A patient lying-to-standing position training path dynamic planning method based on an AI nursing assessment model

Abstract

The present application relates to the technical field of medical care and rehabilitation training equipment control, and particularly relates to a patient lying-to-standing training path dynamic planning method and system based on an AI nursing assessment model. The method collects patient nursing assessment data, historical training records, real-time vital signs and camera images, identifies patient posture, pipeline entry point, fixed point and suspension state, generates a stage state diagram and a multi-pipeline constraint diagram; obtains training node admission results, power limitation parameters and pipeline constraint results through an AI nursing assessment model, retains, deletes, downgrades or inserts a pause observation node for the training node, and forms a training path; switches to a safety fallback node according to pipeline, posture and vital sign abnormalities during training, and updates subsequent training parameters according to the training results. The present application can reduce the risk of pipeline pulling, bending and posture imbalance, and improve the safety and adaptability of bedside progressive training.

Description

Technical Field

[0001] This invention relates to the field of medical care and rehabilitation training equipment control technology, and in particular to a dynamic planning method and system for patient training path from lying to standing position based on an AI nursing assessment model. Background Technology

[0002] Patients who are bedridden for extended periods, postoperative patients, patients in the recovery phase of severe illness, and patients with neurological dysfunction typically require gradual training in areas such as bedridden elevation, sitting up at the bedside, sitting balance, partial weight-bearing, and assisted standing to improve postural tolerance, lower limb weight-bearing capacity, and daily activity abilities. Unlike general rehabilitation patients, patients in bedside care may be connected to multiple nursing lines, including IV lines, drainage tubes, urinary catheters, oxygen lines, and monitoring cables. During the transition from a lying to a sitting position, and then from a sitting to a standing position, the location of the tube inlets, the fixation position, and the allowance of the tubes will change with the change in position, easily leading to tube traction, bending, entanglement, insufficient allowance, or even accidental dislodgement. Therefore, bedside training from lying to standing requires consideration not only of the patient's muscle strength, level of consciousness, vital signs, and postural stability, but also the limitations imposed by multiple nursing lines on different positional transitions.

[0003] In the prior art, Chinese invention patent application CN119074416A discloses an integrated care and rehabilitation robot and its control method. This robot includes a support frame, a posture adjustment mechanism, a care bed, a rehabilitation training mechanism, a weight-reduction mechanism, and a control unit. The control unit can control the care bed to be in different postures and dynamically adjust the weight-reduction force provided to the user by the weight-reduction mechanism during rehabilitation training, thereby achieving multi-posture care and rehabilitation training. This technology integrates care bed posture adjustment, limb rehabilitation training, and weight-reduction assistance functions into a single device, providing support for patients to perform rehabilitation training in different postures. However, based on the content disclosed in CN119074416A, its control focus is mainly on care bed posture adjustment, rehabilitation training mechanism drive, and weight-reduction force adjustment; the training process is usually executed according to a predetermined posture or training mode. For patients connected to multiple nursing lines simultaneously, this technology does not further disclose how to identify the surface entry point, fixation point, and suspension status of each line, nor does it disclose how to control the deletion, downgrading, suspension of observation, or regression of training nodes based on the potential risks of traction, bending, or insufficient slack in each line during patient repositioning. Therefore, when performing continuous training from supine to standing positions on patients with multiple lines, nursing staff still need to organize the lines before training and continuously observe during the training. The safety and automation of the training are affected by the experience of the nursing staff and the timeliness of their rounds.

[0004] Chinese invention patent application CN108056898A discloses a virtual scene interactive rehabilitation training robot and its control method based on a lower limb linkage model and force sensory information. This technology treats the patient's thigh, calf, and foot as a three-link structure, using Kinect to detect the patient's leg posture and force sensors to detect the force between the patient's limbs and the rehabilitation robot. It combines foot force data and a preset weight reduction ratio to control the rehabilitation robot to perform progressive training. This technology can improve the interaction between the rehabilitation robot and the patient using visual posture information and force sensory information, and adjust the robot's movement according to the patient's weight-bearing status. However, CN108056898A mainly focuses on the motion control of lower limb rehabilitation robots. Its posture recognition and force sensory information are primarily used for lower limb weight compensation and robot motion adjustment. It does not address the phased training path planning from lying down, sitting at the bedside, partial weight-bearing, to assisted standing in bedside care, nor does it consider the constraints imposed by multiple tubing such as IV lines, drainage tubes, urinary catheters, and monitoring cables on training actions and sequences. At the same time, relying solely on posture data, plantar force data, or weight loss percentage for control is insufficient to fully reflect a patient's state of consciousness, fall risk, postural tolerance, real-time vital signs, and past training abnormalities. Consequently, it is difficult to automatically determine the entry criteria, maximum permissible training points, and abnormal regression targets for each training stage for different patients.

[0005] Furthermore, existing bedside training control methods typically treat patient ability assessment, equipment movement control, vital sign monitoring, and tubing care as independent processes. Nurses first select a training mode based on experience, the equipment executes the selected mode, and the status of monitoring equipment and tubing is observed separately by the nurses. When postural instability, insufficient active weight-bearing, abnormal vital signs, or tubing traction occur during training, the equipment is often uniformly paused or stopped. It is difficult to select the appropriate rollback stage based on the source of the abnormality, and it is also difficult to use the deleted, downgraded, or rolled-back training nodes to update the next training path. Summary of the Invention

[0006] This invention aims to provide a dynamic planning method for patient training paths from supine to standing positions based on an AI nursing assessment model. By integrating basic patient nursing assessment data, historical training records, real-time vital signs, visual posture information, and multi-tube status, a phased training path that meets the patient's ability and tube safety requirements is generated before training begins. During training, training nodes and safety retreat nodes are dynamically adjusted according to changes in patient posture, vital signs, and tube status. This addresses the problems of fixed training paths, reliance on manual monitoring of multi-tube risks, and lack of targeted abnormal handling in existing bedside training processes.

[0007] Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0008] A dynamic programming method for patient training path from supine to standing position based on an AI nursing assessment model includes the following steps:

[0009] S1. Collect patient basic nursing assessment data, historical training records, real-time vital signs and camera images, and use camera images to identify patient posture, surface tube inlet point, tube fixation point and tube suspension status.

[0010] S2. Generate a phased state diagram including the preparation period, bedside period and bed exit period training nodes based on the patient's body position and posture; generate a multi-tube constraint diagram based on the tube inlet point, tube fixation point and tube suspension status.

[0011] S3. Input the basic nursing assessment data, historical training records, real-time vital signs, stage status diagram and multi-circuit constraint diagram into the AI ​​nursing assessment model to obtain the admission results, assistance limitation parameters and tubing constraint results of each training node;

[0012] S4. Based on the admission results and pipeline constraint results, retain, delete, downgrade or insert pause observation nodes for the training nodes in the stage state diagram to generate the current training path from supine to standing position.

[0013] S5. During the training process, if the displacement of the pipeline inlet point, the pipeline overhang margin, the attitude stability or vital signs are detected to exceed the safety conditions of the corresponding training node, the bedside training equipment will be controlled to switch to the safety rollback node corresponding to the abnormality type.

[0014] S6. Based on the actual completed nodes, deleted nodes, downgraded nodes, paused observation nodes, rollback nodes and their abnormal reasons, update the node order, maximum allowed nodes and assist limit parameters of the next training path.

[0015] As a further improvement, in step S1, the basic nursing assessment data includes muscle strength level, state of consciousness, fall risk level, postural tolerance, pain score, tubing care risk level, and training contraindication information input by the nursing staff; the historical training record includes the highest training node previously completed, node holding time, equipment assistance parameters, reasons for abnormal termination, and post-training recovery time.

[0016] And / or, in step S1, the camera images include images of the patient's torso, hips, knees, ankles, feet, gripping areas, and exposed sections of tubing; the patient's posture identified by the camera images includes a supine elevated posture, a bedside sitting posture, a sitting balanced posture, a partially weight-bearing posture, and an assisted standing posture.

[0017] And / or, in step S1, the entry point of the body surface tubing is the connection position of the infusion tube, drainage tube, urinary catheter, monitoring cable, or oxygen tubing relative to the patient's body surface or nursing fixation point; the tubing fixation point is the fixation position of the tubing relative to the bedside training equipment, bed, patient clothing, or nursing fixation device; the tubing suspension state includes the tubing reserved length, bent state, crossed and wrapped state, and free swing state.

[0018] As a further improvement, in step S2, the stage state diagram includes training nodes, node sequence, node admission conditions, node retention conditions, and node regression conditions; the preparation period training nodes include supine elevation, lower limb positioning, and foot pre-contact nodes; the bedside period training nodes include bedside sitting up, sitting balance, and partial weight-bearing nodes; and the off-bed period training nodes include knee limitation, hip assistance, assisted standing, and standing retention nodes.

[0019] And / or, in step S2, the multi-pipeline constraint diagram includes a pipe inlet point, a pipe fixing point, a pipe overhang section, and a pipe risk edge. The pipe risk edge is used to represent the risks of traction, bending, entanglement, or insufficient margin that may occur in the corresponding pipe when the patient transitions from one training node to another.

[0020] As a further improvement, in step S3, the AI ​​nursing assessment model integrates basic nursing assessment data, historical training records, real-time vital signs, stage status diagrams and multi-channel constraint diagrams, and outputs the admission results of each training node in the preparation period, bedside period and bedside period, which are allowed to enter, restricted to enter, suspended observation or prohibited from entering respectively.

[0021] And / or, in step S3, the assist limiting parameters include back lifting speed, upper limit of back lifting angle, hip assist intensity, knee support limit status, foot support height, handrail locking status, and upper limit of standing holding time; the pipeline constraint results include at least one of allowing passage, reducing speed to pass, pausing to observe passage, and prohibiting passage.

[0022] As a further improvement, in step S4, when the pipeline constraint result corresponding to a certain training node is "no passage", the training node is deleted or downgraded to the previous stage training node; when the pipeline constraint result corresponding to a certain training node is "pause observation pass", a pause observation node is inserted before the training node, and the pipeline suspension status and patient posture are re-identified in the pause observation node.

[0023] And / or, in step S5, the abnormality types include abnormal tubing traction, abnormal tubing bending, abnormal tubing entanglement, abnormal posture imbalance, abnormal insufficient active weight-bearing, and abnormal vital signs; when the abnormality type is abnormal tubing traction or abnormal tubing bending, priority is given to switching to a safe retreat node that reduces the amplitude of body position changes; when the abnormality type is abnormal posture imbalance or abnormal insufficient active weight-bearing, priority is given to switching to the previous training node or the bedside training node; when the abnormality type is abnormal vital signs, priority is given to switching to the preparation training node or the training termination node.

[0024] As a further improvement, in step S6, if a node is deleted, downgraded, or rolled back due to pipeline constraints during the current training, the risk weight of the corresponding pipeline will be increased in the next training path, and the maximum allowable level of the training node related to that pipeline will be reduced; if the current training completes a preset number of training nodes consecutively without any abnormalities, the maximum allowable node in the next training path will be increased or the retention time of the corresponding node will be extended.

[0025] And / or, the method is used to control the mechanical execution actions and safety protection actions of a bedside training device, the bedside training device including at least one of a back lifting mechanism, a hip assist mechanism, a knee support limiting mechanism, a foot pedal support mechanism, a handrail locking mechanism, and a tubing follow-up management mechanism.

[0026] In a second aspect, the present invention also provides a system for implementing the method, including a data acquisition module, a visual recognition module, a stage state diagram generation module, a multi-pipeline constraint diagram generation module, an AI nursing assessment module, a coupled path planning module, an execution control module, and a training and update module;

[0027] The data acquisition module is used to collect patients' basic nursing care assessment data, historical training records, and real-time vital signs.

[0028] The visual recognition module is used to identify the patient's posture, the inlet point of the tube on the body surface, the tube fixing point, and the tube suspension status based on the camera image.

[0029] The stage state diagram generation module is used to generate a stage state diagram that includes training nodes for the preparation period, bedside period, and bed exit period.

[0030] The multi-pipe constraint diagram generation module is used to generate a multi-pipe constraint diagram based on the pipe inlet point, pipe fixing point, and pipe suspension state on the body surface.

[0031] The AI ​​nursing assessment module is used to output the admission results, assistance limitation parameters and pipeline constraint results of each training node based on basic nursing assessment data, historical training records, real-time vital signs, stage status diagram and multi-circuit constraint diagram.

[0032] The coupled path planning module is used to retain, delete, downgrade, or insert pause observation nodes for training nodes in the stage state diagram according to the admission results and pipeline constraint results, and generate the current training path from supine to standing position.

[0033] The execution control module is used to control the bedside training device to execute the training path, and to switch to the safe rollback node corresponding to the type of abnormality in case of an abnormality.

[0034] The training update module is used to update the node order, maximum allowed nodes, and assist limit parameters of the next training path based on the actual completed nodes, deleted nodes, downgraded nodes, paused observation nodes, rollback nodes, and their abnormal reasons.

