Trauma emergency treatment patient posture monitoring method and system

Abstract

The invention provides a trauma emergency treatment patient posture monitoring method and system, and relates to the technical field of posture monitoring and artificial intelligence, and the method comprises the steps: collecting the first posture information and environment context information of a patient according to a first target when the trauma emergency treatment patient is received, inputting a first AI model which is responsible for reverse deduction of the damage mechanism; wherein the first target does not interfere with medical treatment; when the first AI model outputs a candidate injury mechanism to be cured, collecting second posture information of the patient according to a second target, and inputting the second posture information into the first AI model; wherein the second target is to enable the second attitude information to be used for the first AI model to carry out calibration or exclusion on the candidate lesion mechanism, and to allow moderate interference to medical treatment. According to the method, the posture information of the patient is collected in stages, the environment context is combined, reverse deduction and verification of the injury mechanism are conducted through the first AI model, and the accuracy and efficiency of judgment of the injury mechanism in the emergency treatment scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of posture monitoring and artificial intelligence technology, and in particular to a method and system for posture monitoring of trauma emergency patients. Background Technology

[0002] When trauma patients receive emergency medical treatment, accurate assessment of the injury mechanism is crucial for developing an effective treatment plan.

[0003] However, in traditional emergency care, medical staff mainly rely on the patient's complaints, physical examinations, and imaging studies to infer the mechanism of injury. This approach often faces limitations when time is tight or the patient cannot express themselves accurately.

[0004] Currently, the application of artificial intelligence (AI) technology in the medical field is becoming increasingly profound. In particular, the ability of AI models to process complex data and perform reverse engineering provides new possibilities for the precise analysis of trauma mechanisms. However, existing technologies still have shortcomings in the dynamic posture information acquisition and analysis of trauma emergency patients, especially in the difficulty of acquiring and integrating patient posture information and environmental context information in real time to assist in diagnosis without interfering with medical treatment. Furthermore, existing methods often lack the ability to dynamically adjust acquisition strategies during the verification phase of injury mechanism deduction, thus limiting the accuracy of the deduction results.

[0005] Therefore, there is an urgent need for a method that can collect patient posture information in real time in emergency scenarios, perform inverse deduction of injury mechanisms in combination with environmental context, and dynamically adjust the collection targets based on the deduction results. This method needs to maximize the accuracy of injury mechanism deduction without interfering with or only moderately interfering with medical treatment, providing medical personnel with reliable auxiliary diagnostic information, thereby optimizing the emergency treatment process and improving treatment outcomes. Summary of the Invention

[0006] One of the objectives of this invention is to provide a method and system for monitoring the posture of trauma patients in emergency care, in order to solve the problems mentioned in the background art.

[0007] In a first aspect, an embodiment of the present invention provides a method for monitoring the posture of traumatic emergency patients, comprising: When treating patients with trauma emergencies, the primary objective is to collect the patient's initial posture information and environmental context information, and input them together into the primary AI model responsible for inverse deduction of the injury mechanism; the primary objective is to avoid interfering with medical treatment. When the first AI model outputs a candidate injury mechanism to be confirmed, according to the second objective, the patient's second posture information is collected and input into the first AI model; wherein, the second objective is to enable the second posture information to allow the first AI model to confirm or exclude the candidate injury mechanism, and to allow for moderate interference with medical treatment.

[0008] Optionally, the first posture information includes a sequence of depth images of the patient collected by the posture monitoring module according to the first target, time-series data of three-dimensional coordinates of skeletal joints calculated based on the depth image sequence, and body pressure distribution data.

[0009] Optionally, the environmental context information includes voice commands and keyword fragments of dialogue from medical personnel performing medical procedures, collected by audio acquisition devices, and status signals of medical equipment collected by environmental sensors.

[0010] Optionally, the non-interference with medical treatment includes: The potential interference sensors in the attitude monitoring module are controlled to operate in an intermittent acquisition mode based on environmental perception; specifically, by analyzing real-time environmental data to identify the intervals in medical treatment operations, the potential interference sensors are triggered to perform rapid data acquisition only within the time window of the interval.

[0011] Optionally, enabling the second pose information to allow the first AI model to confirm or exclude candidate damage mechanisms includes: This ensures that the second attitude information meets the first specific data acquisition requirements set to confirm or exclude the candidate damage mechanism; The first specific data acquisition requirement is dynamically generated by the first AI model based on the expected biomechanical characteristics corresponding to the candidate damage mechanism.

[0012] Optionally, the permission to moderately interfere with medical procedures includes: The attitude monitoring module is allowed to provide avoidance prompts to medical personnel who obstruct the monitoring path when it is working along the first necessary monitoring path; The first necessary monitoring path is the working path planned in real time by the attitude monitoring module according to the first specific data acquisition requirements.

[0013] Optionally, the attitude monitoring module includes at least one depth vision sensor mounted on an electronically controlled gimbal, which is fixedly installed on the ceiling above the patient's bed. The electronically controlled gimbal is configured to change the pitch angle, horizontal orientation, and optical focal length of the depth vision sensor according to control commands, so as to achieve coverage of the first necessary monitoring path.

[0014] Optional methods for monitoring posture in trauma emergency patients also include: When the first AI model is performing a sit-down analysis on the candidate injury mechanism, it collects the patient's third posture information according to the third objective; the third objective is to enable the third posture information to provide the second AI model responsible for secondary injury warning to the patient for secondary injury warning. Specifically, collecting the patient's third posture information according to the third objective includes: The candidate injury mechanisms identified by the first AI model are input into the second AI model. The second AI model then determines the set of undesirable postures that may cause secondary injury to the patient based on the knowledge base of the mapping between injury mechanisms and undesirable postures. Based on the set of undesirable postures, a second specific data acquisition requirement is determined to capture the key biomechanical features corresponding to each posture in the set of undesirable postures. The operation performed by the attitude monitoring module includes: planning a second necessary monitoring path in real time according to the second specific data acquisition requirements; wherein, the second necessary monitoring path refers to the optimal sensor working trajectory planned by the attitude monitoring module to fully cover the second specific data acquisition requirements; working along the second necessary monitoring path to collect the patient's third posture information, and simultaneously collecting one or more influencing factors that may hinder the second necessary monitoring path; influencing factors include medical staff, mobile medical devices, or caregivers; Collect identity information and dynamic behavioral data of influencing factors; Based on the collected identity information and dynamic behavior data of the influencing factors, the time window of obstruction of the second necessary monitoring path by the influencing factors and the corresponding reminder response time are predicted; the reminder response time refers to the expected time from the issuance of the avoidance prompt to the influencing factor taking the avoidance action; The second necessary monitoring path is continuously and dynamically corrected to generate a corrected second necessary monitoring path. The attitude monitoring module is then controlled to operate along the corrected path, ensuring that the following conditions are met when the attitude monitoring module operates along the corrected second necessary monitoring path: Condition 1: Within the time window of obstruction, the collection of the patient's third posture information must not be affected if the second specific data collection requirement is met. Condition 2: If Condition 1 cannot be met, if an influencing factor causes obstruction at any time within the obstruction time window, there is sufficient capability to continue predicting and supplementing the patient's third posture information within the reminder response time after the obstruction occurs.

[0015] Optionally, the step of predicting the time window of obstruction of the second necessary monitoring path by the influencing factors and the corresponding reminder response time based on the collected identity information and dynamic behavior data of the influencing factors includes: Based on identity information, the corresponding inherent behavioral attributes are matched from a pre-established object type-behavioral attribute database. The inherent behavioral attributes include typical movement speed, path preference, and typical dwell time. Based on dynamic behavioral data and matched path preferences, the future movement trajectory of influencing factors is predicted by a Kalman filter algorithm with behavioral habit constraints; wherein, the behavioral habit constraints use path preferences as the prior probability distribution for trajectory prediction. The predicted future trajectory is spatiotemporally overlaid with the second necessary monitoring path to identify the intersection point of the future paths; Based on the current speed of the influencing factors, the typical speed of movement in the inherent behavioral attributes, the spatial location of the path intersection point, and the correction coefficient of movement speed for the path preference, the estimated arrival time of the influencing factors to the path intersection point is calculated. Based on the typical dwell time in the inherent behavioral attributes, and combined with the adjustment weight of dwell time by path preference and the environmental noise collected in real time by environmental sensors, the dwell time is jointly corrected to determine the expected dwell time of influencing factors at the path intersection point. Starting from the estimated arrival time, and using the jointly corrected estimated stay duration as the time window length, a barrier time window is generated; Based on identity information, the corresponding baseline reminder response time is obtained from a pre-established object type-response time database; Obtain current environmental status parameters, including environmental noise level, ambient light intensity, and real-time attention concentration index of the influencing factor; By using a pre-trained attention-reaction time mapping model, the baseline reminder response time is adjusted by multiple factors based on the current environmental state parameters to generate the final dynamic reminder response time.

[0016] Secondly, an embodiment of the present invention provides a posture monitoring system for trauma emergency patients, comprising: The first posture monitoring module is used to collect the patient's first posture information and environmental context information when receiving patients with trauma emergencies, according to the first objective, and input them together into the first AI model responsible for inverse deduction of the injury mechanism; wherein, the first objective is not to interfere with medical treatment. The second posture monitoring module is used to collect the patient's second posture information and input it into the first AI model according to the second objective when the first AI model outputs a candidate injury mechanism to be confirmed. The second objective is to enable the second posture information to be used by the first AI model to confirm or exclude the candidate injury mechanism, and to allow for moderate interference with medical treatment.

