0119] The system regularly generates risk reports based on patient risk assessment results. These reports use visual charts, such as line graphs and bar charts, to display changes in characteristics and risk history. These charts allow healthcare professionals to intuitively understand patient risk trends, enabling more effective monitoring of changes in the patient's condition. The reports are seamlessly connected to the hospital information system and electronic medical record database, ensuring that patient risk data is archived long-term and readily accessible.

[0120] In addition, the system records the response time of medical staff to alarms and calculates the time interval from alarm triggering to medical staff confirmation. This data provides feedback for model optimization and retraining, ensuring more accurate subsequent risk identification and improving the system's intelligence level.

[0121] Finally, the system dynamically adjusts the priority of alerts based on the different risk levels. High-risk levels correspond to high-priority alerts, ensuring that healthcare workers are alerted first when the risk is severe. Medium-risk levels correspond to lower-priority alerts to avoid overreacting by healthcare workers. This setting helps to allocate resources rationally and ensures that the handling of each risk level is targeted.

[0122] The embodiments of the present invention can effectively improve the response efficiency and accuracy of the risk identification system, reduce medical accidents caused by risk neglect or delayed response, and improve the treatment effect and safety of patients.

[0123] In one possible implementation, the extraction process of abnormal limb movement and the extraction process of attempted hip flexion frequency employ advanced depth camera data analysis technology, designed to accurately capture and assess the behavioral patterns of patients with postoperative delirium of the hip joint in order to accurately identify their constraint risks.

[0124] The extraction of limb movement abnormality is assessed by calculating the ratio of the amplitude of movement of the affected limb to that of the healthy limb. The amplitude of movement is calculated using the displacement vector magnitude of key points on the limb acquired by a depth camera. Specifically, the positional change of each key point captured by the camera is used to quantify the amplitude of limb movement. To ensure the accuracy of the analysis, a sliding time window is used for the ratio calculation. The window length is adaptively adjusted based on the patient's historical behavioral pattern data, dynamically responding to different patient behavioral characteristics. The threshold for limb movement abnormality is set based on clinical data statistics, specifically determined by the distinguishing points between delirious behavior and normal behavior in historical data, ensuring that this threshold can effectively differentiate between abnormal behavior and normal activity.

[0125] The extraction of hip flexion frequency is achieved by identifying changes in hip joint angles in depth camera data to monitor whether the patient is engaging in hip flexion. Hip joint angles are calculated from skeletal key points, reflecting actual changes in hip movement. Angle change sequences are detected using a differential method to ensure that every change in movement is captured. Subsequently, frequency statistics are calculated based on the number of movements per unit time to determine the frequency of the patient's attempted hip flexion.

[0126] This invention provides precise behavioral monitoring and risk assessment for patients with postoperative delirium. Detailed analysis of affected limb movement and hip flexion can promptly detect abnormal behaviors, preventing further complications caused by limited mobility or loss of behavioral control. Simultaneously, this technology provides clinicians with more scientific data support, thereby improving treatment accuracy and patient recovery efficiency.

[0127] In one possible implementation, the training process of the risk identification model includes multiple steps, from data collection to model update mechanisms, to ensure accurate identification and adaptability to clinical needs.

[0128] Specifically, firstly, multi-parameter data were collected from multiple patients with post-hip surgery delirium. This data included physiological parameters (such as heart rate, blood pressure, and blood oxygen saturation) and behavioral data (such as range of motion, gesture frequency, and facial expression changes). Each dataset was labeled with relevant constraint events, and these labels were independently verified by multiple medical professionals to ensure accuracy and consistency. This approach ensured data reliability and provided a solid foundation for subsequent risk identification model training.

[0129] After data collection, key features need to be extracted from historical data. These features include heart rate variability, blood pressure fluctuation coefficient, frequency of decrease in blood oxygen saturation, body movement amplitude, gesture frequency, facial expression change rate, degree of abnormal movement of the affected limb, and frequency of attempted hip flexion. These features reflect changes in the patient's physiological and behavioral state and are important basis for predicting delirium behavior and restraint risk.