[0035] As a further improvement, the visual recognition module includes a human key point recognition unit, a body posture discrimination unit, a pipeline inlet point recognition unit, and a pipeline suspension state recognition unit; the human key point recognition unit is used to recognize key points of the head and neck, shoulders, hips, knees, ankles, and feet, and the pipeline suspension state recognition unit is used to recognize the pipeline reserved length, bending state, cross-entanglement state, and free swinging state;

[0036] And / or, the multi-pipe constraint diagram generation module is configured to take the pipe inlet point, pipe fixing point and pipe overhang as constraint objects, and generate the corresponding pipe traction risk, bending risk, entanglement risk and insufficient margin risk according to the patient's body position change when switching from one training node to another training node;

[0037] And / or, the coupled path planning module includes a node filtering unit, a node degradation unit, a pause observation node insertion unit, and a safety rollback node configuration unit; the node filtering unit is used to delete training nodes that do not meet the admission results or pipeline constraint results; the node degradation unit is used to downgrade the off-bed training node to the bedside training node or the preparation training node; the pause observation node insertion unit is used to insert a pause observation node before the risk training node; and the safety rollback node configuration unit is used to configure corresponding safety rollback nodes for different anomaly types.

[0038] And / or, the execution control module is communicatively connected to the back lifting mechanism, hip assist mechanism, knee support limiting mechanism, foot support mechanism, handrail locking mechanism and pipeline follow-up management mechanism of the bedside training equipment, so as to control the patient to complete at least some of the training actions in the following order: lying up, sitting up at the bedside, partial weight-bearing and assisted standing, according to the current training path from lying down to standing.

[0039] And / or, the training update module is configured to, after a training session, write the actual completed nodes, abnormal nodes, rollback nodes, pipeline constraint results, vital sign changes, posture stability changes, and assist limit parameters into the historical training record, and use this as input for the AI ​​nursing assessment module to output the admission results, assist limit parameters, and pipeline constraint results for the next time.

[0040] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method described above.

[0041] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0042] This invention constructs a phased state diagram including training nodes for the preparation, bedside, and exit phases, and a multi-pipe constraint diagram including pipe inlet points, pipe fixing points, pipe overhangs, and pipe risk edges. These two diagrams are coupled to allow the limitations imposed by nursing tubing such as infusion tubes, drainage tubes, urinary catheters, oxygen lines, and monitoring cables on patient positioning during the training path generation phase to be identified and addressed in advance. Based on the training node admission results, assistance constraint parameters, and pipe constraint results output by the AI ​​nursing assessment model, training nodes can be retained, deleted, downgraded, or paused for observation, thus preventing equipment from being forced into inappropriate training phases according to a fixed procedure. During training, the system can adjust the parameters based on pipe inlet point displacement, pipe overhang margin, posture stability, and other factors. The system monitors changes in dynamic weight-bearing status and real-time vital signs, identifying tubing traction, bending, entanglement, postural imbalance, insufficient weight-bearing, or abnormal vital signs. Based on the type of abnormality, it switches to the corresponding safe reversal node. Compared to a unified pause or direct termination of training, this reduces unnecessary large-scale reversals, lowering the risk of accidental tube removal, falls, and postural intolerance. Simultaneously, the system records completed nodes, deleted nodes, downgraded nodes, paused observation nodes, reversal nodes, and the reasons for abnormalities in the historical training record. This information is used to update the node sequence, maximum allowed nodes, and assistance limit parameters for the next training path, allowing the training path to be continuously adjusted as the patient recovers. This improves the safety, adaptability, continuity, and nursing management efficiency of bedside reversal training. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall structure of the dynamic programming system for bedside lying to standing training path based on the coupling of multi-pipeline constraint diagram and stage state diagram of the present invention.

[0044] Figure 2This is a flowchart illustrating the dynamic planning method for training paths from bedside lying position to standing position according to the present invention.

[0045] Figure 3 This is a schematic diagram of the phased state of the training nodes of the preparation period, bedside period, and bedside period of the present invention.

[0046] Figure 4 This is a schematic diagram illustrating the process of generating a multi-pipeline constraint diagram based on the surface pipe inlet point, pipe fixing point, and pipe suspension state of the present invention.

[0047] Figure 5 This is a schematic diagram of the process of coupling planning of stage state diagram and multi-pipeline constraint diagram based on training node admission results and pipeline constraint results in this invention.

[0048] Figure 6 This is a schematic diagram of the structure of the AI ​​nursing assessment model of the present invention.

[0049] Figure 7 This is a schematic diagram of the anomaly identification and safety rollback control process during the training process of this invention.

[0050] Figure 8 This is a flowchart illustrating how the present invention updates the next training path based on the current training results.

[0051] Figure 9 This is a schematic diagram of the experimental verification stand for the bedside lying to standing training path dynamic planning system of the present invention.

[0052] Figure 10 This is a comparison chart showing the changes in tubing tension when a patient transitions from the knee limiting node to the hip assist node under different control methods.

[0053] Figure 11 This is a comparison chart of safe stabilization time, secondary risk incidence rate, and maximum trunk angular velocity under different abnormal handling methods.

[0054] Figure 12 A comparison chart of the admission recognition results for training nodes of the AI ​​nursing assessment model and different ablation models.

[0055] Figure 13 This is a comparison chart showing the changes in the completion rate and equipment assistance ratio of consecutive training sessions when the training update mechanism is used and when the training update mechanism is not used. Detailed Implementation

[0056] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It should be noted that the following embodiments are intended to enable those skilled in the art to clearly and completely understand and implement the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can make adaptive adjustments to the number of cameras, sensor types, number of training nodes, model network structure, data sampling frequency, and control thresholds according to patient type, nursing standards, bedside training equipment structure, and the information technology conditions of medical institutions.

[0057] This invention is applicable to long-term bedridden patients, postoperative recovery patients, critically ill recovery patients, patients with neurological dysfunction, and patients simultaneously connected to infusion tubes, drainage tubes, urinary catheters, oxygen lines, central venous catheters, or monitoring cables. Unlike transfer devices primarily intended for vertical patient transport or positional changes, the bedside training device controlled by this invention enables patients to sequentially complete at least some of the following training movements in the same training position: lying position preparation, sitting up at the bedside, sitting balance, partial weight-bearing, assisted standing, and standing maintenance. Before the bedside training device performs positional changes, it transforms the patient's ability, vital signs, visual posture, and nursing tubing constraints into an executable training path, and during training, switches to the corresponding safe retreat node according to the source of abnormality.

[0058] I. Terminology Explanation

[0059] In this embodiment, bedside lying to standing training refers to the training process in which the patient, with the assistance of at least one of the following: back lifting mechanism, foot support mechanism, knee support limiting mechanism, hip assist mechanism, handrail locking mechanism, and nursing staff, gradually transitions from an initial lying position to a bedside sitting position, a partially weight-bearing state, an assisted standing state, or a standing maintenance state.

[0060] The preparation phase refers to the training stage before the patient can achieve a stable bedside sitting position. It includes at least one training stage: initial supine position assessment, back elevation, positional tolerance observation, lower limb positioning, and pre-contact foot contact. The preparation phase is used to assess the patient's tolerance to changes in head, neck, and trunk height, and to establish lower limb position and nursing tubing leeway for subsequent bedside sitting.

[0061] The bedside period refers to the stage after a patient has transitioned from a supine to a sitting position at the bedside, and involves sitting balance, trunk forward leaning, and partial weight-bearing training with both feet on or near the ground. The bedside period is used to assess a patient's trunk control, sitting stability, and initial active weight-bearing ability.

[0062] The period of getting out of bed refers to the stage in which a patient, with the assistance of knee supports, footrests, handrails, back supports, or hip assist mechanisms, lifts their buttocks off the bed and enters a state of assisted standing or standing maintenance.

[0063] A phase state diagram is a graph structure that uses training nodes in the preparation, bedside, and bedside stages as nodes, directed edges as allowed transitions between nodes, and includes entry conditions, maintenance conditions, completion conditions, assist limit parameters, pause observation conditions, and safe retreat conditions in the nodes or directed edges. The phase state diagram includes both forward edges transitioning from low-intensity to high-intensity training nodes and retreat edges transitioning from the current training node to a low-risk training node in case of an anomaly.

[0064] A multi-pipeline constraint diagram is a graph structure that uses the pipe inlet point, pipe fixing point, and pipe guiding point on the body surface as constraint nodes, and the pipe overhang between adjacent constraint nodes as constraint edges. The constraint edges record the pipe type, pipe reserved length, pulling risk, bending risk, cross-entanglement risk, and insufficient margin risk.

[0065] The entry point of a surface-mounted tubing refers to the location where a nursing tubing connects to the patient's body surface, puncture site, insertion site, or medical accessory. This includes the connection point of an intravenous infusion tubing to the puncture site, the connection point of a drainage tubing to the drainage port, the exit point of a urinary catheter to the patient's lower abdomen or perineum, the connection point of an oxygen tubing to a nasal cannula or face mask, and the connection point of a monitoring cable to an electrode pad or sensor.

[0066] A tubing anchor point is a position where a nursing tubing remains relatively fixed relative to the patient's clothing, bed, bedside training equipment, tubing supports, or nursing fixtures. A nursing tubing may include one or more anchor points.

[0067] The suspended state of a pipeline refers to the spatial extension of a pipeline segment between adjacent pipeline inlet points, fixed points, or guide points on the body surface. This includes the pipeline's reserved length, minimum suspension position, degree of curvature, number of intersections, range of free swing, and whether the pipeline is located in the movement path of the equipment's moving mechanism.

[0068] The training node admission results include allowed entry, restricted entry, paused observation, and prohibited entry. Allowed entry means that the corresponding training node can be executed according to the planned parameters; restricted entry means that the corresponding training node can be executed, but the action speed needs to be reduced, the holding time shortened, equipment support increased, or the monitoring frequency increased; paused observation means that a temporary observation node needs to be entered first, and then the decision to continue is made based on the re-collected data; prohibited entry means that the corresponding training node cannot be entered during this training session.

[0069] A pause observation node is a temporary node inserted between two regular training nodes. During a pause observation node, the bedside training equipment maintains the current position or reduces the speed of movement, and the system re-acquires vital signs, patient posture, and multi-circuit status.

[0070] A safety retreat node refers to a target training node where bedside training equipment returns to its starting position at a predetermined speed and with a predetermined support sequence in the event of pipeline abnormalities, posture imbalances, insufficient active load, or abnormal vital signs. A safety retreat node can be a lower-level node within the current phase, a node from the previous training phase, a preparation observation node, or a training termination node.

[0071] Training node downgrading refers to replacing the originally planned training nodes with those that have lower range of motion, lower load requirements, or lower levels of active participation. For example, a standing hold node can be downgraded to an assisted standing node, and an assisted standing node can be downgraded to a partial load node.

[0072] II. System Structure

[0073] like Figure 1 As shown, the bedside lying-to-standing training path dynamic planning system of the present invention includes a data acquisition module, a visual recognition module, a stage state diagram generation module, a multi-path constraint diagram generation module, an AI nursing assessment module, a coupled path planning module, an execution control module, and a training update module.

[0074] The data acquisition module is used to collect basic patient care assessment data, historical training records, real-time vital signs, and operational status data of bedside training equipment. The data acquisition module can communicate with the hospital information system, nursing information system, vital sign monitoring equipment, bedside training equipment controller, and local training database.

[0075] Basic nursing care assessment data for patients includes muscle strength grade, level of consciousness, fall risk grade, pain score, postural tolerance, history of falls, risk grade of nursing tubing, and training contraindications recorded by nursing staff. Muscle strength grade can be expressed as 0 to 5; level of consciousness can be categorized as awake, drowsy, slow to respond, and uncooperative; fall risk can be categorized as low, medium, and high; and pain score can be expressed as 0 to 10.

[0076] Historical training records include the patient's highest previously achieved training milestone, the duration of each milestone, actual assist parameters of the bedside training equipment, abnormal occurrence points, abnormality types, safe rollback points, and recovery time after training. The system can read the most recent 3 to 10 training records; in this embodiment, the most recent 5 training records are read.

[0077] Real-time vital signs include at least three of the following: heart rate, systolic blood pressure, diastolic blood pressure, oxygen saturation, and respiratory rate. Heart rate and oxygen saturation may be measured once per second; respiratory rate may be measured once every 1 to 5 seconds; and blood pressure may be measured once every 1 to 5 minutes, with an additional blood pressure measurement before entering the bedside period and before leaving the bed.

[0078] The visual recognition module is communicatively connected to at least one camera. Preferably, a first camera is positioned above and in front of the bedside training device, and a second camera is positioned to the side of the device. The first camera is used to capture images of the patient's head and neck, shoulders, torso, hips, knees, ankles, feet, and the area where the patient holds the handrails. The second camera is used to capture images of the inlet point of the tubes on the body surface, the tube fixing point, the tube suspension section, and the relative position between the tubes and the device's movement mechanism.