[0017] The present invention has achieved the following beneficial effects: This invention significantly improves the accuracy and efficiency of injury mechanism identification in emergency scenarios by collecting patient posture information in stages and combining it with environmental context, and then using a first AI model to reverse-engineer and verify the injury mechanism. In the first stage, the method aims to minimize interference with medical procedures, ensuring minimal impact on the smoothness of the treatment process. In the second stage, moderate interference is used to obtain more accurate second posture information, further verifying or eliminating candidate injury mechanisms, thereby improving diagnostic reliability. This multi-target, staged acquisition strategy not only provides medical personnel with scientific auxiliary decision-making support but also optimizes the emergency treatment process, shortens diagnostic time, reduces the risk of misdiagnosis, and ultimately improves patient prognosis.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for monitoring the posture of traumatic emergency patients according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a trauma emergency patient posture monitoring system according to an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] The research and development approach of this application first focuses on how to achieve non-invasive posture information acquisition in emergency scenarios. A first-stage acquisition strategy, with the primary goal of not interfering with medical procedures, is designed. This strategy acquires the patient's initial posture information and environmental context through sensors or imaging technology and inputs it into a first AI model for preliminary damage mechanism deduction. Secondly, to address the accuracy issue in candidate mechanism verification, a second-stage acquisition strategy is designed. This strategy allows for moderate interference with medical procedures to obtain more detailed second posture information, thereby confirming or ruling out candidate mechanisms. The entire research and development approach revolves around the combination of the AI ​​model's reverse engineering capabilities and a dynamic acquisition strategy. It aims to balance the need for smooth medical procedures with the need for diagnostic accuracy through phased and goal-oriented information acquisition, providing efficient and reliable technical support for emergency medicine.

[0023] Figure 1 A flowchart of a method for monitoring the posture of trauma patients in emergency departments is provided in this application embodiment, as follows: Figure 1 As shown, the method includes: Step A: When receiving patients with traumatic injuries, according to the first objective, the patient's first posture information and environmental context information are collected and input together into the first AI model responsible for inverse inference of the injury mechanism; wherein, the first objective is not to interfere with medical treatment. The first posture information consists of the patient's depth image sequence collected by the posture monitoring module according to the first objective, the time-series data of the three-dimensional coordinates of the skeletal joints calculated based on the depth image sequence, and the body pressure distribution data.

[0024] Using a posture monitoring module and environmental sensors, the system collects the patient's initial posture information and environmental context information according to the primary objective (i.e., without interfering with medical procedures). This data is then input into a first AI model for inverse deduction of the injury mechanism. The initial posture information includes depth image sequences, temporal data of 3D coordinates of skeletal joints, and body pressure distribution data. These data reflect the patient's static forced posture characteristics (such as fixed posture due to fractures) and dynamic pain avoidance characteristics (such as posture adjustments due to pain). The depth image sequences are 3D point cloud data of the patient's body surface captured by a depth vision sensor at a frequency of 30 frames per second, specifically generated using Time-of-Flight (ToF) or structured light technology. The point cloud data resolution is no less than 640×480, and the depth accuracy is controlled within ±5 mm. The 3D coordinate temporal data of skeletal joints was obtained by processing depth image sequences with a skeletal tracking algorithm to extract the 3D coordinates (x, y, z) of at least 20 key joints (such as the shoulder, elbow, hip, and knee). The temporal resolution was 30 times per second, and the coordinate accuracy was ±10 mm. A deep learning-based skeletal tracking algorithm (such as a 3D extension of OpenPose or MediaPipe) was used, which extracts image features and maps them to 3D space through a convolutional neural network (CNN). Body pressure distribution data was collected by a pressure sensor array installed on the hospital bed. The array resolution was no less than 16×16, the pressure range was 0-200 kPa, and the accuracy was ±1 kPa, reflecting the contact pressure distribution between the patient's body and the bed surface. Environmental context information included voice commands from medical staff, keyword fragments of dialogue, and status signals from medical equipment. Voice commands and keyword fragments are extracted by audio acquisition devices (such as high-sensitivity microphones, sampling rate 44.1kHz, signal-to-noise ratio ≥60dB) using speech recognition algorithms (such as Transformer-based speech-to-text models). Keywords include medical terms (such as "fracture fixation" and "analgesics") and action commands (such as "turn over" and "elevate"). Medical device status signals are collected by environmental sensors (such as infrared sensors or Bluetooth modules), including device on / off status and operating parameters (such as infusion pump flow rate, in mL / h), with a sampling frequency of once per second. Non-interference with medical procedures is achieved through an intermittent acquisition mode. Specifically, potential interference sensors in the posture monitoring module (such as high-frequency infrared light from depth vision sensors) analyze real-time environmental data (including the location of medical staff and medical device operation signals) to identify the intervals between medical procedures (such as a 5-10 second idle window after a medical staff member completes an operation), triggering data acquisition only during the intervals, with the acquisition duration controlled within 1-2 seconds.The first AI model is a deep learning-based inverse reasoning model that employs a hybrid architecture combining a convolutional neural network (CNN) and a long short-term memory network (LSTM). Its inputs include first-position information (depth images of approximately 640×480×30 frames, a 20×3×30 skeletal coordinate sequence, and a 16×16 pressure distribution matrix) and environmental context information (keyword vectors and device status sequences). The outputs are candidate injury mechanisms (e.g., "femoral fracture," "rib fracture") and their probability distributions. The model learns the mapping relationship between biomechanical features and injury mechanisms using a pre-training dataset (containing the poses and injury labels of at least 1000 trauma patients). Training uses the Adam optimizer with a learning rate of 0.001 and a cross-entropy loss function. Environmental data is acquired in real-time by parsing the positions of medical personnel (using target detection algorithms based on depth vision sensors, such as YOLOv5, with a detection accuracy of ±20 mm) and device status signals. Intermittent period identification is achieved using time-series analysis algorithms (e.g., the sliding window method, with a window size of 5 seconds).

[0025] In a real-world emergency room scenario, suppose a patient with a lower limb fracture due to a car accident is brought into the emergency room. The posture monitoring module and environmental sensors on the bed immediately activate to execute step A. The posture monitoring module includes an electrically controlled pan-tilt unit fixed to the ceiling above the bed (tilt angle range -45° to 45°, horizontal rotation 0° to 360°, response time <0.5 seconds) and a depth vision sensor (such as an Intel RealSense D435, depth resolution 1280×720, frame rate 30fps). The sensor uses Time-of-Flight (ToF) technology to acquire depth image sequences of the patient's body, generating 640×480 resolution point cloud data with a depth accuracy of ±5 mm. The depth images are processed by a pre-trained MediaPipe3D skeletal tracking model (CNN-based, training dataset containing 5000 human posture samples, model inference time <50ms) to extract coordinate sequences of 20 joint points (e.g., left knee coordinates [x=0.5m, y=0.3m, z=0.1m]), with a temporal resolution of 30Hz. A 16×16 pressure sensor array (e.g., Tekscan, 10Hz sampling frequency) on the hospital bed recorded the pressure distribution in the patient's hips and legs. The results showed significantly lower pressure in the right leg compared to the left (10kPa for the right leg, 50kPa for the left leg), suggesting possible pain avoidance behavior. Environmental context information was obtained by collecting the speech of medical staff through a high-sensitivity microphone (44.1kHz sampling rate) installed beside the bed. Keywords such as "fracture fixation" and "X-ray examination" were extracted and processed using a Transformer-based speech-to-text model (pre-trained on a medical speech dataset, word error rate <5%). Simultaneously, an infrared sensor installed beside the bed monitored the X-ray machine's on / off status (1Hz sampling frequency), and a Bluetooth module acquired the infusion pump flow rate (50mL / h). To ensure non-interference with medical procedures, the system uses the YOLOv5 object detection algorithm (trained on 10,000 emergency room images, mAP@0.5 is 0.95) to detect the location of medical staff in real time. Combined with time-series analysis (5-second sliding window), it identifies the 8-second interval after medical staff complete X-ray positioning. Within this window, a depth vision sensor is triggered to collect 2 seconds of data (approximately 60 frames). The collected first pose information (depth image, skeletal coordinates, pressure distribution) and environmental context information (keyword vectors, device status) are input into the first AI model (CNN+LSTM architecture, pre-trained on 1000 fracture patient data). The model infers and outputs candidate injury mechanisms: femoral fracture (probability 0.85) and tibial fracture (probability 0.10). The entire process ensures that the sensor operation does not interfere with medical staff operations, and data acquisition is efficient and accurate.