[0130] The training process uses the Support Vector Machine (SVM) algorithm. First, the data is divided into training and test sets using stratified random sampling to ensure data balance across different classes. To optimize model performance, parameter tuning employs a grid search method to find the most suitable parameter combination. During validation, k-fold cross-validation is used to further verify the model's generalization ability on different datasets and avoid overfitting.

[0131] When evaluating the model, accuracy, recall, and F1 score are used as evaluation metrics. These metrics comprehensively assess the model's classification performance, including its balance and accuracy across different classes. The evaluation results allow for model selection and improvement, ensuring maximum effectiveness in real-world applications.

[0132] As clinical data accumulates, the model needs to be updated regularly. New data will be used to retrain the model, and the frequency of retraining will be set according to clinical needs. This mechanism allows the model to continuously optimize as patient data changes, improving the accuracy and adaptability of risk identification.

[0133] The embodiments of this invention ensure that the risk identification model possesses high accuracy and real-time performance, effectively identifying behavioral patterns and constraint risks in patients with postoperative delirium after hip arthroplasty, and providing a scientific basis for clinical decision-making. This method not only improves the quality of patient care but also provides important support for the intelligent management of medical systems.

[0134] In one possible implementation, firstly, multi-parameter data from multiple patients with post-hip surgery delirium are collected, including physiological data (such as heart rate, blood pressure, and blood oxygen saturation) and behavioral data (such as range of motion, gesture frequency, and facial expression changes). Each data sample is independently validated and labeled with constraint events by multiple medical staff to ensure the accuracy and consistency of the labeling. This validation mechanism improves the reliability of the data and provides high-quality basic data for subsequent model training.

[0135] Key features were extracted from the collected historical data, including heart rate variability, blood pressure fluctuation coefficient, frequency of decrease in blood oxygen saturation, body movement amplitude, gesture frequency, facial expression change rate, degree of abnormal movement of the affected limb, and frequency of attempted hip flexion. These features reflect the patient's physiological and behavioral changes from multiple dimensions and are important evidence for predicting delirious behavior and restraint risk.

[0136] The Support Vector Machine (SVM) algorithm was used for model training. Data was first partitioned using a stratified random sampling method to ensure class balance between the training and test sets. The model parameters were then tuned using a grid search method to find the optimal parameter combination. To avoid overfitting, k-fold cross-validation was used to verify the model's generalization ability, ensuring high accuracy even on unseen data.

[0137] Model evaluation uses a comprehensive set of metrics, including accuracy, recall, and F1 score. These metrics comprehensively measure the model's performance, particularly its accuracy in classification and its balance across different classes. This evaluation helps determine whether the model can effectively distinguish between delirious behavior and normal behavior, and provides a basis for further model improvement.

[0138] As new clinical data is continuously collected, the model is periodically retrained using this new data to maintain its timeliness and adaptability. The frequency of retraining can be adjusted according to clinical needs to ensure that the model can promptly reflect changes in patient status and the impact of new data, further improving the accuracy of risk identification.

[0139] This invention ensures that the risk identification model can accurately capture the behavioral characteristics of patients with postoperative delirium after hip arthroplasty, improves the ability to predict constraint risks, provides a scientific basis for clinicians, effectively improves the quality of patient care, and promotes the development of intelligent medical management.

[0140] In one possible implementation, embodiments of the present invention achieve real-time monitoring and intelligent risk assessment by integrating with existing hospital monitoring systems. Specific implementation details are as follows:

[0141] Medical monitoring devices (such as electrocardiographs, blood pressure monitors, and pulse oximeters) and behavioral monitoring devices (such as cameras and sensors) connect to a central processing unit via standard interfaces to collect patients' physiological and behavioral data in real time. The completeness and accuracy of the collected data provide the foundation for subsequent processing and analysis.

[0142] After data acquisition, the embedded processor first performs real-time preprocessing to remove noise and standardize various data points. Next, the embedded processor extracts features from the data, identifying key physiological and behavioral characteristics from multiple dimensions, such as heart rate variability, facial expression changes, and gesture frequency. These features are crucial factors in determining whether a patient is in a state of delirium.

[0143] The risk identification model is integrated with the hospital's monitoring system and deployed on the server side, ensuring that the computation and analysis process is not limited by device performance. The server communicates with the embedded processor via a network to achieve cloud-based analysis. This deployment method can handle large amounts of data and perform deep learning training and inference, thereby improving the model's identification accuracy and adaptability.