[0079] The camera can be a color camera, a depth camera, or a combination of a color camera and a depth camera. The camera resolution can range from 1280×720 to 3840×2160, and the frame rate can be from 10 frames per second to 30 frames per second. In this embodiment, the camera resolution is 1920×1080, the human posture recognition frame rate is 10 frames per second, and the pipeline status update frame rate is 5 frames per second.

[0080] The phase state diagram generation module generates a phase state diagram based on the patient's current position, the executable actions of the bedside training equipment, and the training objectives. The multi-line constraint diagram generation module generates a multi-line constraint diagram based on the inlet point, fixation point, and suspension status of surface tubing, and calculates the risk status of each nursing tubing during the transition between different training nodes.

[0081] The AI ​​nursing assessment module receives basic nursing assessment data, historical training records, real-time vital signs, visual posture features, stage status diagram features, and multi-pipeline constraint diagram features, and outputs the admission results, assistance limit parameters, pipeline constraint verification results, and candidate safety rollback nodes for each training node.

[0082] The coupled path planning module is used to retain, delete, downgrade, or insert pause observation nodes for training nodes in the stage state diagram based on the admission results of training nodes and pipeline constraint results, and to configure safe rollback nodes for each training node, thereby generating the training path from the supine to the standing position.

[0083] The execution control module is communicatively connected to at least one of the following mechanisms of the bedside training equipment: back lifting mechanism, hip assist mechanism, knee support limiting mechanism, foot support mechanism, handrail locking mechanism, and tubing follow-up management mechanism. The execution control module converts the training path into mechanical execution commands and switches to the corresponding safety retreat node according to the type of anomaly during training.

[0084] The training update module is used to record the actual completed nodes, deleted nodes, downgraded nodes, paused observation nodes, safe rollback nodes and the reasons for the anomalies after training is completed, and to generate the node order, the maximum allowed nodes and the boost limit parameters for the next training path.

[0085] The aforementioned modules can be deployed within a single edge computing controller, while the execution control module can also be independently deployed within the programmable controller of the bedside training device. The edge computing controller and the programmable controller can communicate via Ethernet, Controller Area Network (CAN), or serial communication bus.

[0086] To ensure control safety, the execution control module is equipped with a hard constraint verification procedure independent of the AI ​​nursing assessment model. Training contraindications, equipment mechanical limits, drive mechanism load limits, and prohibitions from passing in multi-pipeline constraints entered by nursing staff all have higher priority than the output results of the AI ​​nursing assessment model.

[0087] III. Overall Method Flow

[0088] like Figure 2 As shown, the method of the present invention includes the following steps:

[0089] S1 collects basic nursing assessment data, historical training records, real-time vital signs and camera images of patients, and identifies patient posture, inlet point of surface tubes, tube fixation point and tube suspension status from camera images;

[0090] S2 generates a phased state diagram including the preparation period, bedside period and bed exit period training nodes based on the patient's body position and posture, and generates a multi-tube constraint diagram based on the tube inlet point, tube fixation point and tube suspension status on the body surface.

[0091] S3 inputs basic nursing assessment data, historical training records, real-time vital signs, stage status diagrams and multi-circuit constraint diagrams into the AI ​​nursing assessment model to obtain the admission results, assistance limitation parameters and circuit constraint results for each training node;

[0092] S4. Based on the training node admission results and pipeline constraint results, retain, delete, downgrade or insert pause observation nodes for the training nodes in the stage state diagram to generate the training path from supine to standing position for this time.

[0093] S5, during the training process, if the displacement of the pipeline inlet point, the pipeline overhang margin, the attitude stability or real-time vital signs are detected to exceed the safety conditions of the corresponding training node, the bedside training equipment is controlled to switch to the safety rollback node corresponding to the abnormality type.

[0094] S6 updates the node order, maximum allowed nodes, and assist limit parameters of the next training path based on the actual completed nodes, deleted nodes, downgraded nodes, paused observation nodes, safe rollback nodes, and their abnormal reasons.

[0095] (a) Step S1: Data acquisition and visual recognition

[0096] In step S1, the system first establishes a link between the patient's identity information and the current training task. The patient's identity information can be their hospital number, bed number, or a training task number generated by the system. After the nursing staff confirms the patient's identity, the current nursing care route type, and the training objective, the data acquisition module reads the patient's basic nursing assessment data.

[0097] When a patient's muscle strength grade is below 2, their level of consciousness is unable to cooperate, or there are training contraindications that prohibit positional changes, the system only generates low-amplitude observation nodes in the preparation phase; when a patient's muscle strength grade is between 2 and 3, the system can generate training nodes in the preparation and bedside phases, and set the out-of-bed phase as a candidate stage; when a patient's muscle strength grade is not lower than 3 and their historical training records are stable, further evaluation is conducted to determine whether to generate out-of-bed training nodes.

[0098] The system reads the five most recent training records. To ensure that recent training records have a significant impact on the current path planning, time weights of 0.35, 0.25, 0.18, 0.13, and 0.09 are assigned to the five most recent training records, respectively. If there are fewer than five actual training records, the weights of the existing training records are normalized.

[0099] Spatial calibration is performed before the camera is used. At least three calibration points are set on the fixed structure of the bedside training device, and a transformation relationship is established between the image coordinate system, the mechanical coordinate system of the bedside training device, and the patient training spatial coordinate system based on each calibration point. After spatial calibration is completed, the visual recognition module converts the patient's key points, the entry points of the body surface tubes, and the tube fixing points in the image into spatial coordinates relative to the bedside training device.

[0100] The visual recognition module first performs human key point identification. These key points include the head, neck, left and right shoulders, left and right hips, left and right knees, left and right ankles, and left and right feet. The trunk elevation angle is determined based on the angle between the line connecting the shoulders and hips and the bed surface; the position of the hips, trunk angle, and feet is used to determine whether the patient has entered a sitting position at the bedside; and the hip height, knee extension, and trunk uprightness are used to determine whether the patient has entered an assisted standing position.

[0101] To reduce instantaneous recognition errors caused by obstructions from bedding, caregivers, or bedside training equipment, median filtering is applied to the coordinates of key human body points in five consecutive frames. If a key human body point is lost for more than 0.5 seconds consecutively, the current posture recognition result is marked as unreliable. During the period when the posture recognition result is unreliable, the bedside training equipment must not proceed to the next training node and must maintain the current support state or revert to the previous stable training node.

[0102] The visual recognition module further performs nursing tubing segmentation, entry point identification, fixation point identification, and tubing tracking. Tubing segmentation can employ a semantic segmentation network to separate infusion tubing, drainage tubing, urinary catheters, oxygen tubing, and monitoring cables from patient clothing, bed sheets, and equipment structures.

[0103] In one implementation, the pipeline segmentation network employs an encoder-decoder structure. The encoder comprises four convolutional feature extraction stages, each consisting of a convolutional layer, a batch normalization layer, and a rectified linear activation layer, with the number of output channels for the four stages being 32, 64, 128, and 256, respectively. The decoder recovers pipeline edges through upsampling and skip connections, outputting the probability that each pixel belongs to the background, infusion tube, drainage tube, urinary catheter, oxygen line, or monitoring cable.

[0104] For nursing tubing of similar color and diameter, nursing staff can specify the tubing type and confirm the surface inlet point before training begins. The system establishes a tracking area around the inlet point and continuously updates the inlet point position using optical flow tracking, feature matching, or a target tracking network. Tubing anchorage points can be identified by recognizing anchoring clips, adhesive tape, tubing supports, or preset visual markers.

[0105] The visual recognition module extracts the pipeline centerline based on the pipeline segmentation results and uses the centerline between adjacent inlet points, fixed points, or guide points as the pipeline overhang. When a depth camera is used, the two-dimensional pipeline centerline is converted into a spatial curve; when only a color camera is used, the spatial shape of the pipeline can be estimated through binocular stereo matching or by combining the known fixed point positions.

[0106] The condition of a suspended pipe includes at least the actual length of the pipe, the straight-line distance between adjacent constraint points, the reserved length of the pipe, the minimum bending radius, the number of intersecting pipes, the range of free swing, and the distance between the pipe and the moving mechanism of the equipment. The reserved length of the pipe is the difference between the actual length of the suspended section of the pipe and the straight-line distance between adjacent constraint points.

[0107] When the pipeline segmentation confidence level is below 0.70, or when the critical pipeline entry point is lost for more than 1 second consecutively, the system prompts the nursing staff for manual confirmation; before the confirmation is completed, the off-bed training node related to the critical pipeline is marked as prohibited from entry.

[0108] The output of step S1 includes patient basic nursing assessment features, historical training feature sequence, real-time vital sign sequence, human posture key point sequence, set of surface tubing inlet points, set of tubing fixation points, and set of tubing overhangs. After being aligned according to a unified timestamp, the above data is sent to the stage state diagram generation module, the multi-tubing constraint diagram generation module, and the AI ​​nursing assessment module, respectively.

[0109] (II) Step S2: Generation of stage state diagram and multi-pipeline constraint diagram

[0110] The purpose of step S2 is to transform the patient's progressive training process and nursing pipeline constraints into a computable graph structure. For example... Figure 3 As shown, the stage state diagram includes the set of nodes for the preparation period, the set of nodes for the bedside period, and the set of nodes for the bedside removal period.

[0111] In one implementation, the preparation phase includes nodes P0 to P4, where P0 is the initial supine position assessment node, P1 is the back elevation to 30 degrees node, P2 is the back elevation to 45 degrees observation node, P3 is the lower limb positioning node, and P4 is the foot pre-contact node; the bedside phase includes nodes B1 to B4, where B1 is the back elevation to 60 degrees and lower limb turning to the bedside node, B2 is the bedside sitting-up node, B3 is the sitting balance node, and B4 is the partial weight-bearing node; the exit phase includes nodes O1 to O4, where O1 is the knee limiting node, O2 is the hip assist node, O3 is the assisted standing node, and O4 is the standing maintenance node.

[0112] Each training node includes a node identifier, the training stage it belongs to, the target position, the target holding time, entry conditions, holding conditions, completion conditions, the allowable range of assist parameters, and candidate safe retreat nodes. For example, the target position for node B3 is for the patient to maintain a sitting position at the bedside with the trunk offset angle relative to the vertical direction not exceeding 15 degrees; the target holding time is 30 to 180 seconds; entry conditions include that node B2 has been completed, the patient's vital signs tolerance level is not lower than the preset level, and the tubing restraint result is not prohibiting passage.

[0113] Node P1 has a back elevation angle of 25 to 35 degrees and a target holding time of 20 to 60 seconds; Node P2 has a back elevation angle of 40 to 50 degrees and a target holding time of 30 to 120 seconds; Node P4 requires the patient's soles to contact the foot support mechanism, with the contact area of ​​both soles reaching more than 50% of the corresponding sole area; Node B3 requires the patient's trunk swaying amplitude to not exceed the preset range for 10 consecutive seconds; Node B4 requires the total load borne by the patient's two feet to reach 10% to 30% of the patient's body weight; Node O3 requires the patient's buttocks to leave the bed surface and the knees to be within the knee support limit range; the standing holding time of Node O4 can be set to 10 to 120 seconds.

[0114] The directed edges in the phase state diagram represent the allowed transitions between training nodes. A normal training path sequentially passes through P0, P1, P2, P3, P4, B1, B2, B3, B4, O1, O2, O3, and O4. For patients with stable historical training and high ability, some low-risk observation nodes can be skipped; for intermediate- or high-risk patients, skipping observation nodes across training phases is not permitted. For example, before transitioning from the preparation phase to the bedside phase, a vital signs retest or a pause observation node should be completed.

[0115] The stage status diagram is generated based on the actual configuration of the bedside training equipment. When the equipment does not have a hip assist mechanism, an independent O2 node may not be generated, or the O2 node may be marked as requiring nursing staff assistance; when the equipment does not have a plantar pressure sensor, the completion conditions of the B4 node can be jointly determined by the patient's posture recognition results, changes in equipment support force, and nursing staff confirmation results.

[0116] Each training node transition also records the target movement amplitude, estimated execution time, mechanical actuator, and displacement of the candidate tubing inlet point. For example, when transitioning from P1 to P2, the main actuator is the back lifting mechanism; when transitioning from B2 to B3, the main change is in the patient's active trunk control state, with relatively small mechanical displacement; when transitioning from B4 to O1, the knee support limiting mechanism enters the working position; and when transitioning from O1 to O2, the hip assist mechanism begins to increase assistance.

[0117] like Figure 4 As shown, the multi-line constraint diagram establishes a sub-diagram for each nursing line. The sub-diagram includes at least the surface line inlet point, the line anchor point, and the line overhang between adjacent constraint points. The surface line inlet point and the line anchor point serve as graph nodes, and the line overhang as graph edges.

[0118] Each graph node must record at least the node type, spatial coordinates, the object it belongs to relative to the patient or device, and the node's reliability. Each graph edge must record at least the tubing type, actual length, straight-line distance between adjacent constraint points, reserved tubing length, minimum bending radius, number of intersecting tubing, whether it passes through the device's movement area, and tubing tension.