[0026] Step B: When the first AI model outputs a candidate injury mechanism to be confirmed, according to the second objective, the patient's second posture information is collected and input into the first AI model. The second objective is to enable the second posture information to allow the first AI model to confirm or exclude candidate injury mechanisms, and to allow for moderate interference with medical procedures. The second posture information consists of a patient depth image sequence collected by the posture monitoring module according to the second objective, time-series data of three-dimensional coordinates of skeletal joints calculated based on the depth image sequence, and body pressure distribution data. The environmental context information consists of voice commands from medical personnel performing medical procedures, keyword fragments of dialogue content collected by audio acquisition devices, and status signals of medical equipment collected by environmental sensors. The non-interference with medical procedures includes: controlling the potential interference sensors in the posture monitoring module to operate in an intermittent acquisition mode based on environmental perception; specifically: identifying intervals in medical procedures by analyzing real-time environmental data, and triggering the potential interference sensors to rapidly acquire data only within the time window of the interval. The provision enabling the second posture information to be used by the first AI model to confirm or exclude candidate injury mechanisms includes: ensuring that the second posture information meets a first specific data acquisition requirement set for confirming or excluding the candidate injury mechanism; wherein the first specific data acquisition requirement is dynamically generated by the first AI model based on the expected biomechanical characteristics corresponding to the candidate injury mechanism. The provision allowing moderate interference with medical procedures includes: allowing the posture monitoring module to provide avoidance prompts to medical personnel obstructing the first necessary monitoring path when working along the first necessary monitoring path; wherein the first necessary monitoring path is a working path planned in real time by the posture monitoring module based on the first specific data acquisition requirement. The posture monitoring module includes at least one depth vision sensor mounted on an electronically controlled gimbal, which is fixedly installed on the ceiling above the patient's bed; the electronically controlled gimbal is configured to change the pitch angle, horizontal orientation, and optical focal length of the depth vision sensor according to control commands to achieve coverage of the first necessary monitoring path. The first posture information and the second posture information can reflect the patient's static forced posture characteristics and dynamic pain avoidance characteristics.

[0027] After the first AI model outputs candidate injury mechanisms, the patient's second posture information is collected according to the second objective (i.e., ensuring that the second posture information meets the requirements for confirming or excluding candidate injury mechanisms, while allowing for moderate interference with medical procedures), and then input into the first AI model for further analysis. The second posture information is similar to the first posture information, including depth image sequences, temporal data of three-dimensional coordinates of skeletal joints, and body pressure distribution data. The collection method is the same as in step A, but the collection parameters are dynamically adjusted according to the first specific data collection requirements. The first specific data collection requirements are generated by the first AI model based on the expected biomechanical characteristics of the candidate injury mechanism. For example, for a femoral fracture, the model may require the collection of dynamic rotation angles of the hip and knee (accuracy ±5°) or pressure distribution changes under specific postures (accuracy ±2 kPa). Biomechanical characteristics include the motion trajectory of skeletal joints (e.g., hip rotation angle range of 0°-30°) and pressure distribution patterns (e.g., pressure on the fracture side below 20 kPa). The accuracy of the injury mechanism is determined by comparing the deviation between the actual data and the expected characteristics. The depth vision sensor in the attitude monitoring module adjusts the pitch angle (-45° to 45°), horizontal orientation (0° to 360°), and optical focal length (range 28-50mm) via an electronically controlled pan-tilt unit to cover the first necessary monitoring path. This path is calculated in real-time by a path planning algorithm (such as the A* algorithm, based on a 3D spatial grid of the hospital bed with a resolution of 0.1m) according to the first specific data acquisition requirements, ensuring that the sensor's field of view covers key areas (such as the lower limbs). Moderate interference with medical procedures is achieved through avoidance prompts: when the position of medical personnel (detected via YOLOv5 with an accuracy of ±20mm) obstructs the first necessary monitoring path, the system prompts the medical personnel to move via a voice module (output volume 60-80dB) or an LED indicator (flashing frequency 2Hz), with a prompt duration of <3 seconds. The acquisition frequency of the second attitude information is adjusted according to the first specific data acquisition requirements (e.g., increased to 60Hz to capture rapid movements), with an acquisition duration of 5-10 seconds. After data is input into the first AI model, the model compares the matching degree between the second posture information and the expected biomechanical features, and outputs a confirmation or exclusion result (e.g., "Femoral fracture confirmed, probability 0.95"). The reasoning process of the first AI model is based on a pre-trained biomechanical feature library, combined with gradient boosting decision tree (GBDT) to rank the importance of features and optimize the selection of candidate mechanisms. The avoidance prompt is triggered by real-time analysis of the position of medical staff and the path planning results. The path planning algorithm takes as input the 3D spatial grid of the hospital bed and sensor field of view constraints, and outputs the optimal path (path length error < 0.1m).

[0028] Continuing with the emergency scenario described above, assuming the first AI model outputs a femoral fracture as the candidate injury mechanism in step A (probability 0.85), the system proceeds to step B to collect second posture information for verification. The depth vision sensor (Intel RealSense D435) of the posture monitoring module receives the first specific data acquisition requirement: capturing the dynamic rotation angles (accuracy ±5°) of the patient's right hip and knee, and the changes in pressure distribution in the right leg (accuracy ±2kPa). The electronically controlled pan-tilt unit (response time <0.5 seconds) adjusts the sensor's pitch angle to -30° using the A* path planning algorithm (based on a 0.1m resolution 3D mesh, computation time <100ms), aligning it horizontally with the patient's right leg, with a focal length set to 35mm to ensure the field of view covers the lower limb. The sensor acquires a 10-second depth image sequence (1280×720 resolution, approximately 600 frames) at a frequency of 60Hz, extracts the joint coordinates of the right hip and knee (accuracy ±10mm) using the MediaPipe 3D model, and calculates the rotation angle (e.g., hip rotation 15°). A pressure sensor array recorded a change in pressure in the right leg from 15 kPa to 18 kPa, indicating potential pain avoidance actions. During data acquisition, YOLOv5 detected a healthcare worker standing in front of the sensor's field of view, obstructing the necessary path for initial monitoring. The system issued a "Please move aside" prompt via a voice module (70 dB volume), while an LED flashed (2 Hz, for 2 seconds). After the healthcare worker moved, the sensors continued data acquisition. Environmental context information was updated synchronously, the microphone recorded keywords related to "fracture fixation" discussed by the healthcare worker, and the infrared sensor confirmed that the X-ray machine was off. Secondary posture information (depth image, skeletal coordinates, pressure distribution) was input into the first AI model. The model compared the data with the expected biomechanical characteristics of a femoral fracture (hip rotation <20°, pressure <20 kPa), confirming a femoral fracture (probability 0.95) and ruling out a tibial fracture (probability <0.05). Model inference combined with the GBDT algorithm (trained on 500 cases of fracture biomechanical data, feature importance ranking time <50 ms) optimized the output results. The entire process was completed efficiently under moderate interference, and the acquired data accurately supported the verification of the injury mechanism.

[0029] This invention significantly improves the accuracy and efficiency of injury mechanism identification in emergency scenarios by collecting patient posture information in stages and combining it with environmental context, and then using a first AI model to reverse-engineer and verify the injury mechanism. In the first stage, the method aims to minimize interference with medical procedures, ensuring minimal impact on the smoothness of the treatment process. In the second stage, moderate interference is used to obtain more accurate second posture information, further verifying or eliminating candidate injury mechanisms, thereby improving diagnostic reliability. This multi-target, staged acquisition strategy not only provides medical personnel with scientific auxiliary decision-making support but also optimizes the emergency treatment process, shortens diagnostic time, reduces the risk of misdiagnosis, and ultimately improves patient prognosis.

[0030] In the treatment of emergency trauma patients, real-time monitoring of patient posture is crucial for preventing secondary injuries. Traditional posture monitoring methods typically rely on fixed sensors or manual observation, which are ill-suited to the dynamic factors in the emergency environment, such as medical staff, mobile medical equipment, and caregivers. This leads to incomplete or interrupted data collection, affecting the accuracy of secondary injury warnings. Furthermore, current technologies lack systematic analysis and prediction of dynamic environmental factors (such as personnel movement trajectories and dwell time), making it difficult to achieve accurate path planning and data completion in complex environments.

[0031] Therefore, in some embodiments, the posture monitoring method for trauma emergency patients further includes: Step C: When the first AI model is performing a sitting motion analysis on the candidate injury mechanism, the third posture information of the patient is collected according to the third objective; wherein, the third objective is to enable the third posture information to provide the second AI model responsible for secondary injury warning to the patient for secondary injury warning. In step C, the patient's third posture information is collected according to the third target, specifically including: Step C1: Input the candidate injury mechanisms confirmed by the first AI model into the second AI model. The second AI model determines the set of undesirable postures that may cause secondary injury to the patient based on the mapping knowledge base between injury mechanisms and undesirable postures.