[0144] The entire identification process is implemented in software, with data flowing seamlessly between each step. The first step is the acquisition and transmission of raw multi-parameter data; the second step is data preprocessing; the third step is feature vector extraction; the fourth step is to output the risk level through the risk identification model; and finally, the output risk level provides decision support for doctors.

[0145] All patient data strictly adheres to medical data privacy regulations. Data is anonymized during collection and transmission using identifier replacement to ensure the security and confidentiality of patient identity information.

[0146] The embodiments of this invention can significantly improve the accuracy of delirium risk identification, reduce risks in patient care, and ensure patient privacy protection, aligning with the development trend of modern medical informatization and intelligentization. Through seamless integration with existing monitoring systems, it can monitor patient status in real time and provide scientific evidence to assist clinicians in making timely and effective judgments.

[0147] Example

[0148] The scenario is an orthopedic ward in a general hospital, where a 72-year-old patient (24 hours post-surgery) is being monitored after undergoing hip replacement surgery. The patient exhibits mild delirium symptoms, characterized by intermittent agitation and confusion, requiring real-time identification of behavioral patterns to assess restraint risks.

[0149] System integration and environment setup:

[0150] This method integrates with the hospital's existing monitoring system, including medical monitoring equipment, behavioral monitoring equipment, a central processing unit (CPU), and a server. The medical monitoring equipment includes electrocardiogram (ECG) monitors, non-invasive blood pressure monitors, and pulse oximeters, which connect to the CPU via standard interfaces. The behavioral monitoring equipment includes a depth camera (mounted 1.5 meters directly above the patient's bed), an inertial measurement unit (IMU) sensor (worn on the patient's wrist), and a high-definition camera (focusing on the patient's face). The CPU runs preprocessing and feature extraction algorithms, while the server deploys a risk identification model and communicates with the nurses' station monitoring terminal via the hospital network. All data acquisition and processing adhere to medical data privacy guidelines; patient identifiers are immediately replaced with anonymous codes after acquisition, and the anonymization process uses a hash algorithm (SHA-256) to generate unique identifiers.

[0151] Detailed explanation of the steps in the implementation example:

[0152] Step 1: Patient Multi-Parameter Data Acquisition This step involves collecting the patient's physiological parameters and behavioral data. The specific process is as follows:

[0153] Physiological parameter collection:

[0154] Heart rate data: Continuous monitoring was performed using an electrocardiogram (ECG) monitor at a sampling frequency of 250 Hz (based on clinical standards to ensure capture of heart rate variability details). The ECG monitor was connected to the patient's chest via electrode patches and outputs a digital heart rate sequence.

[0155] Blood pressure data: Measured using a non-invasive blood pressure monitor at preset intervals. The preset intervals are dynamically adjusted based on the patient's postoperative condition (in this case, the initial interval is 15 minutes, which is automatically shortened to 5 minutes when an abnormal heart rate is detected). Blood pressure data is output in millimeters of mercury (mmHg).

[0156] Blood oxygen saturation data: continuously acquired using a pulse oximeter at a sampling frequency of 100 Hz. Blood oxygen saturation data is output as a percentage and transmitted to the central processing unit via Bluetooth.

[0157] Behavioral data collection:

[0158] Body motion data: Captured using a depth camera, which generates 3D point cloud data at a rate of 30 frames per second, recording the movement trajectories of the patient's torso and limbs. The point cloud data includes key skeletal points (such as the 3D coordinates of joints like the shoulder, hip, and knee).

[0159] Gesture data: Acquired using an inertial measurement unit (IMU) sensor worn on the patient's wrist, measuring triaxial acceleration (range ±16g) and triaxial angular velocity (range ±2000 degrees / second) at a sampling frequency of 50 Hz. Data is transmitted wirelessly via a protocol such as Wi-Fi.

[0160] Facial expression data: Captured using a high-definition camera, recorded as a 1080p resolution video stream at 30 frames per second, focusing on the patient's facial area to ensure clear imaging under ward lighting conditions (illuminance 200-500 lux).