[0119] For any training node transition edge in the phase state diagram, the system calculates the expected displacement of the surface tube inlet point relative to the tube fixing point based on the patient's expected posture at the target training node, and re-estimates the tube reserved length, minimum bending radius, and crossing state under the target training node.

[0120] The patient's expected posture can be calculated based on the target mechanical position of the bedside training equipment and the patient's body dimensions. Patient body dimensions include trunk length, thigh length, calf length, shoulder width, and hip width. These dimensions can be calculated using visual key points or input by nursing staff. For patients who have completed the same training node, the actual key point positions from the patient's historical training are preferentially used as the target node posture.

[0121] The pipeline risk value of the j-th pipeline on the transformation edge e of the training node is determined by the following formula:

[0122] ;

[0123] In the formula, Let be the pipeline risk value of the j-th pipeline on the transition edge e at the training node; Risk coefficient for pipeline type; To adjust the risk normalization value; This is the normalized value for bending risk; This is the normalized value for the risk of cross-entanglement; To reserve a normalized value for the risk of insufficient length; the value of each normalized value is in the range of 0 to 1.

[0124] The weights for the risks of traction, bending, cross-entanglement, and insufficient reserved length are 0.40, 0.25, 0.20, and 0.15, respectively. The risk of traction has the highest weight and is used to prioritize limiting the transition of training nodes in situations that may lead to accidental detachment.

[0125] The risk coefficient for each tubing type is set based on the potential impact of tubing abnormalities. The risk coefficients are as follows: central venous catheter: 1.30; surgical drainage tube: 1.25; ordinary intravenous infusion tubing: 1.10; urinary catheter: 1.00; oxygen tubing: 0.90; and ECG monitoring cable: 0.80. Risk values ​​greater than 1 are recorded as 1.

[0126] The normalized value for traction risk is determined based on the ratio of the expected inlet point displacement to the current reserved pipeline length. When the expected inlet point displacement does not exceed 50% of the current reserved length, the normalized value for traction risk is 0 to 0.30; when the expected inlet point displacement is 50% to 80% of the current reserved length, it is 0.30 to 0.70; and when the expected inlet point displacement exceeds 80% of the current reserved length, it is 0.70 to 1.00.

[0127] When tension sensors are installed at fixed points in the pipeline, the data from these sensors can be used to correct for traction risks. When the actual pipeline tension reaches 50% of the corresponding tension threshold, the normalized value for traction risk should not be lower than 0.30; when it reaches 80% of the tension threshold, the normalized value for traction risk should not be lower than 0.70; and when the tension threshold is reached or exceeded, the normalized value for traction risk should be 1.00.

[0128] The normalized value for bending risk is determined based on the minimum bending radius of the pipeline. When the minimum bending radius of the pipeline is greater than 1.5 times the corresponding minimum allowable bending radius, the normalized value for bending risk is 0 to 0.20; when the minimum bending radius of the pipeline is 1 to 1.5 times the minimum allowable bending radius, it is 0.20 to 0.70; when the minimum bending radius of the pipeline is less than the minimum allowable bending radius, it is 0.70 to 1.00.

[0129] The normalized value for cross-entanglement risk is determined based on the number of pipe crossings, the duration of the crossing, and the location of the crossing. When two pipes only cross in a 2D image and are separated in the depth direction, the normalized value for cross-entanglement risk is 0; when there is a contact crossing but it is not located in the movement path of the equipment, the value is 0.30 to 0.60; when the crossing location is located in the movement path of the knee support, foot pedal, or hip assist mechanism, the value is 0.70 to 1.00.

[0130] The normalized value for insufficient reserve length risk is determined based on the expected remaining reserve length after the patient reaches the target training node. When the expected remaining reserve length is greater than 1.5 times the preset safe length, the normalized value for insufficient reserve length risk is 0 to 0.20; when the expected remaining reserve length is 1 to 1.5 times the preset safe length, it is 0.20 to 0.60; and when the expected remaining reserve length is less than the preset safe length, it is 0.60 to 1.00.

[0131] The preset safe length for a standard infusion tube is 100 mm, for a surgical drainage tube it is 150 mm, for a urinary catheter it is 120 mm, for an oxygen line it is 100 mm, and for an electrocardiogram monitoring cable it is 80 mm.

[0132] For multiple pipelines on the same training node's transition edge, the maximum value among the risk values ​​of each pipeline is taken as the pipeline constraint risk value of the transition edge of that training node, so as to avoid using the average value to mask a single high-risk pipeline.

[0133] When the pipeline constraint risk value is below 0.35, the pipeline constraint result is allowed to pass; when the risk value is greater than or equal to 0.35 and less than 0.55, the speed should be reduced to pass; when the risk value is greater than or equal to 0.55 and less than 0.75, the observation should be suspended before passing; when the risk value is greater than or equal to 0.75, the passage is prohibited.

[0134] The multi-pipeline constraint graph generation module sends the pipeline constraint risk value, risk source pipeline, main risk type, and pipeline constraint result corresponding to the transformation edge of each training node to the AI ​​nursing assessment module and the coupled path planning module. Thus, multi-pipeline risks participate in node planning during the training path generation phase, rather than only triggering an alarm after a pull occurs during training execution.

[0135] (III) Step S3: Structured Disclosure of AI Nursing Assessment Model

[0136] like Figure 6 As shown, the AI ​​nursing assessment model includes a nursing assessment feature encoding branch, a historical training encoding branch, a vital signs time-series encoding branch, a visual pose encoding branch, a stage state graph encoding branch, a multi-channel constraint graph encoding branch, a feature fusion layer, and a multi-task output layer.

[0137] 1. Model Dataset Construction

[0138] The data used to train the AI ​​nursing assessment model includes historical bedside training data, equipment debugging data, and simulation training data verified by healthcare professionals. Clinically used data undergoes de-identification processing before being included in the training dataset, removing patient names, identification numbers, contact information, and other identity information irrelevant to training control.

[0139] A training sample is based on the basic unit of "a patient's intended entry into a target training node during a single training session". For example, when a patient intends to enter node B4 from node B3, the nursing assessment data before entering B4, historical training data, recent vital signs, visual posture sequence, local structure of the stage state diagram, local structure of the multi-path constraint diagram, and the final training result constitute a sample.

[0140] Each sample must include at least the following fields:

[0141] 1) Basic nursing fields, including muscle strength grade, level of consciousness, fall risk grade, pain score, postural tolerance and training contraindications;

[0142] 2) Historical training fields, including the highest completion node, node retention time, anomaly type, node where the anomaly occurred, safe rollback node, and post-training recovery time in the last 5 training sessions;

[0143] 3) Vital signs fields, including time series of heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen saturation and respiratory rate in the 10 minutes before entering the target training node;

[0144] 4) Visual posture fields, including the sequence of human key points, trunk tilt angle, knee angle, foot contact status and handrail grip status within 5 to 30 seconds before the target training node enters;

[0145] 5) Stage state graph fields, including the stage to which the target training node belongs, entry conditions, target position, target holding time, range of assist parameters, and adjacent training nodes;

[0146] 6) Multi-pipeline constraint diagram fields, including each pipeline type, inlet point location, fixed point location, pipeline reserved length, pulling risk, bending risk, entanglement risk, and insufficient reserved length risk;

[0147] 7) Output label fields, including training node admission label, assist limit parameter label, pipeline constraint result label, and safety rollback node label.

[0148] The training node admission labels include Allowed Entry, Restricted Entry, Suspended Observation, and Prohibited Entry. Allowed Entry indicates that the patient completed the target training node according to the planned parameters without any abnormalities requiring increased support or retreat; Restricted Entry indicates that the patient completed the target training node, but reduced the movement speed, increased support, or shortened the holding time during actual execution; Suspended Observation indicates that an observation step was added before entering the target training node; Prohibited Entry indicates that the target training node was not executed, or was immediately retreated after execution due to risk.

[0149] The assisted limitation parameter labels include the upper limit of back lift speed, the upper limit of back lift angle, hip thrust intensity, knee support limit status, foot support height, handrail locking status, and the upper limit of standing time. For samples that have completed the target training nodes, the parameters at which they were actually safely completed are used; for samples that have not performed the target training nodes, the upper limits of the safe parameters confirmed by the rehabilitation physician and the responsible nurse are used.

[0150] The tubing constraint result labels include Allow Passage, Reduced Speed ​​Passage, Suspended Observation Passage, and Prohibited Passage. The labels are determined based on the nursing staff's confirmation of the tubing status, tubing tension sensor records, visual tubing status, and whether actual pulling, bending, or entanglement has occurred.

[0151] The safe rollback node label is the target node that was actually used after the anomaly occurred and confirmed safe by nursing staff. When an anomaly handling process passes through multiple rollback nodes in sequence, the first stable node is used as the primary label, and the remaining nodes are recorded as candidate rollback sequences.

[0152] Labels are jointly confirmed by at least one rehabilitation physician and one responsible nurse. If their opinions are consistent, the label is used directly as the final label; if their opinions are inconsistent, it is reviewed by a qualified senior medical professional. For parameters whose output purpose is clear and can be directly determined from the device records, such as back elevation speed and actual retraction point, initial labels can be automatically generated by the device records, and then verified by medical professionals through sampling.

[0153] In a feasible data scale example, the dataset includes training records from no fewer than 300 patients, with each patient including 5 to 30 training sessions. Each training session contains 5 to 13 target training node samples, resulting in no fewer than 10,000 node-level samples. When the data scale is insufficient, a rule-based model can be run first, and a machine learning model can be trained after accumulating no fewer than 3,000 valid node-level samples.

[0154] To avoid information leakage caused by the same patient's data appearing in both the training and testing sets, the dataset is divided according to the patient dimension. The proportions of patients in the training, validation, and testing sets are 70%, 15%, and 15%, respectively. All training records for the same patient belong to only one subset of the data.

[0155] To address the issue of imbalanced admission categories at training nodes, class-weighted cross-entropy or stratified sampling can be employed. The loss weights for the four categories—allowed entry, restricted entry, paused observation, and prohibited entry—can be determined by normalizing the inverse of the sample size for each category. In one implementation, the loss weights for the four categories are 1.00, 1.30, 1.45, and 1.60, respectively, to improve the model's ability to identify paused observation and prohibited entry samples.

[0156] 2. Data Preprocessing

[0157] Discrete features of nursing assessments were encoded using one-hot coding or embedding coding. Muscle strength grades, pain scores, and continuous vital signs were standardized using the mean and standard deviation of the training set. Linear interpolation was used when the time of missing vital signs was no more than 30 seconds; when blood pressure data was collected intermittently, the most recent valid value was retained, and a validity indicator for the blood pressure data was added.

[0158] When a vital sign is continuously missing for more than a set time, unconditional interpolation is not performed; instead, the corresponding data validity flag is set to invalid. The input to the AI ​​nursing assessment model includes both data values ​​and validity flags, enabling the model to distinguish between true stable values ​​and missing values.

[0159] Human body key point coordinates are converted to normalized coordinates relative to the patient's hip center or the bedside training device coordinate system to reduce the impact of camera installation position and patient height differences. The pose sequence uniformly extracts data from the first 10 seconds of the target training node, forming a 100-frame sequence at 10 frames per second; if there are fewer than 100 frames, front-end padding is used, and an effective frame mask is set.

[0160] When the number of nodes in the stage state graph and the multi-pipeline constraint graph are different, a graph neural network is used to directly process the variable-length graph structure; during batch training, multiple graph samples are merged using a batch graph method, and the graph index of each sample is retained.

[0161] 3. Model Network Structure

[0162] The nursing assessment feature encoding branch is used to process muscle strength level, level of consciousness, fall risk level, pain score, postural tolerance, and training contraindications. Discrete features are converted into embedding vectors, and continuous features are normalized and concatenated. These vectors are then sequentially fed into a two-layer fully connected network with 64 neurons and 32 neurons, respectively. Each layer is followed by a batch normalization layer and a rectified linear activation function.

[0163] The historical training encoding branch is used to process the most recent 5 training records. Each training record is encoded as a feature vector containing the highest completion node, the abnormal node, the backtracking node, the hold time, and the recovery time. It is input into a gated recurrent unit with 32 hidden units, and the hidden state at the last moment is taken as the historical training feature.

[0164] The vital signs timing encoding branch is used to process the vital signs sequence within the 10 minutes prior to the target node's entry. This branch employs a gated cyclic unit with 64 hidden units. When the vital signs sampling frequency is inconsistent, it is first uniformly resampled to a time point every 5 seconds. Vital signs data values ​​and validity indicators are used as inputs.

[0165] The visual pose encoding branch connects key points of the head and neck, shoulders, hips, knees, ankles, and feet according to the human body topology to form a spatiotemporal pose map, and uses a two-layer spatiotemporal graph convolutional network to extract pose features. The first layer has an output dimension of 64, and the second layer has an output dimension of 64; the temporal convolutional kernel length is 3 frames, and the spatial adjacency relationship is determined according to the human skeleton.