[0032] Step C1 aims to input the injury mechanism (e.g., femoral fracture) identified by the first AI model into the second AI model. The second AI model analyzes the correlation between the injury mechanism and the undesirable posture, generating a set of undesirable postures that may lead to secondary injury. The undesirable posture set refers to combinations of positions or movements that may aggravate the patient's injury or trigger new injuries; for example, improper turning over in a femoral fracture patient may lead to fracture displacement. The mapping knowledge base is a pre-built database that stores the correspondence between injury mechanisms and undesirable postures, including injury type (e.g., fracture, ligament tear), description of the undesirable posture (e.g., hip rotation exceeding 30°), biomechanical characteristics (e.g., joint stress exceeding 100N), and risk of secondary injury (e.g., fracture displacement probability 0.7). The mapping knowledge base is generated through clinical data and biomechanical simulations, with data sources including at least 5000 emergency patient cases and finite element analysis results. The data is stored in a relational database format, with fields including injury ID, posture description, and risk level. The second AI model is a deep learning-based classification model constructed using a multilayer perceptron (MLP) combined with an attention mechanism. The input consists of the injury mechanism (encoded as a vector, e.g., femoral fracture encoded as [1, 0, 0]) and environmental context information (keyword vectors, device status). The output is a set of undesirable poses (represented by pose labels and probabilities, e.g., "hip rotation > 30°, probability 0.85"). The model is trained through supervised learning. The dataset includes injury mechanisms, undesirable pose labels, and secondary injury cases. Training uses the Adam optimizer with a learning rate of 0.0005, and the loss function is weighted cross-entropy to handle imbalanced data. During inference, the model extracts biomechanical features of the injury mechanism (e.g., stress distribution at the fracture site), focuses on key features through the attention mechanism, and combines this with a mapping knowledge base query to match undesirable poses, generating a set of undesirable poses. Environmental context information was provided by the audio acquisition device (sampling rate 44.1 kHz, signal-to-noise ratio ≥ 60 dB) and environmental sensors (sampling frequency 1 Hz) in steps A and B, including healthcare personnel keywords (such as "turning over") and equipment status (such as the tightness of fixation devices, in N). The accuracy of the poor posture set generation was evaluated by F1 score, with a target F1 ≥ 0.9.

[0033] In an emergency scenario, assuming the first AI model confirms a femoral fracture (probability 0.95), the system initiates step C1, encoding the injury mechanism as a vector [1, 0, 0] and inputting it into the second AI model. The second AI model is an MLP+Attention model built on the PyTorch framework, containing three fully connected layers (512 neurons per layer) and one attention layer. The training dataset includes poor posture data from 6000 fracture patients (derived from hospital emergency room records and biomechanical simulations, trained for 100 epochs, F1 score 0.92). The model loads a mapping knowledge base (stored in a local SQLite database, containing 10,000 injury-posture mapping records) and queries for poor postures corresponding to femoral fractures, such as "hip rotation > 30° leading to fracture displacement, risk 0.8" and "knee hyperextension > 45° leading to ligament sprain, risk 0.6". The environmental context information is collected via microphone from medical staff instructions (such as "prepare to turn over"), and keyword vectors are extracted using a Transformer-based speech-to-text model (pre-trained on a medical speech dataset, word error rate <4%). An infrared sensor records the status of the fixation device (tightness 50N). A second AI model focuses on the biomechanical characteristics of the femoral fracture (such as the hip stress threshold of 120N) through an attention mechanism, and combines this with knowledge base matching to generate a set of abnormal postures: {"hip rotation > 30°, probability 0.85" "knee extension > 45°, probability 0.65"}. The inference process takes approximately 50ms, and the output is stored in the system cache for use in step C2. The entire process efficiently matches abnormal postures, ensuring highly targeted subsequent data collection and reducing the risk of secondary injury.

[0034] Step C2: Based on the set of undesirable postures, determine the second specific data acquisition requirements for capturing the key biomechanical features corresponding to each posture in the set of undesirable postures.

[0035] Step C2 aims to generate a second specific data acquisition requirement based on the set of undesirable postures. This ensures that the posture monitoring module can capture key biomechanical features corresponding to the undesirable postures for subsequent acquisition of third posture information. The second specific data acquisition requirement refers to the sensor parameter configuration needed to monitor undesirable postures (e.g., hip rotation > 30°), including acquisition frequency (Hz), field of view coverage (m²), joint point accuracy (mm), and pressure distribution resolution (kPa). Key biomechanical features include joint angles (e.g., hip rotation angle, accuracy ±5°), movement speed (e.g., rotational speed, ° / s), and pressure distribution patterns (e.g., fracture side pressure, kPa), determined by the biomechanical definition of the undesirable posture. For example, hip rotation > 30° requires the acquisition of dynamic angles of the hip joints and pressure changes in related areas. The process of generating the second specific data acquisition requirement is completed by a second AI model. The model derives the required parameters by analyzing posture labels and probabilities in the set of undesirable postures and combining this with a biomechanical knowledge base (containing posture-feature mappings, data from clinical trials and simulations, storing approximately 2000 records). For example, for hip rotation >30°, the model requires a sampling frequency ≥60Hz to capture rapid movements, a field of view covering the hip area (approximately 0.5m²), and joint accuracy of ±10mm. The biomechanical knowledge base is generated by parsing the mechanical constraints of poor postures (such as stress thresholds and joint range of motion). Data acquisition requirements are generated using a rule engine (based on the Drools framework), with rules defined as "if the posture label contains X, then the acquisition parameter is Y." Environmental context information (such as the status of the fixation device) is used to adjust the acquisition requirements; for example, if the fixation device tightness is less than 50N, the pressure distribution resolution needs to be increased to ±1kPa. The final output is a list of acquisition requirements containing specific parameters, stored in system memory for use in step C3.

[0036] Continuing with the emergency scenario, the second AI model outputs a set of undesirable postures: {"hip rotation > 30°, probability 0.85" and "knee extension > 45°, probability 0.65"}. The system then proceeds to step C2 to generate the second specific data collection requirements. The second AI model loads a biomechanical knowledge base (SQLite database, containing 2500 posture-feature mapping records, such as "hip rotation > 30° corresponds to stress > 120N") and analyzes the undesirable postures using the Drools rule engine (100 predefined rules, inference time < 20ms) to determine the collection requirements: hip rotation > 30° requires collecting the dynamic angle of the hip joint (accuracy ±5°, frequency 60Hz, field of view 0.5m²), and knee extension > 45° requires collecting the knee angle (accuracy ±5°) and leg pressure distribution (resolution ±1kPa). The environmental context information shows that the tightness of the fixation device is 40N (below the threshold of 50N), so the rule engine increases the pressure distribution collection frequency to 20Hz to monitor pressure changes caused by potential loosening. The final data acquisition requirements list is generated as follows: {“Hips: Angle ±5°, Frequency 60Hz, Field of view 0.5m²”, “Knees: Angle ±5°, Frequency 60Hz”, “Legs: Pressure ±1kPa, Frequency 20Hz”}. These requirements are transmitted to the posture monitoring module via system memory. The module adjusts the field of view of the depth vision sensor (Intel RealSense D435) to the hip area (0.5m², focal length 35mm), and sets the frequency of the pressure sensor array (16×16, Tekscan model) to 20Hz. The entire process precisely defines the acquisition parameters to ensure that subsequent monitoring can effectively capture the biomechanical characteristics of poor posture.

[0037] Step C3, controlling the posture monitoring module to perform operations includes: planning the second necessary monitoring path in real time according to the second specific data acquisition requirements; wherein, the second necessary monitoring path refers to the optimal sensor working trajectory planned by the posture monitoring module to fully cover the second specific data acquisition requirements; working along the second necessary monitoring path, collecting the patient's third posture information, and simultaneously collecting one or more influencing factors that may hinder the second necessary monitoring path; influencing factors include medical staff, mobile medical devices, or caregivers.

[0038] Step C3 aims to control the posture monitoring module to plan a second necessary monitoring path according to the second specific data acquisition requirements, collect the patient's third posture information along this path, and simultaneously identify factors that may obstruct the path (such as medical staff, mobile medical devices, and caregivers). These two processes are independent of each other (e.g., simultaneously determining the patient's posture and influencing factors in the depth vision sensor image; similarly, the subsequent acquisition of identity information and dynamic behavior data is independent of the acquisition of third posture information). The second necessary monitoring path refers to the optimal sensor working trajectory planned by the posture monitoring module to meet the acquisition requirements (e.g., hip angle accuracy ±5°). It is defined as a temporal sequence of the pitch angle (-45° to 45°), horizontal orientation (0° to 360°), and focal length (28-50mm) of the electronically controlled pan-tilt unit, with a path length error <0.1m. The path planning uses the A* algorithm, with inputs being a 3D spatial grid of the hospital bed (resolution 0.1m), acquisition requirements (e.g., field of view 0.5m²), and sensor constraints (e.g., field of view angle 60°), and outputs being a sequence of optimal trajectory points (each point includes angle and focal length). The third pose information includes depth image sequences (resolution 1280×720, frame rate 60Hz, depth accuracy ±5mm), 3D coordinates of skeletal joints (20 joints, accuracy ±10mm), and body pressure distribution data (resolution ±1kPa), acquired through a depth vision sensor (ToF technology) and a pressure sensor array. Influencing factor identification uses a target detection algorithm (e.g., YOLOv5, mAP@0.5 is 0.95), with inputs of depth images and infrared sensor data (sampling frequency 1Hz), and outputs of the influencing factor's category (medical staff, equipment, caregivers) and location (accuracy ±20mm). Path planning and data acquisition are achieved through real-time control loops at a frequency of 10Hz to ensure dynamic path adjustment. During acquisition, the system synchronously records the dynamic positions of influencing factors and stores them in the system cache for subsequent C4 analysis.