[0161] Data synchronization: All devices synchronize timestamps via a central clock module (based on the NTP protocol) to ensure that physiological parameters and behavioral data are aligned in time. The collected data is temporarily stored in a local buffer (1GB capacity), and the buffer data is transferred to the central processor in batches every 5 minutes.

[0162] Step 2: Multi-parameter data preprocessing; This step involves cleaning and standardizing the collected data. The specific process is as follows:

[0163] Physiological parameter filtering:

[0164] Heart rate data: High-frequency noise was removed using a low-pass filter, with a cutoff frequency set to 0.5 Hz (based on clinical heart rate variability analysis standards, the cutoff frequency was adaptively adjusted using the power spectral density of the heart rate data to ensure the retention of effective low-frequency components). After filtering, the heart rate data was converted into a heart rate per minute (BPM) sequence.

[0165] Blood pressure data: The moving average method is used to smooth fluctuations, and the sliding window size is set to 5 data points (dynamically adjusted based on the coefficient of variation of the blood pressure data. The coefficient of variation is calculated as the ratio of the standard deviation to the mean. When the coefficient of variation exceeds 10%, the window size is increased to 7 points).

[0166] Blood oxygen saturation data: Median filtering was used to eliminate transient outliers, and the window length was set to 10 samples (based on a sampling rate of 100 Hz, the window length corresponds to 0.1 seconds to ensure the filtering out of transient interference).

[0167] Denoising and segmentation of behavioral data:

[0168] Body motion data: Environmental interference was reduced using a point cloud denoising algorithm, which was based on neighborhood statistical analysis (neighborhood radius set to 0.1 meters) to remove outliers. Then, a skeletal keypoint extraction algorithm (based on Azure Kinect SDK) was used to segment the patient's torso and limb motion sequences, with keypoints including 15 points such as the hip and knee joints.

[0169] Gesture data: The zero-point drift of the inertial measurement unit sensor was calibrated. The calibration process was performed while the patient was stationary (lasting 10 seconds), and a baseline value was measured. Then, a threshold method was used to identify valid gestures. The thresholds were set based on historical data distribution (acceleration threshold was set to 0.5g, angular velocity threshold was set to 50 degrees / second, and gestures exceeding the thresholds were considered valid).

[0170] Facial expression data: Facial regions were cropped using a face detection algorithm (based on the OpenCV Haar feature classifier), which was trained on the FER2013 dataset. Then, lighting conditions were normalized, and histogram equalization was used to enhance contrast.

[0171] Data standardization: A min-max normalization method is used to scale each feature value to the range [0,1]. For example, the minimum and maximum values ​​of heart rate data are based on historical patient data statistics (in this example, the minimum is set to 50 BPM and the maximum to 120 BPM). The standardized data is organized by time series to form a structured dataset (stored in CSV format, with each row containing a timestamp and all parameter values).

[0172] Step 3: Behavioral Pattern Feature Extraction; This step extracts key features from the preprocessed data. The specific process is as follows:

[0173] Physiological parameter feature extraction:

[0174] Heart rate variability index: obtained by analyzing the time-domain standard deviation of heart rate data. The time window length for calculating the time-domain standard deviation is set to 5 minutes (based on a sampling frequency of 250 Hz, the window contains 75,000 samples), and the standard deviation value reflects the degree of heart rate fluctuation.

[0175] Blood pressure variability coefficient: Calculated using the coefficient of variation of blood pressure data, which is the ratio of the standard deviation to the mean of the blood pressure data. The calculation is based on a sliding window (window size of 10 measurement points), and a ratio greater than 0.15 is considered abnormal fluctuation.

[0176] Frequency of oxygen saturation decline: This is obtained by counting the number of times oxygen saturation falls below a preset threshold in the blood oxygen saturation data. The preset threshold is dynamically adjusted based on the patient's baseline oxygen saturation value (in this example, the baseline value is 95%, and the threshold is set to 90%), and the statistical window length is 1 minute.

[0177] Behavioral data feature extraction:

[0178] Body motion amplitude: obtained from the displacement vector magnitude of key points on the torso in depth camera data. The displacement vector magnitude is calculated based on changes in three-dimensional coordinates (e.g., the Euclidean distance of hip joint key points across consecutive frames), and the average value is taken as the amplitude value.