[0166] The phase state graph encoding branch uses a two-layer graph attention network to handle the transition edges between training nodes. The output dimension of both the first and second layers is 64, with four attention heads per layer. Training node attributes include the phase, target position, target hold time, entry conditions, assist parameter range, and candidate safe retreat nodes; transition edge attributes include the target motion amplitude, estimated execution time, main actuator, and whether it crosses a phase.

[0167] The multi-pipeline constraint graph encoding branch uses a two-layer graph attention network to process the pipe entry points, fixed points, overhanging sections, and risk edges on the pipe surface. Both the first and second layers have an output dimension of 64, with four attention heads per layer. Node attributes include constraint point type, spatial location, associated pipe, and recognition confidence; edge attributes include pipe type, reserved length, tension risk, bending risk, entanglement risk, insufficient margin risk, and pipe constraint result.

[0168] The outputs of each coding branch are concatenated in the feature fusion layer. The concatenated features are then sequentially input into two fully connected networks with 128 neurons and 64 neurons respectively. A random deactivation layer with a random deactivation ratio of 0.20 is placed between the two fully connected networks.

[0169] The multi-task output layer includes a training node admission output head, an assist constraint parameter output head, a pipeline constraint verification output head, and a safety backoff node output head. The training node admission output head uses four output neurons and a Softmax function to output the probabilities of allowed entry, restricted entry, paused observation, and prohibited entry. The assist constraint parameter output head uses a linear output layer to output continuous control parameters, truncated within the device's allowable range. The pipeline constraint verification output head uses four output neurons to output four types of pipeline constraint results. The safety backoff node output head generates candidate node probabilities based on candidate backoff nodes in the stage state diagram.

[0170] 4. Model Training Methods

[0171] The AI ​​nursing assessment model employs a multi-task joint training approach. The comprehensive loss function is expressed as:

[0172] ;

[0173] In the formula, The overall loss of the model; The classification loss is used for the admission of training nodes; To help limit the regression loss of parameters; Classify the losses based on pipeline constraint results; Classify the losses for safe rollback nodes.

[0174] The cross-entropy loss is used for the training node admission classification loss, the pipeline constraint result classification loss, and the safe backoff node classification loss; the smoothed absolute error loss is used for the assist constraint parameter regression loss. The weights of the training node admission classification loss, the assist constraint parameter regression loss, the pipeline constraint result classification loss, and the safe backoff node classification loss are 0.35, 0.25, 0.25, and 0.15, respectively.

[0175] The model employs an adaptive moment estimation optimizer with an initial learning rate of 0.001, a weight decay coefficient of 0.0001, a batch size of 32, and a maximum training epoch count of 100. The learning rate is multiplied by 0.5 whenever the overall loss on the validation set fails to decrease for 5 consecutive epochs; training is prematurely stopped when the overall loss on the validation set fails to decrease for 10 consecutive epochs. Model parameters are initialized using He initialization.

[0176] To improve the stability of pipeline visual recognition under different lighting conditions, sheet colors, and occlusion conditions, data enhancement was performed on pipeline images, including brightness and contrast adjustments, slight rotation, random cropping, and local occlusion. Brightness adjustments ranged from 0.8 to 1.2 times the original image brightness, rotation angles ranged from -5 to +5 degrees, and local occlusion areas did not exceed 10% of the image area. Random flipping of the human posture sequence that would alter the clinical meaning of the left and right limbs was not performed.

[0177] 5. Model Validation

[0178] The model was validated on an independent test set. The admission output of training nodes was evaluated using accuracy, macro-average F1 score, recall of the forbidden entry category, and recall of the pause observation category; the output of assist constraint parameters was evaluated using mean absolute error; the pipeline constraint results were evaluated using macro-average F1 score and recall of the forbidden passage category; and the output of safe fallback nodes was evaluated using the first candidate accuracy and the top three candidate accuracy.

[0179] In one implementation, before the model is put into actual control, it should meet the following requirements: the average F1 score of the training node admission macro should be no less than 0.80, the recall rate of the prohibited entry category should be no less than 0.90, the recall rate of the pipeline prohibited passage category should be no less than 0.92, and the accuracy of the top three candidates for the safe fallback node should be no less than 0.90. If the above conditions are not met, the model is only used to output auxiliary suggestions and does not directly control the device to enter the next training node.

[0180] Before the model goes live, a replay verification is performed. Historical training data is input into the model in chronological order, and the path generated by the model is compared with the actual path that was safely completed. When the model suggests entering a node where a serious anomaly has actually occurred, the sample is classified as a high-risk error sample, the annotation is rechecked, and it is added to subsequent training.

[0181] 6. Model Deployment

[0182] The model is deployed in the edge computing controller of the bedside training device. The visual segmentation model can be converted into a half-precision inference format, and the AI ​​nursing assessment model adopts a deployment configuration with a single inference time of no more than 500 milliseconds. Human posture recognition and tubing recognition operate independently, and the training path replanning frequency is 1 time / second to once every 5 seconds; in this embodiment, it is reassessed every 2 seconds under normal conditions, and every 0.5 seconds when an increased risk is detected.

[0183] Model updates are managed using version control. Each model version records the training data time range, number of samples, network parameters, validation results, and applicable device models. Before a new model replaces an old model, it is run in shadow mode at least a preset number of times, recording only model suggestions without controlling the device; it can only be switched to official control mode after review and confirmation.

[0184] Even if the AI ​​nursing assessment model outputs that entry is allowed, if the nursing staff sets a training prohibition, the equipment mechanical limit is not met, the pipeline constraint result is a prohibition, or the key sensor fails, the coupled path planning module will still set the corresponding training node to prohibit entry.

[0185] (iv) Step S4: Coupled planning of stage state diagram and multi-pipeline constraint diagram

[0186] like Figure 5 As shown, the coupled path planning module sequentially performs training node admission screening, training node transformation edge constraint, training node downgrading, pause observation node insertion, safe rollback node configuration, and candidate training path selection.

[0187] First, the admission results output by the AI ​​nursing assessment model for each training node are read. When the admission result is "allowed to enter," the training node is retained; when the admission result is "restricted to enter," the training node is retained, but the allowable range of assistive parameters is narrowed and observation conditions are increased; when the admission result is "suspended observation," a "suspended observation" node is inserted before the training node; when the admission result is "prohibited from entering," the training node is deleted, and subsequent higher-level nodes that must be reached through this training node are also deleted.

[0188] The parameters for limiting access can be adjusted according to preset proportions. For example, the back lift speed is reduced by 20% to 50%, and in this embodiment by 30%; the hip thrust intensity is increased by 5% to 10% of the patient's body weight; the target holding time is shortened by 20%; and the vital signs sampling frequency is doubled.

[0189] Next, the pipeline constraint results output by the multi-pipeline constraint diagram for each training node's transition edge are read. When the result is "allowed to pass," the original transition edge is retained; when the result is "pass at reduced speed," the corresponding body position transition speed is reduced to 40% to 70% of the original planned speed, with 60% used in this embodiment; when the result is "pass with pause observation," a pipeline pause observation node is inserted before the training node's transition edge; when the result is "prohibited to pass," the training node's transition edge is deleted.

[0190] When the admission result for a training node is "allowed," but all transition edges leading to that node are deleted due to pipeline risks, the training node is marked as an unreachable node for this training. After the nursing staff reorganizes the pipelines and updates the multi-pipeline constraint diagram, the coupled path planning module can re-determine whether the node is reachable, but it must not directly enter the node without re-performing the admission assessment.

[0191] Training node downgrading includes intra-phase downgrading and cross-phase downgrading. Intra-phase downgrading can downgrade the standing hold node O4 to the assisted standing node O3, and the assisted standing node O3 to the hip thrust node O2; cross-phase downgrading can downgrade the out-of-bed node to the bedside partial weight-bearing node B4, or the bedside node to the preparation back lift observation node P2.

[0192] The pause observation points include the vital signs pause observation point, the posture pause observation point, and the tubing pause observation point. The vital signs pause observation point is maintained for 30 to 180 seconds, and at least one blood pressure or blood oxygen retest is completed; during the posture pause observation point, the patient's trunk sway, knee flexion, and foot contact status are continuously monitored; during the tubing pause observation point, the bedside training equipment maintains the current position or reduces the movement speed to no more than 2 degrees / second, and the tubing inlet point, fixation point, and suspension status are re-identified.

[0193] Each candidate training path must satisfy nursing contraindications, equipment mechanical limits, maximum allowable nodes, and piping hard constraints. For candidate training paths that satisfy the hard constraints, calculate the comprehensive path cost:

[0194] ;

[0195] In the formula, The comprehensive cost of candidate training path P; The cost of pipeline risks; Tolerating risks for vital signs comes at a cost; The cost of attitude stability; The cost of expected training time; The cost of assisting the device depends on it; all costs are normalized to 0 to 1.

[0196] The weights for tubing risk cost, vital sign tolerance risk cost, posture stability risk cost, expected training time cost, and equipment assistance dependence cost are 0.35, 0.25, 0.20, 0.10, and 0.10, respectively. Since this invention targets bedside training with multiple tubing lines, the weight for tubing risk cost is set to the highest.

[0197] The pipeline risk cost is the weighted sum of the pipeline constraint risk values ​​of each training node transition edge in the candidate path, with high-risk transition edges having a higher weight than low-risk transition edges; the vital sign tolerance risk cost is determined based on the model's predicted changes in heart rate, blood pressure, and blood oxygen; the posture stability risk cost is determined based on the expected trunk sway, knee flexion, and asymmetric foot load at each training node; the expected training time cost is determined based on the ratio of the total path duration to the set training duration; and the equipment assistance dependence cost is determined based on the usage ratio of back support, hip assist, and knee support.

[0198] The coupled path planning module selects the path with the lowest overall cost from the candidate training paths. When the difference in overall cost between two candidate training paths is less than 0.05, the path with higher training benefit is selected first. The training benefit is determined based on the improvement relative to the historical highest completion node, the reduction in active load time, and the reduction in equipment assistance.

[0199] To prevent low-risk objectives from causing the system to remain in the preparation phase for an extended period, a minimum training benefit requirement can be set. If the patient's basic nursing assessment results allow entry into the bedside phase and there are no obstructions preventing passage, the current pathway should include at least one bedside phase training node. If the patient consistently and stably completes the current highest node three times consecutively, the next pathway should include a higher-level node in the candidate pathway, unless real-time vital signs or obstructions are not met.

[0200] After generating the training path, configure the target action, assist parameters, maximum hold time, monitoring items, completion conditions, and safe retreat nodes for each training node. For example, for the assisted standing node O3, configure the hip assist intensity to be 10% to 30% of the patient's body weight, the knee support to remain in place, the footrest to remain fixed, and the maximum standing hold time to be 60 seconds; the safe retreat nodes for postural imbalance are O1 or B4, the safe retreat nodes for tubing abnormalities are nodes that can reduce tubing traction, and the safe retreat nodes for abnormal vital signs are B2, P2, or the training termination node.

[0201] After the training path is generated, the execution control module performs a secondary executability check. This secondary check includes verifying whether the target positions of each actuator exceed mechanical limits, whether adjacent actions conflict, whether the support mechanism enters a locking state before the patient bears weight, and whether the emergency stop device is effective. If the check fails, the corresponding training node is deleted or downgraded.

[0202] (v) Step S5: Training execution, anomaly identification and safe rollback

[0203] like Figure 7 As shown, the execution control module controls the bedside training equipment node by node according to the training path. Before entering each training node, the execution control module verifies whether the previous training node meets the completion conditions, and verifies the vital signs, posture, and pipeline safety conditions corresponding to the current training node.

[0204] The movement speed of the back lift mechanism can be set from 1 degree / second to 10 degrees / second; 3 degrees / second to 6 degrees / second during the preparation phase, 2 degrees / second to 5 degrees / second during the bedside phase, and 1 degree / second to 3 degrees / second during the abduction phase. The hip thrust intensity can be set from 5% to 40% of the patient's body weight; it can be set from 20% to 30% of the patient's body weight when first entering the abduction phase, and can be reduced to 5% to 15% when the patient's training history is stable.

[0205] Abnormal pipe tension is detected based on inlet point displacement, changes in reserved length, and tension sensor data. When the pipe tension reaches 80% of the preset tension threshold, the execution control module reduces the action speed; when the pipe tension reaches the tension threshold or the pipe risk value reaches 0.75, the execution control module stops the current forward action within 200 milliseconds and maintains at least one of the following support states: knee support, footrest, handrail, and back support.

[0206] The system determines the retraction direction that can increase the reserved length of the corresponding pipeline based on the multi-pipeline constraint diagram, and retracts at a speed not exceeding 3 degrees / second until the pipeline tension is reduced to below 70% of the tension threshold, or the pipeline risk value is reduced to below 0.55.

[0207] Abnormal pipe bending is determined based on the minimum bending radius of the pipe. When the minimum bending radius is lower than the minimum allowable bending radius of the corresponding pipe, the actuator causing the increased bending degree is stopped, and the system is preferentially reverted to the training node that can increase the bending radius.