[0039] In an emergency scenario, the posture monitoring module receives a second specific data acquisition requirement (hip angle ±5°, frequency 60Hz, field of view 0.5m²; leg pressure ±1kPa, frequency 20Hz), initiating step C3. The electronically controlled pan-tilt unit in the module (response time <0.5 seconds) uses the A* algorithm (based on a 0.1m resolution 3D mesh, computation time <100ms) to plan the necessary path for the second monitoring, outputting the trajectory as: {t=0s: pitch -30°, orientation 90°, focal length 35mm; t=1s: pitch -25°, orientation 95°}, covering the hip area. A depth vision sensor (Intel RealSense D435) acquires depth image sequences at 60Hz, a MediaPipe 3D model extracts the coordinates of the hip joint points (e.g., [x=0.5m, y=0.3m, z=0.1m], accuracy ±10mm), and a pressure sensor array (16×16, frequency 20Hz) records leg pressure (e.g., 20kPa). Simultaneously, the YOLOv5 algorithm (trained on 10,000 emergency room images) analyzed depth images and infrared sensor data (1Hz), identifying a healthcare worker (position [x=1.2m, y=0.8m, z=1.5m]) and a mobile IV stand (position [x=1.0m, y=0.9m, z=0.5m]) as influencing factors, with a positional accuracy of ±20mm. The acquired third-person pose information and influencing factor data were stored in the system cache. The acquisition process took 10 seconds, with a path error of <0.1m. The entire process efficiently acquired patient pose information and accurately identified influencing factors, providing support for subsequent path correction.

[0040] Step C4: Collect identity information and dynamic behavior data of influencing factors.

[0041] Step C4 aims to collect identity information and dynamic behavior data of influencing factors that may hinder the necessary path for the second monitoring, providing a basis for subsequent obstacle prediction. Influencing factors include healthcare workers, mobile medical devices, or caregivers. Identity information refers to category labels (e.g., "healthcare worker," "IV stand") and unique identifiers (e.g., "healthcare worker ID001"), obtained through target detection algorithms and RFID tag readers. The target detection algorithm (e.g., YOLOv5, mAP@0.5 = 0.95) is based on image data from a depth vision sensor (1280×720 resolution, 30Hz frame rate), outputting the category and location of the influencing factor (accuracy ±20mm). An RFID reader (reading distance 2m, frequency 13.56MHz) scans pre-assigned tags to obtain unique identifiers. Dynamic behavior data includes the real-time location (x, y, z coordinates, accuracy ±20mm), movement speed (unit m / s, accuracy ±0.1m / s), and direction (angle, accuracy ±5°) of the influencing factor, collected through a depth vision sensor and an infrared sensor (sampling frequency 1Hz). Position and velocity are calculated from the depth image sequence using an optical flow algorithm (based on the Lucas-Kanade method), while orientation is obtained through vector analysis of the coordinate sequence. Data acquisition is performed at a frequency of 10Hz to ensure real-time dynamic behavior data. The acquired data is stored in a system cache in time-series format (each record includes time, category, identifier, position, velocity, and orientation) for use in step C5. The entire process is controlled in a real-time loop to coordinate sensor operation at a frequency of 10Hz.

[0042] For example, in an emergency room scenario, step C3 identifies a medical staff member and an IV stand as influencing factors, and the system proceeds to step C4 to collect their identity information and dynamic behavior data. A depth vision sensor (Intel RealSense D435, 30Hz frame rate) uses the YOLOv5 algorithm (trained on 10,000 emergency room images) to identify the medical staff member (category label "medical staff", location [x=1.2m, y=0.8m, z=1.5m]) and the IV stand (category label "equipment", location [x=1.0m, y=0.9m, z=0.5m]), with an accuracy of ±20mm. An RFID reader (frequency 13.56MHz) installed next to the bed scans the medical staff member's name tag (ID001) and the IV stand's tag (ID002) to confirm unique identification. Dynamic behavioral data was extracted from depth image sequences using an optical flow algorithm (Lucas-Kanade, computation time <30ms). The speed of medical personnel was 0.5m / s, with an orientation of 30° (relative to the bed coordinate system); the speed of the IV stand was 0.2m / s, with an orientation of 45°. An infrared sensor (1Hz) assisted in verifying the location. Data was acquired at a frequency of 10Hz, generating the following time series: {t=0s: Medical personnel, ID001, [x=1.2m, y=0.8m, z=1.5m], 0.5m / s, 30°; IV stand, ID002, [x=1.0m, y=0.9m, z=0.5m], 0.2m / s, 45°}. The data was stored in the system cache. The acquisition process took 5 seconds, efficiently providing the identity and behavioral information of influencing factors.

[0043] Step C5: Based on the collected identity information and dynamic behavior data of the influencing factors, predict the time window of obstruction of the second necessary monitoring path by the influencing factors and the corresponding alert response time; the alert response time refers to the expected time from issuing an avoidance warning to the influencing factor taking an avoidance action. Step C5 specifically includes the following sub-steps: Step C51: Based on identity information, match the corresponding inherent behavioral attributes from the pre-established object type-behavioral attribute database. The inherent behavioral attributes include typical movement speed, path preference, and typical dwell time.

[0044] The purpose of step C51 is to match the inherent behavioral attributes of influencing factors from the object type-behavioral attribute database based on the factors' identity information, providing a foundation for subsequent trajectory prediction. Inherent behavioral attributes include typical movement speed (m / s, accuracy ±0.1m / s), path preference (described by probability distribution, e.g., a tendency to move along the left side of the bed, probability 0.7), and typical dwell time (s, accuracy ±1s). The object type-behavioral attribute database is a pre-built SQLite database containing at least 1000 records. Fields include object type (medical staff, equipment, caregivers), typical movement speed (e.g., medical staff 1.0m / s), path preference (e.g., medical staff tend to walk around the head of the bed, probability 0.8), and typical dwell time (e.g., medical staff 10s). The database was constructed by analyzing historical emergency room surveillance data (at least 500 hours of video, covering 1000 medical procedures). Data extraction used statistical analysis (mean, variance) and a clustering algorithm (K-means, k=3). The matching process is implemented through database queries. The input consists of the category and identifier of the influencing factor (e.g., "Medical Staff ID001"), and the output is the corresponding behavioral attribute. The query time is <10ms, and the matching accuracy is ≥95%. Obtaining the behavioral attribute ensures that the behavioral habits of the target are considered in subsequent predictions, improving the accuracy of the time window of obstacles.

[0045] In the emergency room scenario, step C4 identifies a medical staff member (ID001) and an IV stand (ID002), and the system proceeds to step C51. The object type-behavioral attribute database (SQLite, 1200 records) is generated using K-means clustering (trained on 600 hours of emergency room video, k=3), containing attributes for the medical staff: typical movement speed 1.0 m / s, path preference "around the head of the bed, probability 0.8", typical dwell time 12 s; and attributes for the IV stand: speed 0.3 m / s, path preference "straight along the side of the bed, probability 0.9", dwell time 30 s. The system queries the database using "medical staff member ID001" and finds a match with the same attributes: speed 1.0 m / s, path preference "around the head of the bed, probability 0.8", dwell time 12 s; and queries using "IV stand ID002" and finds a match with the same attributes: speed 0.3 m / s, path preference "along the side of the bed, probability 0.9", dwell time 30 s. The query took 8ms with a matching accuracy of 98%. Behavioral attributes are stored in the system cache for use in subsequent steps. The entire process efficiently extracts behavioral features, providing reliable input for trajectory prediction.

[0046] Step C52: Based on dynamic behavior data and matched path preferences, predict the future movement trajectory of influencing factors using a Kalman filter algorithm with behavioral habit constraints; wherein, the behavioral habit constraints use path preferences as the prior probability distribution for trajectory prediction.

[0047] Step C52 uses dynamic behavioral data (position, velocity, direction) and path preferences to predict the future trajectory of influencing factors using a Kalman filter algorithm with behavioral habit constraints. The trajectory is defined as the position sequence of the influencing factors within the next 5 seconds (x, y, z coordinates, accuracy ±20 mm, time resolution 0.1 s). Dynamic behavioral data is provided by step C4, including real-time position (accuracy ±20 mm), velocity (accuracy ±0.1 m / s), and direction (accuracy ±5°). Path preferences (e.g., medical staff circling the bedside, probability 0.8) serve as the prior probability distribution for the Kalman filter, constraining the state transition matrix. The Kalman filter algorithm consists of two stages: state estimation and update. The state vector includes position and velocity (6-dimensional), and the observation vector is the position measurement from the depth vision sensor (3-dimensional). Behavioral habit constraints are implemented by adjusting the process noise covariance; for example, a probability of 0.8 for medical staff circling the bedside causes the trajectory to bias towards the bedside area (radius 0.5 m). The algorithm parameters include process noise (Q = 0.01 m² / s²), measurement noise (R = 0.02 m²), an iteration frequency of 10 Hz, and a prediction time window of 5 s. The algorithm is implemented in PyTorch, trained on at least 1000 emergency room trajectory datasets, and has a prediction error of <0.1 m. The output trajectory is a time series and stored in the system cache.

[0048] Continuing with the emergency room scenario, step C4 provides dynamic behavioral data for medical staff (ID001): position [x=1.2m, y=0.8m, z=1.5m], speed 0.5m / s, direction 30°; IV stand (ID002): position [x=1.0m, y=0.9m, z=0.5m], speed 0.2m / s, direction 45°. Step C51 provides path preferences: medical staff "around the head of the bed, probability 0.8", IV stand "along the side of the bed, probability 0.9". The Kalman filter algorithm (implemented in PyTorch, trained on 1200 trajectory data points, with an error of 0.08m) adjusts for process noise based on path preference to predict the trajectory of medical staff: {t=0.1s: [x=1.22m, y=0.82m, z=1.5m], t=0.2s: [x=1.24m, y=0.84m, z=1.5m]}, biased towards the head of the bed; the trajectory of the IV stand: {t=0.1s: [x=1.02m, y=0.91m, z=0.5m], t=0.2s: [x=1.04m, y=0.92m, z=0.5m]}, along the side of the bed. The prediction takes 50ms, and the trajectories are stored in a cache. The entire process accurately predicts the trajectory for the next 5 seconds, providing a basis for obstacle analysis.