[0179] Gesture frequency: obtained by counting peak acceleration values ​​in the inertial measurement unit sensor data. The peak acceleration count uses a sliding window detection method (window length 1 second), and the peak value is defined as the acceleration value exceeding twice the standard deviation of the historical average.

[0180] Facial expression change rate: obtained by averaging the motion optical flow of facial feature points in high-definition camera data. The motion optical flow calculation is based on the Lucas-Kanade algorithm (implemented using OpenCV, window size 15x15 pixels), and the change rate is the average optical flow amplitude per unit time.

[0181] Extraction of postoperative specific behavioral features of hip joint surgery:

[0182] Abnormality of movement in the affected limb: This is obtained by comparing the range of motion of the affected limb with that of the healthy limb. The range of motion is based on the displacement of key points of the limb in depth camera data (e.g., the magnitude of the displacement vector of the knee joint key point), and the ratio is calculated using a sliding time window (window length 30 seconds). A ratio greater than 1.5 is considered abnormal (based on clinical data statistics, the ratio is usually higher than 1.5 in patients with delirium).

[0183] Hip flexion movement frequency: This is obtained by identifying hip joint angle change sequences from depth camera data. Hip joint angles are calculated based on skeletal key points (vector dot product formula), and angle change sequences are detected using a difference method (changes exceeding 10 degrees are considered hip flexion movements). Frequency statistics are based on the number of movements per unit time (1 minute). All features are aggregated by time window (5 minutes) to form a feature vector (8-dimensional vector), which is stored in the feature database.

[0184] Step 4: Application of the constraint risk identification model;

[0185] This step inputs the feature vector into the risk identification model and outputs a constrained risk level. The specific process is as follows:

[0186] Risk identification model: A three-class classification model based on the support vector machine algorithm, with radial basis function kernels (based on the nonlinear characteristics of feature distribution) selected as the kernel type. The model was trained using historical data, which included multi-parameter features and corresponding constraint event annotations of 200 patients with postoperative delirium of hip arthroplasty (the annotations were independently verified by 3 medical staff, with a consistency of over 90%).

[0187] Model training process:

[0188] Data partitioning: Using stratified random sampling, historical data were divided into training, validation and test sets in a 70:15:15 ratio.

[0189] Parameter tuning: Optimize the support vector machine parameters (such as the penalty parameter C and the kernel function parameter γ) using a grid search method, with the search range C being [0.1, 10] and γ being [0.01, 1].

[0190] Validation: k-fold cross-validation (k=5) was used, and the evaluation metrics included accuracy, recall and F1 score (in this example, the model achieved an accuracy of 92% and a recall of 89% on the test set).

[0191] Model Application: Real-time feature vectors are input into the model, and the support vector machine decision function calculates the classification score. The score threshold is determined through cross-validation during training (low risk: score < 0.3, medium risk: 0.3 ≤ score < 0.7, high risk: score ≥ 0.7). The model application is updated in real time, with the update frequency based on the data collection sampling rate (once every 5 minutes).

[0192] In this embodiment, after the patient's real-time feature vector is input into the model, the score is 0.75, triggering a high-risk level.

[0193] Step 5: Risk Result Output and Alerts; This step generates output information based on the risk level. The specific process is as follows:

[0194] Audiovisual alarm generation: When the constraint risk level is medium or high, an audiovisual alarm is automatically triggered. The alarm signal is sent to the nurse station monitoring terminal via the hospital network (the terminal display interface includes the patient's bed number, risk level, and timestamp). The audiovisual alarm includes an audible prompt (frequency 1000 Hz, lasting 3 seconds) and a flashing visual signal (red for high risk, yellow for medium risk).

[0195] Risk report generation: Regularly (every 30 minutes), behavioral pattern trend reports are generated, displaying characteristic changes and risk history in the form of visual charts. The visual charts include line graphs (showing trends in heart rate variability and abnormality of affected limb movement) and bar charts (showing gesture frequency and frequency of attempted hip flexion movements). The reports are transmitted to the electronic medical record database through the hospital information system (such as the EPIC system).