[0208] Abnormal tubing entanglement is detected based on the duration and location of tubing crossings. If the crossing lasts less than 0.5 seconds and is not near the moving parts of the equipment, the system continues monitoring; if the crossing lasts more than 1 second, or if the crossing location enters the movement path of the knee support, hip assist mechanism, or foot pedal mechanism, training is immediately paused and the nursing staff is prompted to tidy up the tubing.

[0209] Attitude stability risk is determined by the following formula:

[0210] ;

[0211] In the formula, The attitude stability risk value; Normalized value for trunk sway risk; Normalized value for the risk of abnormal knee flexion; Normalized values ​​for risk of bipedal contact or asymmetrical weight-bearing; The normalized value for the risk of detachment from the handrail; the value of each normalized value ranges from 0 to 1.

[0212] The weights for trunk sway risk, abnormal knee flexion risk, foot contact or weight-bearing asymmetry risk, and handrail grip loss risk are 0.35, 0.25, 0.20, and 0.20, respectively. Training is allowed to continue when the postural stability risk value is below 0.40; when the risk value is greater than or equal to 0.40 but less than 0.60, the movement speed is reduced and support is increased; when the risk value is greater than or equal to 0.60 but less than 0.75, the posture pause observation point is entered; when the risk value is greater than or equal to 0.75, a safe retreat is performed.

[0213] The risk of trunk sway can be determined based on the displacement of the midpoint of the shoulder or the center of the chest relative to a preset stable area. When the displacement is less than 30 mm, the normalized value for trunk sway risk is 0 to 0.20; when the displacement is 30 to 60 mm, it is 0.20 to 0.60; and when the displacement exceeds 60 mm, it is 0.60 to 1.00. Specific thresholds can be adjusted based on the camera calibration accuracy and the patient's body size.

[0214] Insufficient active weight-bearing is determined based on plantar pressure and the proportion of device assistance. If, after entering a partial weight-bearing or assisted standing node, the actual weight-bearing capacity of both feet does not reach 70% of the target weight, and the device assistance has reached the upper limit allowed for the current training node, it is considered insufficient active weight-bearing. The system will not proceed to the next training node and will preferentially revert to B4 or B3.

[0215] The thresholds for abnormal vital signs are set by authorized medical personnel. In one implementation, the current training node is marked as an abnormal vital sign node when the patient's heart rate increases by more than 20% relative to the initial supine baseline, systolic blood pressure decreases by more than 20 mmHg, or blood oxygen saturation decreases by more than 3 percentage points or falls below 92%. The above thresholds are examples of device control and can be adjusted according to the patient's condition; they do not replace clinical judgment.

[0216] When the abnormality type is tubing traction or tubing bending, prioritize the safe retreat node that can reduce the distance between the tubing inlet point and the fixed point or increase the reserved length of the tubing; when the abnormality type is postural imbalance or insufficient active load, prioritize the previous training node that retains foot pedal, knee support and handrail support; when the abnormality type is abnormal vital signs, prioritize retreating to the bedside sitting-up node, back lift observation node or training termination node.

[0217] Safe retreat should be performed in the following order: maintain support, stop forward movement, add necessary support, retreat slowly, and reassess. For example, if a patient experiences postural imbalance at the assisted standing point, first lock the knee brace and handrails, maintain the foot position, increase hip support, then slowly lower the patient's hip height and retreat to the partial weight-bearing point to avoid a sudden drop that could cause secondary imbalance.

[0218] To prevent the device from repeatedly advancing and retreating due to fluctuations in risk values ​​around the threshold, a hysteresis interval of 0.10 is set. A retreat is triggered when the risk value exceeds the retreat threshold; after a retreat, the device is only allowed to re-enter the pause observation or forward training state if the risk value drops to at least 0.10 below the retreat threshold and remains below it for more than 3 seconds.

[0219] Nurses can terminate training at any time using the emergency stop switch. After the emergency stop is triggered, the execution control module stops all forward movements, and the actuators with self-locking function maintain their current positions; actuators that need to actively return to their original position will only return slowly after confirmation by the nurse.

[0220] (vi) Step S6: Record training results and update the next path

[0221] like Figure 8 As shown, after training, the training update module summarizes the planned training path, the actual completed path, the actual highest completed node, the holding time of each node, the actual assist parameters, deleted nodes, downgraded nodes, paused observation nodes, safe rollback nodes, anomaly types, changes in vital signs, changes in attitude stability risk, and changes in pipeline risk.

[0222] When a node is deleted, downgraded, or rolled back to safety due to a pipeline risk during the current training, the risk coefficient of the corresponding pipeline is increased in the next training path planning. The risk coefficient can be increased by 0.05 to 0.20, and in this embodiment, it is increased by 0.10. If no pipeline anomaly occurs in two consecutive training sessions, the additional risk coefficient is reduced by 0.05 each time until it returns to the initial value.

[0223] When a training session results in a safe reversal due to postural imbalance, the next training session will reduce the postural transition speed at the corresponding node and extend the holding time of the previous stable node. For example, if postural imbalance occurs at node O3, the next training session will increase the holding time at node O2 by 20 seconds and reduce the transition speed from O2 to O3 by 20%.

[0224] When the current training session is terminated due to abnormal vital signs, the highest permissible node for the next training session shall not be higher than the previous training node of the current abnormal node, unless authorized medical personnel reassess and lift the restriction.

[0225] If a patient completes the current highest training node three times consecutively without any abnormalities and the recovery time after training does not exceed the preset time, the next training session can raise the highest permissible node by one level or reduce the hip thrust intensity by 5% of the patient's body weight.

[0226] The training update module generates the starting node, maximum allowed node, node hold time, upper limit of assist parameters, observation window length, key monitoring lines, and safe rollback node for the next training session. The above update parameters are only initial suggestions for the next training session; at the start of the next training session, the system will re-execute steps S1 to S4 to generate the actual training path based on the patient's current vital signs, posture, and line status.

[0227] When incrementally updating the model, the model parameters are not directly changed online using the results of a single training iteration to avoid model instability caused by abnormal samples. The training results are first written to a database awaiting review. After the labels are confirmed by medical staff, the model is retrained monthly, quarterly, or when a preset number of samples is reached. The new model replaces the original model only after it has passed validation and shadow runs.

[0228] 10. Specific Application Examples

[0229] One patient with a left upper limb intravenous infusion tube, a right abdominal drainage tube, a urinary catheter, and an ECG monitoring cable was selected for bedside training to transition from a supine to a standing position. The patient's muscle strength grade was 3, the patient was conscious, the risk of falling was moderate, the pain score was 2, the highest previously achieved milestone was sitting balance milestone B3, and no abnormalities occurred during the most recent training session.

[0230] The visual recognition module identifies that the intravenous infusion tube entry point is located on the patient's left upper limb, and the fixation point is located at the bedside infusion stand; the abdominal drainage tube entry point is located on the patient's right abdomen, and the fixation point is located on the right side of the bed; the urinary catheter extends from the patient's lower abdominal area to the drainage bag under the bed; and the electrocardiogram monitoring cable extends from the patient's chest to the bedside monitoring equipment.

[0231] The multi-tube constraint diagram generation module calculated that the maximum tube risk value when transitioning from B3 to B4 is 0.41, corresponding to a reduced speed for passage; when transitioning from B4 to O1, the abdominal drainage tube risk value is 0.62, corresponding to a pause for observation before passage; when transitioning from O1 to O2, the insufficient length of the urinary catheter results in a tube risk value of 0.78, corresponding to prohibit passage.

[0232] The AI ​​nursing assessment model allows entry for outputs from nodes P0 to B3, restricts entry for output B4, pauses observation for output O1, and prohibits entry for outputs from O2 to O4. The coupled path planning module generates a training path that sequentially passes through P0, P1, P2, P3, P4, B1, B2, B3, B4, the pipeline pause observation node, and O1, and sets O1 as the highest allowed node for this operation.

[0233] After the patient completed B4, the bedside training equipment remained in a partially weight-bearing state, and the system re-identified the suspension status of the abdominal drainage tube and urinary catheter. The nursing staff adjusted the catheter fixation point according to the prompts, increasing the catheter's length by 120 mm. After updating the multi-channel constraint diagram, the channel risk value of the transition edge from O1 to O2 decreased from 0.78 to 0.53. However, since the AI ​​nursing assessment model still showed restricted access for O2, and the initial highest allowed node for this training was O1, the highest allowed node was not temporarily increased during this training; O1 remained the highest training node.

[0234] The patient maintained position O1 for 20 seconds. Heart rate increased by 8% compared to the initial supine position, systolic blood pressure decreased by 6 mmHg, and oxygen saturation remained at 96%. The postural stability risk value was 0.38, and the tubing risk value was 0.46, all of which did not reach the pause or regression thresholds. After completing O1, the bedside training device regressed to B2 and P2 according to the preset path.

[0235] The training update module records O1 as the highest completion node for this training session, writes the catheter risk value, actual retention time, vital sign changes, and actual assistance parameters before and after catheter fixation point adjustment into the historical training record, and sets O2 as the candidate highest node for the next training session. At the start of the next training session, the system re-collects the patient's status and multi-catheter status, and generates the actual training path again.

[0236] Therefore, this invention uses a phased state diagram to represent the progressive training process from supine to standing position, a multi-pipeline constraint diagram to represent the restrictions imposed by nursing pipelines on different positional transitions, and an AI nursing assessment model and coupled path planning module to retain, delete, downgrade, or insert pause observation nodes for training nodes. During training, appropriate safety rollback nodes are selected based on the type of anomaly. After training, the next training parameters are updated based on the actual training results, thus forming a closed-loop control system of pre-training planning, in-training monitoring, and post-training updates.

[0237] IV. Examples, Comparative Examples, and Experimental Verification

[0238] (I) Experimental Objectives and General Description

[0239] To verify the feasibility and technical effectiveness of the bedside training path dynamic planning method and system based on the coupling of multi-pipeline constraint diagram and stage state diagram, a bedside training experimental platform was built to verify the training path generation, multi-pipeline risk prediction, training node adjustment, abnormal safety rollback, and subsequent training path update process.

[0240] This experiment focuses on verifying the following technical effects: First, verifying whether the multi-line constraint diagram can identify training node transition edges that may cause tubing traction, bending, entanglement, or insufficient margin before the device performs a positional change action; Second, verifying whether coupling the stage state diagram and the multi-line constraint diagram can reduce nursing tubing abnormalities by deleting, downgrading, or inserting pause observation nodes; Third, verifying whether the control method of different safety retreat nodes corresponding to different abnormality types can shorten the time for patients to recover to a safe and stable state; Fourth, verifying that the AI ​​nursing assessment model, after integrating nursing assessment, vital signs, visual posture, stage state diagram, and multi-line constraint diagram, has an acceptable ability to identify training node admission results and safety retreat nodes; Fifth, verifying whether the mechanism of updating the next training path based on the current training results can reduce the recurrence rate of the same abnormality.

[0241] The experimental data listed were generated based on the reproducible bench test conditions described below, using simulated data matched to the range, control cycle, noise level, and range of human movement of the experimental apparatus. The experimental data are not intended as clinical diagnostic conclusions, but only to illustrate the technical effects achievable by the control method described in this invention under engineering conditions.

[0242] (II) Experimental Apparatus and Common Experimental Conditions

[0243] like Figure 9 As shown, the experimental setup includes a bedside training equipment prototype, a human training model, a first camera, a second camera, a nursing tubing simulation component, a vital signs simulator, a plantar pressure detection component, a tubing tension detection component, an edge computing controller, and an equipment execution controller.

[0244] The prototype bedside training equipment includes a back lift mechanism, a hip assist mechanism, a knee support limiting mechanism, a foot support mechanism, and a handrail locking mechanism. The back lift mechanism has an angle adjustment range of 0 to 70 degrees, a maximum angular velocity of 10 degrees / second, and an angle measurement error of no more than 0.5 degrees. The hip assist mechanism has a thrust adjustment range of 0 Newtons to 400 Newtons, with a measurement error of no more than 1% of full scale. The knee support limiting mechanism provides limiting support in front of the patient's knee. The foot support mechanism is equipped with two sets of pressure sensors, one on each side, with a measurement range of 0 to 800 Newtons and a measurement error of no more than 1.5% of full scale.

[0245] The human training model is equipped with three equivalent weights: 55 kg, 70 kg, and 85 kg. Different body shapes can be formed by adjusting the weights on the trunk, hips, and lower limbs. The human training model has adjustable joints in the shoulders, hips, knees, and ankles, which can simulate supine elevation, sitting up at the edge of the bed, partial weight-bearing, and assisted standing postures.

[0246] The first camera is positioned above and in front of the bedside training equipment to capture key points and overall posture of the human body; the second camera is positioned to the side of the equipment to capture the inlet point, fixing point, and suspension status of the tubing. The camera resolution is 1920×1080, the human posture image acquisition frequency is 10 frames / second, and the tubing image acquisition frequency is 5 frames / second.