[0049] Step C53: Perform spatiotemporal overlay analysis on the predicted future trajectory and the second necessary monitoring path to identify the intersection point of the future paths.

[0050] Step C53 compares the future trajectories of influencing factors with the second necessary monitoring path through spatiotemporal overlay analysis to identify the intersection points. An intersection point is defined as the point where the trajectory and path overlap in space (error < 0.1m) and time (error < 0.1s), representing the location and timing of potential obstacles. The future trajectory is provided by step C52 (5-second sequence, accuracy ±20mm), and the second necessary monitoring path is provided by step C3 (trajectory point sequence, error < 0.1m). The spatiotemporal overlay analysis uses a geometric collision detection algorithm. The input is the 3D coordinate sequence of the trajectory and path, and the output is the coordinates and time of the intersection point. The algorithm discretizes the trajectory and path into a set of points at 0.1s intervals, calculates the Euclidean distance between each pair of points, and marks an intersection point if the distance is < 0.1m and the time difference is < 0.1s. The algorithm is implemented using NumPy, with a computational complexity of O(n×m), where n and m are the number of points on the trajectory and path, respectively, with a typical value of approximately 50. Intersection points are stored in the system cache in the format {time, coordinates} for use in step C54. The analysis frequency is 10Hz to ensure real-time performance.

[0051] For example, in an emergency room scenario, the trajectory of medical staff (step C52) is {t=0.1s: [x=1.22m, y=0.82m, z=1.5m], t=0.2s: [x=1.24m, y=0.84m, z=1.5m]}, and the second necessary monitoring path (step C3) is {t=0.15s: [x=1.23m, y=0.83m, z=1.5m]}. A geometric collision detection algorithm (implemented in NumPy, computation time < 20ms) compares the point sets and finds an intersection point: {t=0.15s, [x=1.23m, y=0.83m, z=1.5m]}, with a distance error of 0.05m and a time error of 0.05s. The trajectories of the IV stand do not intersect. The intersection points are stored in a cache, and the analysis frequency is 10Hz. The entire process quickly identifies obstacles, providing accurate input for subsequent time window calculations.

[0052] Step C54: Based on the current movement speed of the influencing factors, the typical movement speed in the inherent behavioral attributes, the spatial location of the path intersection point, and the correction coefficient of the movement speed by the path preference, calculate the estimated arrival time of the influencing factors to the path intersection point.

[0053] Step C54 calculates the estimated arrival time (in seconds, accuracy ±0.1 seconds) of the influencing factors at the intersection point by combining the current movement speed, typical movement speed, path intersection point location, and path preference. The current movement speed is provided by step C4 (accuracy ±0.1 m / s), and the typical movement speed is provided by step C51 (accuracy ±0.1 m / s). The path intersection point is provided by step C53 (coordinate accuracy ±0.1 m). Path preference (e.g., around the head of the bed, probability 0.8) provides a speed correction coefficient (range 0.8-1.2), for example, a detour may reduce speed (coefficient 0.9). The calculation formula is: Arrival Time = Distance / Weighted Speed, Weighted Speed ​​= Current Speed ​​× 0.5 + Typical Speed ​​× 0.5 × Correction Coefficient, and Distance is the Euclidean distance from the current location of the influencing factor to the intersection point (calculated via NumPy, accuracy ±0.01 m). The correction coefficient is obtained through the path preference query rule engine (Drools, 100 rules), for example, around the head of the bed corresponds to a coefficient of 0.9. The calculation frequency is 10Hz, and the result is stored in the cache for use in step C56.

[0054] For example, in an emergency room scenario, the medical staff's current speed is 0.5 m / s (step C4), typical speed is 1.0 m / s (step C51), intersection point [x=1.23m, y=0.83m, z=1.5m] (step C53), current position [x=1.2m, y=0.8m, z=1.5m], path preference "around the bedside" with a correction coefficient of 0.9 (Drools query, time 5ms). The distance is calculated to be 0.042m (NumPy), weighted speed = 0.5×0.5 + 1.0×0.5×0.9 = 0.7 m / s, arrival time = 0.042 / 0.7 = 0.06s. The result is stored in the cache with an accuracy of ±0.1s. The entire process quickly calculates the arrival time, providing support for the generation of obstacle time windows.

[0055] Step C55: Based on the typical dwell time in the inherent behavioral attributes, and combined with the adjustment weight of dwell time by path preference and the environmental noise collected in real time by environmental sensors, the dwell time is jointly corrected to determine the expected dwell time of influencing factors at the path intersection point.

[0056] Step C55 calculates the expected dwell time (in seconds, with an accuracy of ±1 second) at the intersection of influencing factors based on typical dwell time, path preference, and environmental noise level. Typical dwell time is provided by step C51 (e.g., 12 seconds for healthcare workers), and path preference (e.g., circling the bedside, probability 0.8) provides adjustment weights (range 0.8-1.2), for example, circling may shorten the dwell time (weight 0.9). Environmental noise level is collected by an audio sensor (sampling rate 44.1 kHz, signal-to-noise ratio ≥60 dB) and calculated using sound pressure level (in dB, with an accuracy of ±1 dB). High noise level (e.g., >80 dB) increases dwell time (weight 1.1). The joint correction formula is: Expected Dwell Time = Typical Dwell Time × Path Weight × Noise Level Weight, with weights queried from the Drools rule engine (100 rules). The calculation frequency is 10 Hz, and the results are stored in a cache for use in step C56.

[0057] For example, in an emergency room scenario, the typical dwell time for medical staff is 12 seconds (step C51), and the path preference of "around the bedside" corresponds to a weight of 0.9 (Drools query). The audio sensor records a noise level of 85dB (>80dB threshold), corresponding to a weight of 1.1. The estimated dwell time = 12 × 0.9 × 1.1 = 11.88 seconds, approximately 12 seconds, and is stored in the cache. The entire process accurately calculates the dwell time, ensuring accurate prediction of the obstacle time window.

[0058] Step C56: Generate a barrier time window, starting from the estimated arrival time and using the jointly corrected estimated stay duration as the time window length.

[0059] Step C56 generates a time window for the obstacle using the estimated arrival time (step C54) and estimated dwell time (step C55), defined as [start time, start time + dwell time] (unit: seconds, accuracy: ±0.1 seconds). The arrival time (e.g., 0.06 seconds) and dwell time (e.g., 12 seconds) are directly combined to generate the time window, for example, [0.06 seconds, 12.06 seconds]. The generation process is achieved through simple concatenation and stored in the system cache in the format {start time, end time}. The calculation frequency is 10Hz to ensure real-time performance.

[0060] For example, in an emergency room scenario, medical staff arrive at 0.06s (step C54) and stay for 12s (step C55), generating an obstacle time window [0.06s, 12.06s]. The time window is stored in a cache, taking less than 5ms. The process is simple and efficient, providing crucial input for path correction.

[0061] Step C57: Based on the identity information, obtain the corresponding baseline reminder response time from the pre-established object type-response time database.

[0062] Step C57 uses the influencing factor identity information to query the baseline alert response time (in seconds, with an accuracy of ±0.1 seconds) from the object type-response time database. This is the expected time from issuing an avoidance alert to the influencing factor responding. The database (SQLite, 1000 records) contains object type (e.g., healthcare worker) and response time (e.g., 2 seconds for healthcare workers), generated through statistical analysis of 500 hours of emergency room video (mean, variance). The query input is a category and identifier (e.g., "healthcare worker ID001"), and the output is the response time. The query time is <10ms, and the accuracy is ≥95%. The results are stored in a cache for use in step C59.

[0063] For example, in an emergency room scenario, the database is queried using "Medical Staff ID001" to obtain a baseline reaction time of 2 seconds (the database is based on 600 hours of video, and the query takes 8ms). The results are stored in the cache to provide a basis for subsequent adjustments.

[0064] Step C58: Obtain the current environmental state parameters, including environmental noise level, ambient light intensity, and the real-time attention concentration index of the influencing factor.

[0065] Step C58 collects environmental parameters, including ambient noise level (dB, accuracy ±1dB), light intensity (lux, accuracy ±10lux), and attention concentration index (range 0-1, accuracy ±0.1). Noise level is calculated by an audio sensor (44.1kHz) using sound pressure level, and light intensity is measured by a photosensor (sampling frequency 1Hz). Attention concentration is analyzed using a target detection algorithm (YOLOv5) to determine influencing factors such as facial orientation (accuracy ±10°) and movement frequency (Hz, accuracy ±0.1Hz), mapping them to a 0-1 range (high-frequency movements correspond to low concentration). Data is collected at a frequency of 10Hz and stored in a cache.