[0196] Output feedback mechanism: Records the response time of medical staff, calculating the time interval from alarm triggering to confirmation by medical staff on the terminal. The recorded data is used for model optimization and data retraining (e.g., a response time exceeding 5 minutes is considered a delay, triggering model recalibration).

[0197] In this embodiment, after a high-risk alarm is triggered, the nurse arrives at the patient's bedside within 2 minutes to implement interventions (such as comforting the patient), and the response time is recorded and used for subsequent analysis.

[0198] It is understandable that:

[0199] Heart rate variability index: This index reflects the function of the autonomic nervous system and is calculated based on the time-domain standard deviation of heart rate data. The larger the standard deviation, the more drastic the heart rate fluctuations, which may be related to the stress response caused by delirium.

[0200] Affected limb movement abnormality: This feature is specifically designed for patients after hip surgery. By comparing the range of motion of the affected limb and the unaffected limb, it identifies asymmetric behaviors caused by delirium. The ratio calculation is based on 3D point cloud data to ensure accuracy.

[0201] Support Vector Machine Model: Radial basis function kernel is chosen because the feature data may have non-linear relationships. The kernel parameters are optimized through grid search to ensure the model's generalization ability.

[0202] Data standardization: The min-max normalization method ensures that features of different dimensions (such as heart rate and exercise amplitude) are within the same range, avoiding model bias.

[0203] Threshold settings: All thresholds (such as blood oxygen saturation threshold and score threshold) are dynamically adjusted based on historical data statistics and clinical experience to adapt to individual differences.

[0204] This embodiment demonstrates the application of the method of the present invention in a real-world scenario. Through multi-parameter data acquisition, detailed preprocessing, specific feature extraction, and machine learning models, it achieves accurate and real-time identification of behavioral pattern constraint risks in patients with postoperative delirium of the hip joint.

[0205] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0206] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method of identifying a risk of a behavior pattern restraint in a patient with postoperative delirium after hip arthroplasty, characterized by, The method comprises the following steps: Step 1: patient multi-parameter data acquisition, physiological parameters of a patient with postoperative delirium of hip arthroplasty are collected using a medical monitoring device, the physiological parameters include heart rate data, blood pressure data and blood oxygen saturation data, at the same time, behavior data of the patient is collected using a behavior monitoring device, the behavior data includes body movement data, gesture data and facial expression data, all data are synchronized through time stamp; Step 2: multi-parameter data preprocessing, the physiological parameters collected in step 1 are filtered, the behavior data is denoised and segmented, and all data are standardized to form a structured data set; Step 3: behavior pattern feature extraction, key features are extracted from the preprocessed data, the key features include heart rate variability index, blood pressure fluctuation coefficient, blood oxygen saturation drop frequency, body movement amplitude, gesture frequency, facial expression change rate, abnormality degree of affected limb movement and frequency of trying to flexion action; Step 4: application of constraint risk identification model, the key features extracted in step 3 are input into a pre-trained risk identification model, the risk identification model is based on a machine learning algorithm, and a constraint risk level is output, the constraint risk level includes low risk, medium risk and high risk; Step 5: risk result output and alarm, an audio-visual alarm and a risk report are generated according to the constraint risk level of step 4, and a response time of medical staff is recorded.

2. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip arthroplasty as claimed in claim 1, characterized in that, The specific process of the patient multi-parameter data acquisition in step 1 comprises: The heart rate data is continuously monitored using an electrocardiogram monitor, the sampling frequency is set according to clinical standards to ensure real-time data; The blood pressure data is measured at a preset interval using a non-invasive blood pressure meter, the preset interval is adjusted based on the postoperative state of the patient; The blood oxygen saturation data is continuously obtained using a pulse oximeter, and the blood oxygen saturation data is output in the form of a digital signal; The body movement data is captured using a depth camera installed in the bed area, and the depth camera records the movement trajectory of the patient's trunk and limbs in the form of a three-dimensional point cloud; The gesture data is obtained using an inertial measurement unit sensor worn on the patient's wrist, the inertial measurement unit sensor contains an accelerometer and a gyroscope to measure hand acceleration and angular velocity; The facial expression data is captured using a high-definition camera, the high-definition camera focuses on the patient's face area, and records the expression changes in the form of a video stream; The process of synchronizing the time stamps of all devices is realized through a central clock module, ensuring that the physiological parameters and behavior data are aligned in time, and the collected data is temporarily stored in a local buffer.

3. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip arthroplasty as claimed in claim 1, characterized in that, The specific process of the multi-parameter data preprocessing in step 2 comprises: A low-pass filter is used to remove high-frequency noise from the heart rate data, and the cutoff frequency of the low-pass filter is adaptively adjusted based on the clinical fluctuation range of the heart rate data; A moving average method is used to smooth the fluctuations of the blood pressure data, and the size of the sliding window is dynamically set according to the coefficient of variation of the blood pressure data; Median filtering is used to eliminate transient outliers from the blood oxygen saturation data, and the window length of the median filtering is determined based on the sampling rate of the blood oxygen saturation data; A point cloud denoising algorithm based on neighborhood statistical analysis is used to reduce environmental interference, and a skeleton key point extraction algorithm is used to segment the movement sequence of the patient's trunk and limbs; The gesture data is calibrated by the zero-point drift of the inertial measurement unit sensor, the calibration process uses static reference measurement, and the threshold method is used to identify the effective gesture action, and the threshold is set based on the acceleration and angular velocity history distribution of the gesture data; The facial expression data is cropped by the face detection algorithm, the face detection algorithm is based on Haar feature classifier, and the illumination condition is normalized, and the normalization process uses histogram equalization method; The data standardization adopts the minimum-maximum normalization method to convert the physiological parameters and behavior data into a unified numerical range, and the standardized data is organized in time sequence to form a structured data set.

4. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip arthroplasty as claimed in claim 1, characterized in that, The specific process of behavior pattern feature extraction in step 3 includes: The heart rate variability index is obtained by analyzing the time domain standard deviation of heart rate data, and the time window length for calculating the time domain standard deviation is set based on the sampling frequency of heart rate data; The blood pressure fluctuation coefficient is calculated by the coefficient of variation of blood pressure data, and the coefficient of variation is the ratio of the standard deviation to the average value of blood pressure data; The blood oxygen saturation drop frequency is obtained by counting the number of times that the blood oxygen saturation data is below the preset threshold, and the preset threshold is dynamically adjusted based on the patient's baseline blood oxygen saturation value; The body movement amplitude is obtained by the displacement vector module length of the trunk key point in the depth camera data, and the displacement vector module length calculation is based on three-dimensional coordinate change; The gesture frequency is obtained by counting the acceleration peak value in the inertial measurement unit sensor data, and the acceleration peak value count uses a sliding window detection method; The facial expression change rate is obtained by the average value of the motion optical flow of the facial feature points in the high-definition camera data, and the motion optical flow calculation is based on Lucas-Kanade algorithm; The abnormality degree of affected limb movement is obtained by comparing the movement amplitude ratio of the affected limb to the healthy limb, and the movement amplitude ratio calculation is based on the key point displacement of the limb in the depth camera data; The frequency of trying to flex the hip action is obtained by recognizing the hip joint angle change sequence in the depth camera data, and the hip joint angle change sequence is based on the calculation of angle difference based on skeletal key points.

5. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip arthroplasty as claimed in claim 1, characterized in that, The specific process of applying the constraint risk identification model in step 4 includes: The risk identification model is a three-classification model based on support vector machine algorithm, and the kernel function type of support vector machine algorithm is selected based on feature distribution, including linear kernel or radial basis function kernel; The model training uses historical data, which includes multi-parameter features of hip arthroplasty postoperative delirium patients and corresponding constraint event labels, and the constraint event labels are manually labeled by medical staff according to clinical records; The model input is the key features extracted in step 3, including heart rate variability index, blood pressure fluctuation coefficient, blood oxygen saturation drop frequency, body movement amplitude, gesture frequency, facial expression change rate, abnormality degree of affected limb movement and frequency of trying to flex the hip action; The process of outputting the constraint risk level of the model includes calculating the classification score, which is obtained by support vector machine decision function, and dividing the risk level according to the score threshold, and the score threshold is determined by cross-validation during model training; The model application is updated in real time, and the update frequency is set based on the sampling rate of data acquisition to ensure the timeliness of risk identification.

6. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip replacement surgery as claimed in claim 1, wherein, The specific process of risk result output and alarm in step 5 includes: An audio-visual alarm is generated, which is automatically triggered when the constraint risk level is medium risk or high risk, and the audio-visual alarm signal is sent to the nurse station monitoring terminal, and the alarm information includes patient identification, risk level and timestamp; A risk report is generated, which is generated regularly, and the report displays feature changes and risk history in the form of visual charts, including line charts and bar charts, and the report is transmitted to the electronic medical record database through the hospital information system; An output feedback mechanism is provided to record the response time of medical staff, which is calculated from the time interval from alarm triggering to confirmation by medical staff, and the recorded data is used for model optimization and data retraining; Alarm priority is set based on the constraint risk level, and high risk level corresponds to high priority alarm and medium risk level corresponds to medium priority alarm.

7. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip replacement surgery as claimed in claim 1, wherein, The extraction process of the abnormality degree of the affected limb movement in step 3 further includes: The abnormality degree of the affected limb movement is calculated by the ratio of the movement amplitude of the affected limb to the movement amplitude of the healthy limb, and the movement amplitude is obtained based on the key point displacement vector module length of the limb in the depth camera data; The ratio calculation uses a sliding time window, and the length of the time window is adaptively adjusted based on the historical data of the patient behavior pattern; The threshold setting of the abnormality degree of the affected limb movement is based on clinical data statistics, and the threshold is determined by the dividing point between delirium behavior and normal behavior in the historical data; The extraction process of the frequency of attempted hip flexion action further includes: the attempted hip flexion action is identified by the change of the hip joint angle in the depth camera data, the hip joint angle is calculated based on the key points of the skeleton, the angle change sequence is detected by the difference method, and the frequency is counted based on the number of actions in a unit of time.

8. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip replacement surgery as claimed in claim 1, wherein, The training process of the risk identification model in step 4 includes: Historical data collection, collecting multi-parameter data of multiple delirium patients after hip joint surgery, including physiological parameters and behavior data, and corresponding constraint events are labeled, and the constraint event labels are verified by multiple medical staff independently; Feature engineering, extracting key features from historical data, including heart rate variability index, blood pressure fluctuation coefficient, blood oxygen saturation drop frequency, body movement amplitude, gesture frequency, facial expression change rate, abnormality degree of affected limb movement and frequency of attempted hip flexion action; Model training uses support vector machine algorithm, and the training process includes data division, parameter optimization and verification, data division uses random stratified sampling, parameter optimization uses grid search method, and verification uses k-fold cross-validation; Model evaluation is based on accuracy, recall and F1 score, and the evaluation results are used for model selection and improvement; Model updating mechanism, periodically retrain the model using new data, and the retraining frequency is set based on clinical needs.

9. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip replacement surgery as claimed in claim 1, wherein, The data standardization process in step 2 further includes: Specific application of the min-max normalization method, each feature value is scaled to the range of zero to one, and the scaling formula is based on the minimum and maximum values of the feature; The determination of the minimum and maximum values of the feature is based on the statistical analysis of the historical data, which covers multiple clinical scenarios; The standardized data is stored in a structured data set, which is indexed in time sequence, facilitating subsequent feature extraction and model application; Data quality check, data integrity verification before standardization, verification method includes missing value detection and outlier elimination, missing value is treated by interpolation method, and abnormal value is eliminated by statistical method.

10. A method of identifying a risk of a postoperative delirium patient behavior pattern restraint in a hip replacement surgery as claimed in claim 1, wherein, The method is integrated in a hospital monitoring system, and specifically comprises: The data acquisition device is integrated with the existing monitoring system of the hospital, and the medical monitoring device and the behavior monitoring device are connected to the central processor through a standard interface; The preprocessing and feature extraction algorithm runs on an embedded processor configured to process data streams in real time; The risk identification model is deployed on a server side, and the server side communicates with the embedded processor through a network to realize cloud analysis; The whole process is realized in the form of software, and the data flow between the steps is coherent, step 1 outputs the original multi-parameter data to step 2, step 2 outputs the preprocessed data to step 3, step 3 outputs the feature vector to step 4, and step 4 outputs the risk level to step 5.

Images