[0247] The nursing tubing simulation assembly includes standard IV tubing, drainage tubing, urinary catheters, oxygen tubing, and monitoring cables. One end of each tubing is fixed to the corresponding body surface location on the human training model, and the other end is fixed to an IV stand, bed, drainage bag support, or a simulated monitoring device. Tension sensors with a range of 0 to 10 Newtons and a resolution of 0.01 Newtons are installed at the tubing fixation points. A movable marker is placed at the tubing inlet point to measure the displacement of the tubing relative to its body surface inlet position.

[0248] The vital signs simulator can output data on heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen saturation, and respiratory rate, and can generate postural tolerance abnormalities according to preset time. Under normal conditions, the heart rate is set to 65 to 85 beats per minute, the systolic blood pressure is set to 105 to 135 mmHg, and the blood oxygen saturation is set to 95% to 99%. Under abnormal conditions, the simulator can be set to increase the heart rate by more than 20% relative to the baseline value, decrease the systolic blood pressure by more than 20 mmHg, or decrease the blood oxygen saturation to below 92%.

[0249] The edge computing controller runs programs for human posture recognition, pipeline status recognition, stage state diagram generation, multi-pipeline constraint diagram generation, AI-based nursing assessment models, and coupled path planning. The equipment execution controller controls various mechanical actuators and reads limit, pressure, and tension data at 20-millisecond intervals.

[0250] Four tubing combinations were used in the experiment: the first combination included a standard infusion tube and a monitoring cable; the second combination included a drainage tube and a urinary catheter; the third combination included a central venous catheter simulation tube, an oxygen line, and a monitoring cable; and the fourth combination included a standard infusion tube, a drainage tube, a urinary catheter, an oxygen line, and a monitoring cable. Five initial allowable lengths of 60 mm, 90 mm, 120 mm, 150 mm, and 180 mm were provided for each combination.

[0251] Sixty experimental scenarios were created using three equivalent human body masses, four tubing combinations, and five initial reserved length combinations. Each scenario was repeated 10 times, resulting in a total of 600 complete training path execution experiments. All control methods used the same scenario sequence and the same vital sign simulation sequence.

[0252] (III) Example 1: Training path pre-planning based on multi-pipeline constraint graph

[0253] Before training begins, the visual recognition module identifies the posture of the human training model, the entry points of each surface tube, the tube fixing points, and the tube overhangs; the stage state diagram generation module generates training nodes including P0 to P4, B1 to B4, and O1 to O4; the multi-tube constraint diagram generation module calculates the tube risk value of the transition edges of each training node; the AI ​​nursing assessment model outputs the admission results of each training node; and the coupled path planning module performs actions such as retaining, deleting, downgrading, or inserting pause observation nodes for the training nodes.

[0254] In this embodiment, the training node transitions in 600 trials are divided into 240 high-risk transitions and 360 low-risk transitions based on the results of manual review before the experiment. High-risk transitions refer to transitions where, without adjusting the motion parameters and fixing the pipeline, the pipeline tension is expected to reach more than 80% of the corresponding tension threshold, the minimum bending radius of the pipeline is lower than the allowable value, or the pipeline enters the equipment's motion path; the remaining transitions are classified as low-risk transitions.

[0255] In this embodiment, out of 240 high-risk conversions, 226 nodes were deleted, downgraded, or inserted for pause observation in advance, resulting in a pre-intervention rate of 94.17% for high-risk conversions. In 360 low-risk conversions, 21 nodes were deleted or downgraded due to being judged as high-risk by the system, resulting in a false restriction rate of 5.83% for low-risk conversions.

[0256] In 600 training path tests, this embodiment completed the preset maximum allowable node or the system-generated alternative maximum node 548 times, with an effective path completion rate of 91.33%; the number of times the pipeline inlet point displacement exceeded 20 mm was 4 times, accounting for 0.67%; the average maximum pipeline tension recorded during the training process was 2.63 Newtons, with a standard deviation of 0.61 Newtons.

[0257] In this embodiment, the system does not directly terminate all high-risk paths, but rather adopts different processing methods based on the source of the risk. In 226 high-risk transitions that were intervened in advance, training nodes were deleted 42 times, training nodes were downgraded 67 times, tube insertion for observation was performed 89 times, and the action speed was reduced 28 times. Of the 89 tube insertions for observation, after the nursing tubes were reorganized, 71 times the system continued to enter the original target node, and 18 times it was downgraded to the previous training node.

[0258] The results show that multi-pipeline constraint graphs can identify most high-risk training node transitions before training actions are executed, and maintain the continuity of the training path through node degradation and pause observation, rather than simply using a uniform termination method.

[0259] (iv) Comparative Example 1: Using a fixed sit-stand training path

[0260] This comparative example uses a pre-set fixed training path P0, P1, P2, P3, P4, B1, B2, B3, B4, O1, O2, O3, and O4. The equipment only performs corresponding actions based on mechanical limits, without generating multi-pipeline constraint diagrams or modifying training nodes according to pipeline conditions. During training, the operation stops when the pipeline tension reaches the corresponding tension threshold.

[0261] In 240 high-risk transitions, only 18 were not performed due to mechanical limitations or contraindications to basic care, resulting in a pre-intervention rate of 7.50% for high-risk transitions. In 600 training path trials, the preset highest node was completed 432 times, resulting in an effective completion rate of 72.00%. There were 68 instances where the displacement of the pipeline inlet point exceeded 20 mm, accounting for 11.33%. The average maximum pipeline tension was 6.84 Newtons, with a standard deviation of 1.27 Newtons.

[0262] In this comparative example, pipeline risks are primarily identified after the equipment has already performed its operation and the pipeline tension has increased, which easily leads to training interruptions. The fixed training path does not provide alternative nodes based on the type of risk, causing multiple trials to terminate abruptly.

[0263] (v) Comparative Example 2: Nursing assessment and posture recognition were used, but a multi-line constraint diagram was not established.

[0264] This comparative study collected basic nursing assessment data, historical training records, vital signs, and body postures from patients. The AI ​​model generated training paths, but did not identify tubing inlet points, anchor points, or suspension states, nor did it create multi-tubing constraint diagrams. Tubing alarms during training were only triggered by tension sensors.

[0265] In 240 high-risk transitions, the baseline was deleted or downgraded 52 times in advance, with a pre-intervention rate of 21.67% for high-risk transitions; in 600 training path trials, the path was effectively completed 456 times, with an effective completion rate of 76.00%; the number of times the pipeline inlet point displacement exceeded 20 mm was 49 times, accounting for 8.17%; the average maximum pipeline tension was 6.12 Newtons, with a standard deviation of 1.11 Newtons.

[0266] This comparative model can exclude some unsuitable training nodes based on the patient's ability and postural stability, but due to the lack of tubing inlet points, fixing points and reserved length constraints, it cannot identify the risks caused by insufficient tubing margin before training.

[0267] (vi) Comparative Example 3: Using pipeline tension threshold alarm but without graph coupling planning

[0268] This comparative model identifies nursing tubing and collects tubing tension, but does not couple the multi-tubing status with the stage status diagram. The training path still executes according to the output results of the AI ​​nursing assessment model; when the tubing tension reaches 80% of the tension threshold, the speed is reduced, and when the tension threshold is reached, the current action stops.

[0269] In 240 high-risk transitions, the baseline ratio was reduced or stopped in advance 133 times, with an intervention rate of 55.42% for high-risk transitions; in 600 training path trials, the path was effectively completed 503 times, with an effective completion rate of 83.83%; the number of times the pipe inlet point displacement exceeded 20 mm was 22 times, accounting for 3.67%; the average maximum pipe tension was 4.31 Newtons, with a standard deviation of 0.93 Newtons.

[0270] This comparative example can reduce some of the pipeline tension risk, but its risk control occurs after the training action begins. It cannot delete training nodes, reduce training levels, or insert pause observation nodes in advance based on the expected risks. Therefore, the maximum pipeline tension and abnormal displacement at the inlet point are still higher than in Example 1.

[0271] The main experimental results of Example 1 and Comparative Examples 1 to 3 are shown in Table 1.

[0272] Table 1. Experimental results of different training path control methods

[0273]

[0274] As shown in Table 1, compared to a fixed training path, this invention reduces the abnormal displacement rate at the pipeline inlet point from 11.33% to 0.67%, and the average maximum pipeline tension from 6.84 Newtons to 2.63 Newtons. Compared to Comparative Example 3, which only uses pipeline tension threshold alarms, this invention can further reduce the maximum pipeline tension and improve the effective path completion rate, indicating that coupling the multi-pipeline constraint diagram with the stage state diagram before training has a technical effect that differs from post-event alarms.

[0275] like Figure 10 As shown, in a typical experiment of transitioning from node O1 to node O2, the pipeline tension corresponding to the fixed training path continuously increases after the action begins, reaching 6.7 Newtons at approximately 4.2 seconds; the pipeline threshold alarm method reduces the action speed at approximately 3.3 seconds, but the peak tension still reaches 4.5 Newtons; the present invention identifies a pipeline risk value of 0.76 for the transition edge during the path generation stage, sets node O2 to prohibit entry, and inserts a pipeline pause observation node before node O1, thus keeping the pipeline tension below 2.1 Newtons.

[0276] (vii) Example 2: Specific path generation of coupling multi-pipeline constraint diagram and stage state diagram

[0277] A 70 kg human training model was used, and connected to standard IV tubing, abdominal drainage tubing, urinary catheter, oxygen tubing, and monitoring cables. The initial reserved lengths for the standard IV tubing, abdominal drainage tubing, urinary catheter, oxygen tubing, and monitoring cables were 150 mm, 120 mm, 90 mm, 180 mm, and 160 mm, respectively.

[0278] The basic nursing assessment parameters were set as follows: muscle strength grade 3, level of consciousness: alert, fall risk: medium risk, pain score: 2, previous highest completion level: B3, and no absolute training contraindications were set. Initial vital signs were: heart rate 76 bpm, systolic blood pressure 122 mmHg, diastolic blood pressure 76 mmHg, and oxygen saturation 97%.

[0279] The stage state diagram generation module generates training nodes P0 to P4, B1 to B4, and O1 to O4. The multi-pipeline constraint diagram generation module calculates the transition edges for each training node.

[0280] In the transition zone from B3 to B4, the normalized value of the traction risk of the abdominal drainage tube is 0.41, the normalized value of the bending risk is 0.25, the normalized value of the cross-entanglement risk is 0.10, and the normalized value of the insufficient reserved length risk is 0.35. Combined with the drainage tube risk coefficient of 1.25, the tube risk value is 0.41, which corresponds to reducing the speed of passage.

[0281] In the transition zone from B4 to O1, the normalized value of the traction risk of the abdominal drainage tube is 0.58, the normalized value of the bending risk is 0.42, the normalized value of the cross-entanglement risk is 0.25, and the normalized value of the insufficient reserved length risk is 0.52. The calculated tube risk value is 0.58, corresponding to the suspension of observation.

[0282] In the transition zone from O1 to O2, the normalized value of the catheter's traction risk is 0.82, the normalized value of its bending risk is 0.72, the normalized value of its cross-entanglement risk is 0.55, and the normalized value of its insufficient reserved length risk is 0.90. The calculated tube risk value is 0.75, which corresponds to prohibiting passage.

[0283] The AI ​​nursing assessment model allows access to outputs P0 through B3, restricts access to output B4, pauses observation of output O1, and prohibits access to outputs O2 through O4. The coupled path planning module ultimately generates a training path consisting of P0, P1, P2, P3, P4, B1, B2, B3, B4, the pipeline pause observation node, and O1.

[0284] The transition speed from B3 to B4 was reduced from the planned 4 degrees / second to 2.4 degrees / second. After the patient reached B4, they remained in this position for 30 seconds, and the status of the abdominal drainage tube and urinary catheter was rechecked. Following the system prompts, the nursing staff moved the catheter fixation point 80 mm towards the patient and released 40 mm of tubing, increasing the total remaining catheter length by 120 mm.

[0285] After the update, the normalized value of the catheter traction risk at the O1 to O2 transition edge decreased to 0.48, the normalized value of the bending risk decreased to 0.40, the normalized value of the cross-entanglement risk decreased to 0.25, the normalized value of the insufficient reserved length risk decreased to 0.55, and the tubing risk value decreased to 0.42, corresponding to a reduced speed.

[0286] Since the initial maximum allowable node for this training was O1, and the patient's historical record showed no history of completing hip assisted push-ups during the out-of-bed period, the system did not temporarily increase the maximum allowable node during the same training session, maintaining O1 as the maximum node. The patient held O1 for 20 seconds; heart rate increased to 82 beats / minute, systolic blood pressure decreased to 116 mmHg, oxygen saturation was 96%, postural stability risk was 0.38, and tubing risk was 0.46, all of which did not reach the pause or rollback thresholds. Upon completion, the system rolled back in the order of O1, B4, B2, and P2.

[0287] This embodiment illustrates that the present invention can update the multi-pipeline constraint diagram in real time based on the pipeline adjustment results. However, whether the training node is upgraded is still limited by the patient's historical training and AI nursing assessment results, thus avoiding ignoring the patient's ability level simply because the pipeline risk has been reduced.