[0066] For example, in an emergency room scenario, the audio sensor records a noise level of 85dB, and the photosensor records a light intensity of 500 lux. YOLOv5 analyzes the facial orientation of healthcare workers (10° away from the patient) and their movement frequency (0.5Hz), mapping an attention concentration of 0.7. Data is stored in a cache, with a data acquisition time of <50ms, providing support for reaction time correction.

[0067] Step C59: Using a pre-trained attention-reaction time mapping model, the baseline reminder reaction time is adjusted by multi-factor weighting based on the current environmental state parameters to generate the final dynamic reminder reaction time.

[0068] Step C59 uses an attention-reaction time mapping model to correct the baseline reaction time (step C57) based on environmental state parameters (noise level, light intensity, attention concentration) to generate a dynamic alert reaction time (in seconds, with an accuracy of ±0.1 seconds). The model is an XGBoost-based regression model, with environmental parameters (3-dimensional vectors) as input and the corrected reaction time as output. The model is trained on 1000 sets of emergency room scenario data (including noise level, light intensity, attention, and reaction time), using mean squared error loss and a learning rate of 0.01. The correction formula is: Dynamic reaction time = Baseline time × (1 + w1 × Noise normalization + w2 × Light intensity normalization + w3 × Attention), with weights w1, w2, and w3 optimized through grid search (w1 = 0.4, w2 = 0.2, w3 = 0.4). The calculation frequency is 10Hz, and the results are stored in a cache.

[0069] For example, in an emergency room scenario, the baseline reaction time is 2 seconds (step C57), with environmental parameters of 85 dB noise (normalized to 0.85), 500 lux illumination (normalized to 0.5), and attention level of 0.7. The XGBoost model (trained on 1200 datasets, MSE=0.01) calculates the dynamic reaction time as 2 × (1 + 0.4 × 0.85 + 0.2 × 0.5 + 0.4 × 0.7) = 3.44 seconds. The result is stored in a cache, taking 30 ms, providing accurate input for path correction.

[0070] Step C6: Continuously and dynamically correct the second necessary monitoring path to generate a corrected second necessary monitoring path, and control the attitude monitoring module to work along the corrected path, so that when the attitude monitoring module works along the corrected second necessary monitoring path, the following conditions are met: Condition 1: Within the time window of obstruction, the collection of the patient's third posture information must not be affected if the second specific data collection requirement is met. Condition 2: If Condition 1 cannot be met, if an influencing factor causes obstruction at any time within the obstruction time window, there is sufficient capability to continue predicting and supplementing the patient's third posture information within the reminder response time after the obstruction occurs.

[0071] The purpose of step C6 is to dynamically correct the second necessary monitoring path based on the obstruction time window and dynamic reminder response time generated in step C5, generate the corrected path, and control the attitude monitoring module to work along this path, ensuring that two conditions are met: Condition 1 is that the acquisition of the third attitude information is not interrupted within the obstruction time window, meeting the second specific data acquisition requirements; Condition 2 is that if Condition 1 cannot be met, the continuity of the third attitude information acquisition is ensured through predictive completion acquisition. The second necessary monitoring path refers to the sensor working trajectory planned by the attitude monitoring module to meet the second specific data acquisition requirements (such as hip angle accuracy ±5°, acquisition frequency 60Hz), including the time sequence of the pitch angle (-45° to 45°), horizontal orientation (0° to 360°), and optical focal length (28-50mm) of the electronically controlled pan-tilt unit, with a path error of less than 0.1m. The obstruction time window is provided by step C56 (e.g., [0.06s, 12.06s]), representing the time period during which influencing factors (such as medical personnel) may obstruct the path. The dynamic alert response time is provided by step C59 (e.g., 3.44s), representing the expected time from issuing the avoidance alert to the relocation of the influencing factor. Path correction employs a real-time path planning algorithm (based on the RRT algorithm, i.e., Fast Random Tree Optimization). Inputs include the second necessary monitoring path, obstruction time window, influencing factor location (accuracy ±20mm), dynamic behavior data (speed ±0.1m / s, direction ±5°), and environmental context information (medical staff keywords, equipment status). The RRT algorithm generates candidate paths by randomly sampling in a 3D spatial grid (resolution 0.1m), and evaluates path feasibility by combining the obstruction time window and the trajectory of the influencing factor (step C52). The optimization objective is to minimize the path length (error ±0.1m) and avoid the obstruction area (radius 0.2m). If condition one is met, the corrected path bypasses the obstruction area (by adjusting the gimbal angle and focal length) and continues acquisition. If condition one is not met, the system initiates predictive completion acquisition, using a posture prediction model based on a Long Short-Term Memory (LSTM) network. The input is the acquired third posture information (depth image sequence, skeletal coordinates, pressure distribution) and environmental context, and the output is the predicted posture data (e.g., hip angle sequence, accuracy ±5°). The posture prediction model is trained through supervised learning. The dataset contains posture sequences from at least 1000 emergency patients. Training uses the Adam optimizer with a learning rate of 0.001 and a loss function of mean squared error. Predictive completion is completed within the alert reaction time after the obstruction occurs. The completed data is fused with the real-time acquired data (weighted average, weight 0.7:0.3) to ensure acquisition continuity. Environmental context information is provided by an audio sensor (sampling rate 44.1kHz, signal-to-noise ratio ≥60dB) and an infrared sensor (sampling frequency 1Hz), containing keywords of medical staff instructions (e.g., "turn over") and equipment status (e.g., fixation tightness 50N).Path correction and data acquisition are coordinated in real-time control loops at a frequency of 10Hz. The depth vision sensor (resolution 1280×720, frame rate 60Hz, depth accuracy ±5mm) and pressure sensor array (16×16, resolution ±1kPa) of the posture monitoring module work synchronously to ensure that the third posture information meets biomechanical requirements (such as hip rotation angle ±5°).

[0072] For example, in an emergency scenario, assuming step C3 generates the second necessary monitoring path (trajectory points such as {t=0.15s: [x=1.23m, y=0.83m, z=1.5m], pitch -30°, focal length 35mm}), step C56 provides an obstacle time window [0.06s, 12.06s], step C59 provides a dynamic alert response time of 3.44s, and the system proceeds to step C6. The attitude monitoring module's electronically controlled gimbal (response time <0.5 seconds) is equipped with a depth vision sensor (Intel RealSense D435, frame rate 60Hz) and a pressure sensor array (16×16, frequency 20Hz). The RRT* algorithm (implemented in PyTorch, trained on 2000 emergency room path data points, optimization time <100ms) loads the second necessary monitoring path, obstruction time window, and influencing factor trajectory (medical staff position [x=1.23m, y=0.83m, z=1.5m], speed 0.5m / s). Candidate paths are generated by sampling in a 3D spatial grid (resolution 0.1m). Evaluation reveals that the original path overlaps with the medical staff trajectory at t=0.15s. The algorithm adjusts the gimbal tilt to -25° and orientation to 95°, generating a corrected path (error 0.08m) that bypasses the obstruction area (radius 0.2m). A depth vision sensor acquires a 10-second depth image sequence (1280×720, approximately 600 frames) along the corrected path. The MediaPipe 3D model extracts the hip joint coordinates (e.g., [x=0.5m, y=0.3m, z=0.1m], accuracy ±10mm), and a pressure sensor records leg pressure (e.g., 20kPa). If medical staff remain in the obstruction time window (e.g., 12 seconds), condition one cannot be met. The system then activates the LSTM posture prediction model (trained on posture sequences of 1500 patients, inference time <50ms), inputs the collected third posture information (5 seconds of data, approximately 300 frames) and environmental context (keyword "turning over", fixation device tightness 40N), and predicts the subsequent 5-second hip angle sequence (e.g., 15° to 18°, accuracy ±5°). The predicted data is fused with the real-time acquired data (weighted 0.7:0.3) to generate complete third posture information. The environmental context is extracted via the microphone (44.1kHz) using the keyword "turning over," and the infrared sensor confirms the fixation device status. The acquisition and prediction processes are coordinated via a 10Hz control loop, and the correction path takes effect within a 3.44s alert response time, ensuring continuous acquisition. Ultimately, the system efficiently acquires third posture information, meeting the second specific data acquisition requirements (hip angle ±5°, pressure ±1kPa), and supports secondary injury warning.

[0073] This technical solution significantly improves the accuracy and continuity of posture monitoring for emergency trauma patients by integrating a first AI model with a second AI model, combined with an adverse posture mapping knowledge base and dynamic path planning. Its core advantage lies in predicting the motion trajectory and obstruction time window of influencing factors through Kalman filtering and spatiotemporal overlay analysis, and dynamically correcting the alert reaction time using environmental state parameters. This ensures that the posture monitoring module can continuously collect key biomechanical characteristic data in complex emergency environments. Even when the path is obstructed, the system can still meet the needs of secondary injury early warning through predictive and supplementary data collection, reducing false alarm and false negative rates, and improving patient safety and medical efficiency.

[0074] In particular, steps C51 to C59 significantly improve the adaptability and real-time performance of posture monitoring for emergency trauma patients through systematic analysis of influencing factors and obstacle prediction. These steps first use database queries and Kalman filtering algorithms, combined with identity information, dynamic behavioral data, and environmental context, to accurately predict the movement trajectories and obstacle time windows of influencing factors (such as medical personnel and mobile devices). Spatiotemporal overlay analysis identifies path intersection points, and by integrating factors such as movement speed, path preference, and environmental noise, the estimated arrival time and dwell time are calculated to generate accurate obstacle time windows. Simultaneously, a pre-trained attention-reaction time mapping model is used to adjust the alert reaction time based on environmental conditions and attention concentration, ensuring accurate timing of avoidance prompts. This multi-level, multi-factor analysis and prediction mechanism fully considers the dynamic complexity of the emergency environment, significantly improves the robustness of path planning, and provides a reliable basis for the dynamic path correction of the subsequent posture monitoring module. This effectively reduces data acquisition interruptions caused by environmental interference, ensures the continuity and integrity of third-person posture information, and supports the accuracy of secondary injury warnings.

[0075] Step C6, through continuous dynamic correction of the necessary second monitoring path, ensures that the posture monitoring module can effectively cope with dynamic interference in complex environments during posture monitoring of emergency trauma patients, demonstrating significant beneficial effects. Its core advantage lies in ensuring the monitoring module is always in one of two favorable states: first, by dynamically adjusting the path to bypass obstacles such as medical personnel, mobile medical equipment, or caregivers, it satisfies condition one, ensuring continuous and accurate acquisition of third posture information, meeting specific data acquisition requirements; second, when bypassing obstacles is not feasible, relying on predictive and complementary acquisition capabilities, it generates continuous posture data within the alert reaction time after the obstacle occurs, satisfying condition two, thereby maintaining the reliability of secondary injury warnings. This dual-protection mechanism significantly improves the robustness and accuracy of posture monitoring in dynamic emergency environments, effectively reducing the risk of data loss or interruption due to path obstacles, and providing strong technical support for preventing secondary injuries to patients.

[0076] Figure 2This application provides a schematic diagram of a posture monitoring system for trauma emergency patients, as shown in the embodiment of the present application. Figure 2 As shown, the system includes: The first posture monitoring module 1 is used to collect the patient's first posture information and environmental context information when receiving patients with trauma emergencies, according to the first objective, and input them together into the first AI model responsible for inverse deduction of the injury mechanism; wherein, the first objective is not to interfere with medical treatment. The second posture monitoring module 2 is used to collect the patient's second posture information and input it into the first AI model according to the second objective when the first AI model outputs a candidate injury mechanism to be confirmed. The second objective is to enable the second posture information to be used by the first AI model to confirm or exclude the candidate injury mechanism, and to allow for moderate interference with medical treatment.

[0077] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method of monitoring the posture of a trauma emergency patient, characterized in that, Comprise: In the patient with trauma emergency, according to the first target, the first posture information of the patient and the environmental context information are collected and input into the first AI model responsible for the reverse deduction of the injury mechanism, wherein the first target is not to interfere with medical treatment; When the first AI model outputs the candidate injury mechanism to be seated, according to the second target, the second posture information of the patient is collected and input into the first AI model, wherein the second target is to enable the second posture information to be seated or excluded for the first AI model to the candidate injury mechanism, and to allow moderate interference with medical treatment.

2. The trauma emergency patient posture monitoring method of claim 1, wherein, The first posture information is a sequence of depth images of the patient collected by the posture monitoring module according to the first target, skeletal joint three-dimensional coordinate time series data calculated based on the depth image sequence, and body pressure distribution data.

3. The trauma emergency patient posture monitoring method of claim 1, wherein, The environmental context information is the voice instruction of the medical staff performing medical treatment, the keyword fragment of the dialogue content collected by the audio acquisition device, and the medical equipment state signal collected by the environmental sensor.

4. The trauma emergency patient posture monitoring method of claim 1, wherein, The non-interference medical treatment includes: Control the potential interference sensor in the posture monitoring module to work in an intermittent acquisition mode based on environmental perception, specifically: identify the intermittent period in the medical treatment operation by analyzing real-time environmental data, and only in the time window of the intermittent period, trigger the potential interference sensor to perform rapid data acquisition.

5. The trauma emergency patient posture monitoring method of claim 1, wherein, The second posture information enables the first AI model to seat or exclude the candidate injury mechanism, including: Make the second posture information meet the first specific data acquisition requirement set for seating or excluding the candidate injury mechanism; Wherein, the first specific data acquisition requirement is dynamically generated by the first AI model according to the expected biomechanical characteristics corresponding to the candidate injury mechanism.

6. The trauma emergency patient posture monitoring method of claim 5, wherein, The moderate interference with medical treatment includes: Allow the posture monitoring module to give avoidance prompts to medical staff who block the monitoring necessary path when working along the first monitoring necessary path; Wherein, the first monitoring necessary path is the working path of the posture monitoring module planned in real time according to the first specific data acquisition requirement.

7. The trauma emergency patient posture monitoring method of claim 6, wherein, The posture monitoring module includes at least one depth vision sensor mounted on an electric control holder, and the electric control holder is fixedly installed on the ceiling above the patient's bed; The electric control holder is configured to change the pitch angle, horizontal orientation and optical focal length of the depth vision sensor according to the control instruction to realize the coverage of the first monitoring necessary path.

8. The trauma emergency patient posture monitoring method of claim 1, wherein, Also include: When the first AI model seats the candidate injury mechanism, according to the third target, the third posture information of the patient is collected; wherein the third target is to enable the third posture information to be seated for the second AI model responsible for secondary injury warning to perform secondary injury warning on the patient; Wherein, collecting the third posture information of the patient according to the third target specifically includes: Input the candidate injury mechanism seated by the first AI model into the second AI model, and determine the set of adverse postures that may cause secondary injury to the patient based on the mapping knowledge base between injury mechanism and adverse posture by the second AI model; determine, based on the set of bad postures, a second specific data acquisition requirement for capturing key biomechanical features corresponding to each posture in the set of bad postures; controlling the posture monitoring module to perform operations including: planning a second necessary monitoring path in real time according to the second specific data acquisition requirement; wherein the second necessary monitoring path refers to an optimal sensor working trajectory planned by the posture monitoring module to completely cover the second specific data acquisition requirement; working along the second necessary monitoring path to collect third posture information of the patient, and synchronously collecting one or more influencing factors that may hinder the second necessary monitoring path; the influencing factors include medical staff, mobile medical equipment, or accompanying personnel; collecting identity information and dynamic behavior data of the influencing factors; based on the collected identity information and dynamic behavior data of the influencing factors, predicting a hindering time window of the influencing factors to the second necessary monitoring path and a corresponding reminder reaction time; the reminder reaction time refers to the expected time from issuing an avoidance prompt to the influencing factors making an avoidance action; continuously dynamically correcting the second necessary monitoring path, generating a corrected second necessary monitoring path, and controlling the posture monitoring module to work along the corrected path, so that when the posture monitoring module works along the corrected second necessary monitoring path, the following conditions are met: condition one, within the hindering time window, the third posture information of the patient continues to be collected in accordance with the second specific data acquisition requirement without being affected; condition two, when condition one cannot be met, if the influencing factors hinder at any time within the hindering time window, within the reminder reaction time after the hindering time, there is sufficient ability to continue to predict and complete the collection of the third posture information of the patient.

9. The trauma emergency patient posture monitoring method of claim 8, wherein, The prediction of the hindering time window of the influencing factors to the second necessary monitoring path and the corresponding reminder reaction time based on the collected identity information and dynamic behavior data of the influencing factors includes: based on the identity information, matching corresponding inherent behavior attributes from a pre-established object type-behavior attribute database; the inherent behavior attributes include typical moving speed, path preference, and typical stay duration; based on the dynamic behavior data and the matched path preference, predicting the future motion trajectory of the influencing factors through a Kalman filtering algorithm with behavior habit constraints; wherein the behavior habit constraints take the path preference as the prior probability distribution of trajectory prediction; spatiotemporally superimposing and analyzing the predicted future motion trajectory and the second necessary monitoring path to identify future path intersection points; based on the current motion speed of the influencing factors, the typical moving speed in the inherent behavior attributes, the spatial location of the path intersection points, and the correction coefficient of the moving speed based on the path preference, calculating the predicted arrival time of the influencing factors to the path intersection points; based on the typical stay duration in the inherent behavior attributes, and combining the adjustment weight of the stay duration based on the path preference and the real-time environmental noise level collected by the environmental sensor to jointly correct the stay duration, determining the predicted stay duration of the influencing factors at the path intersection points; taking the predicted arrival time as the starting point and the jointly corrected predicted stay duration as the time window length, generating the hindering time window; Based on the identity information, the corresponding reference reminding reaction time is obtained from a pre-established object type-reaction time database; Obtain the current environmental state parameters, including environmental noise, environmental light intensity and real-time attention concentration index of the influencing factors; Using the pre-trained attention-reaction time mapping model, the baseline reminding reaction time is weighted and corrected according to the current environmental state parameters, and the final dynamic reminding reaction time is generated.

10. A trauma emergency patient posture monitoring system, characterized by, Comprise: The first attitude monitoring module is used for collecting the first attitude information and the environmental context information of the patient when receiving the patient with traumatic emergency, and inputting the first AI model responsible for injury mechanism reverse deduction together; wherein the first target is not to interfere with medical treatment; The second attitude monitoring module is used for collecting the second attitude information of the patient when the first AI model outputs the candidate injury mechanism to be seated, and inputting the first AI model; wherein the second target is to enable the second attitude information to be seated or excluded for the first AI model to the candidate injury mechanism, and to allow moderate interference with medical treatment.

Images