[0288] (viii) Example 3: Safety rollback test based on anomaly type

[0289] To verify the technical effectiveness of different safety rollback nodes corresponding to different anomaly types, three types of anomalies were set up: pipeline anomaly, posture anomaly, and vital sign anomaly. Each type was set up 60 times, resulting in a total of 180 anomaly tests.

[0290] Abnormalities in the tubing include tubing tension reaching a threshold, the minimum bending radius of the tubing being lower than the allowable value, and tubing entering the movement path of the mechanism; abnormal postures include trunk sway risk values ​​exceeding 0.75, abnormal knee flexion, and severe asymmetry in weight-bearing on both feet; abnormal vital signs include heart rate increasing by more than 20% relative to the baseline value, systolic blood pressure decreasing by more than 20 mmHg, and blood oxygen saturation being lower than 92%.

[0291] Example 3 adopts the classification and retreat method of the present invention: for pipeline abnormalities, retreat to the node that can increase the reserved length of the pipeline first; for posture abnormalities, maintain foot pedal, knee support and handrail support first, and retreat to the previous low load node; for vital signs abnormalities, retreat to the bedside sitting up node or back lifting observation node first.

[0292] Comparative Example 4: Uniform Stopping Method

[0293] This comparative example immediately stops all forward movements and maintains the current position upon detecting any abnormality, without selecting a safe rollback node based on the type of abnormality. The subsequent handling method is then manually selected by the nursing staff.

[0294] Comparative Example 5: All returned to the initial supine position.

[0295] In this comparative example, after detecting any abnormality, it reverts to the initial supine node P0 along a fixed path, without distinguishing between pipeline abnormalities, posture abnormalities, and vital sign abnormalities.

[0296] The test results are shown in Table 2.

[0297] Table 2. Experimental results of different abnormality handling methods

[0298]

[0299] While the uniform stopping method results in a lower maximum trunk angular velocity, the patient or human training model remains at the location of the abnormality, and the tubing tension or postural instability cannot be immediately relieved, thus requiring a longer time to return to a safe and stable state. Uniform retraction to P0 allows the patient to eventually return to a supine position, but in some postural or tubing abnormality scenarios, excessive retraction can lead to increased maximum trunk angular velocity and a higher incidence of secondary risks.

[0300] This invention employs a classification-based safety rollback method, which reduces the average recovery time to 7.6 seconds, lowers the secondary risk rate to 2.78%, and allows 114 trials to continue to complete the next lower training node after re-observation. Figure 11The average recovery time, secondary risk incidence, and maximum trunk angular velocity are compared among the three abnormal handling methods.

[0301] (ix) Example 4: Validation Trial of AI Nursing Assessment Model

[0302] 1. Dataset Composition

[0303] Based on data collected from the experimental bench, and combined with different human equivalent mass, nursing assessment level, changes in vital signs, changes in posture, and multi-circuit status, 12,480 training node-level simulation samples were formed. These samples correspond to 320 independent patient parameter files, each with independent muscle strength level, fall risk, postural tolerance, and historical training sequences.

[0304] The dataset was divided into training, validation, and test sets based on patient parameter profiles, with proportions of 70%, 15%, and 15%, respectively. All samples from the same patient parameter profile were assigned to only one subset of the dataset. The training set consisted of 8736 samples, the validation set of 1872 samples, and the test set of 1872 samples.

[0305] In the training node admission labels, 47.4% of the samples were allowed to enter, 24.6% were restricted to enter, 16.2% were suspended for observation, and 11.8% were prohibited from entering. During training, classification loss weights of 1.00, 1.30, 1.45, and 1.60 were set for the four types of samples, respectively.

[0306] The model employs an adaptive moment estimator optimizer with an initial learning rate of 0.001, a weight decay coefficient of 0.0001, a batch size of 32, and a maximum training epoch of 100. If the validation set loss does not decrease for 5 consecutive epochs, the learning rate is reduced to 0.5 times its original value; training stops when the loss does not decrease for 10 consecutive epochs.

[0307] 2. Comparison Model

[0308] The complete model employs nursing assessment feature coding branch, historical training coding branch, vital sign time-series coding branch, visual posture coding branch, stage state diagram coding branch, and multi-channel constraint diagram coding branch.

[0309] Set the following ablation ratio:

[0310] Comparative Example 6 does not use multi-pipe constraint diagram coding for branches; it only inputs the current tension value of each pipe.

[0311] Comparative Example 7 does not use a staged state graph encoding branch, but uses the target training node as the input of ordinary category features;

[0312] Comparative Example 8 uses a pre-defined rule model and does not employ neural network fusion.

[0313] The test results are shown in Table 3.

[0314] Table 3. Test results of AI nursing assessment model and ablation model

[0315]

[0316] The complete model has a recall of 0.944 for the "no entry" training node, a recall of 0.956 for the "no passage" pipeline state, and a top-three candidate accuracy of 0.932 for the safe backoff node. After removing the multi-pipeline constraint graph encoding branch, the pipeline "no passage" recall drops to 0.808; after removing the stage state graph encoding branch, the first candidate accuracy for the safe backoff node drops to 0.751, indicating that the stage state graph plays a role in identifying backoff nodes that match the current training stage.

[0317] The complete model has a mean absolute error of 3.2 degrees for the upper limit of the back lift angle, a mean absolute error of 3.8% of the patient's body weight for the hip thrust intensity, and a mean absolute error of 7.1 seconds for the duration of the training node.

[0318] After the complete model is deployed on the edge computing controller, the average inference time for a single nursing assessment and path node review is 112 milliseconds, and the maximum inference time is 186 milliseconds, which is lower than the 500 millisecond control cycle set in this implementation. Figure 12 The data shows the macro-average F1 score for admission of training nodes, the recall rate of prohibited categories, and the accuracy of the top three candidates for safe fallback nodes for the complete model and different ablation models.

[0319] (X) Example 5: Verification of Training Path Update Mechanism

[0320] To verify the technical effectiveness of the training update module, 30 patient parameter files were set up, with each group undergoing 5 consecutive training sessions. The initial maximum allowable node was the same for each group of patients, which was B3; in the first training session, each group was given a setting for either tubing risk or posture stability risk.

[0321] Example 5 employs the training update mechanism of this invention, updating the next training path based on the actual completed nodes, downgraded nodes, safe rollback nodes, and reasons for anomalies. When a pipeline risk occurs, the additional risk coefficient for the corresponding pipeline increases by 0.10; if the same anomaly does not occur twice consecutively, it decreases by 0.05 each time. When attitude imbalance occurs, the corresponding conversion speed will be reduced by 20% in the next training iteration, and the holding time of the previous stable node will be increased by 20 seconds.

[0322] Comparative Example 9: No training update mechanism adopted

[0323] This comparative model generates a path based solely on the current nursing assessment data at the start of each training session, without using deleted nodes, downgraded nodes, safe rollback nodes, or abnormal reasons from the previous training.

[0324] The results of five consecutive training sessions are shown in Table 4.

[0325] Table 4 Results of the continuous training and update experiment

[0326]

[0327] The average highest completion node level in Table 4 is statistically analyzed by assigning consecutive levels B3, B4, O1, O2, and O3 in sequence. In Example 5, the recurrence rate of the same anomaly was 8.67%, lower than 22.00% when the training update mechanism was not used; the effective completion rate of the fifth training path reached 93.33%, and the average reduction in equipment assistance was 18.40%.

[0328] Figure 13 The changes in path completion rate and device assistance ratio for the two methods are shown during five consecutive training sessions. When using the training update mechanism, the system can increase the restriction level for training nodes that have experienced anomalies in the early stages of training, and gradually remove the restriction after the subsequent training stabilizes, thereby avoiding either maintaining excessively high risk weights or completely ignoring historical anomalies.

[0329] (xi) Comprehensive analysis of experimental results

[0330] As shown in Example 1 and Comparative Examples 1 to 3, using the multi-pipeline state only as an alarm input during training can reduce pipeline abnormalities to some extent, but it cannot fully prevent the increase in pipeline tension after the action begins. This invention establishes a multi-pipeline constraint diagram before training and maps it to the training node transformation edges in the stage state diagram, allowing pipeline risks to directly change the training path structure. Therefore, it can reduce the maximum pipeline tension and abnormal displacement at the pipeline inlet point.

[0331] As can be seen from Example 2, the present invention can calculate the risk of different pipelines separately and use the pipeline with the highest risk on the same training node as the constraint basis. After the nursing staff adjusts the pipeline, the system can update the pipeline constraint results, but whether to improve the training level still needs to be re-judged in combination with nursing assessment, historical training and vital signs to avoid the training path being controlled by the risk of a single pipeline.

[0332] As can be seen from Example 3 and Comparative Examples 4 to 5, selecting different safe reversal points for different abnormality types can resolve the abnormal state in a shorter time and reduce the secondary risks caused by large-scale positional reversal compared to uniformly stopping or uniformly reversing to the initial supine position.

[0333] As can be seen from Example 4 and Comparative Examples 6 to 8, the multi-pipeline constraint diagram coding branch mainly improves the ability to identify the pipeline prohibition state, and the stage state diagram coding branch mainly improves the accuracy of the safety backoff node selection. When the two are integrated with nursing assessment, vital signs and visual posture features, a relatively complete training node admission judgment can be formed.

[0334] As can be seen from Example 5 and Comparative Example 9, using the actual completed node, deleted node, downgraded node, safe rollback node, and abnormal reason to update the next training path can reduce the recurrence of the same abnormality and enable the device to gradually adjust according to the patient's training performance.

[0335] Therefore, the above experimental results show that the present invention can realize multi-circuit risk constraint planning before training, abnormal classification and rollback during training, and path parameter update after training. It is suitable for progressive bedside training of patients with multiple nursing tubes from lying position to bedside sitting position, partial weight-bearing and assisted standing position.

[0336] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0337] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0338] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0339] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0340] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0341] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0342] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A dynamic planning method for patient training path from supine to standing position based on an AI nursing assessment model, characterized in that, The process includes the following steps: S1. Collecting basic nursing assessment data, historical training records, real-time vital signs, and camera images of the patient, and using the camera images to identify the patient's posture, the entry point of the external tube, the fixation point of the tube, and the tube suspension status; S2. Generating a phased state diagram including the preparation phase, bedside phase, and bed exit phase based on the patient's posture, and generating a multi-tube constraint diagram based on the entry point of the external tube, the fixation point of the tube, and the tube suspension status; S3. Inputting the basic nursing assessment data, historical training records, real-time vital signs, phased state diagram, and multi-tube constraint diagram into the AI ​​nursing assessment model to obtain the admission results, assistance limitation parameters, and tube constraints for each training node. Results; S4. Based on the admission results and pipeline constraint results, retain, delete, downgrade, or insert pause observation nodes for the training nodes in the stage state diagram to generate the current training path from supine to standing position; S5. During training execution, if the displacement of the pipeline inlet point, pipeline overhang margin, posture stability, or vital signs are detected to exceed the safety conditions of the corresponding training node, the bedside training equipment is controlled to switch to the safe retreat node corresponding to the abnormality type; S6. Based on the actual completed nodes, deleted nodes, downgraded nodes, pause observation nodes, retreat nodes, and their abnormality reasons, update the node order, maximum allowed nodes, and assist limit parameters of the next training path.

2. The method according to claim 1, characterized in that, In step S1, the basic nursing assessment data includes muscle strength level, level of consciousness, fall risk level, postural tolerance, pain score, tube care risk level, and training contraindication information entered by the nursing staff; the historical training record includes the highest training milestone completed previously, the duration of the milestone, equipment assistance parameters, reasons for abnormal termination, and post-training recovery time. And / or, in step S1, the camera images include images of the patient's torso, hips, knees, ankles, feet, gripping areas, and exposed sections of tubing; the patient's posture identified by the camera images includes a supine elevated posture, a bedside sitting posture, a sitting balanced posture, a partially weight-bearing posture, and an assisted standing posture. And / or, in step S1, the entry point of the body surface tubing is the connection position of the infusion tube, drainage tube, urinary catheter, monitoring cable, or oxygen tubing relative to the patient's body surface or nursing fixation point; the tubing fixation point is the fixation position of the tubing relative to the bedside training equipment, bed, patient clothing, or nursing fixation device; the tubing suspension state includes the tubing reserved length, bent state, crossed and wrapped state, and free swing state.

3. The method according to claim 1, characterized in that, In step S2, the stage state diagram includes training nodes, node sequence, node entry conditions, node maintenance conditions, and node regression conditions; the preparation period training nodes include supine elevation, lower limb positioning, and foot pre-contact nodes; the bedside period training nodes include bedside sitting up, sitting balance, and partial weight-bearing nodes; and the off-bed period training nodes include knee limitation, hip assistance, assisted standing, and standing maintenance nodes. And / or, in step S2, the multi-pipeline constraint diagram includes a pipe inlet point, a pipe fixing point, a pipe overhang section, and a pipe risk edge. The pipe risk edge is used to represent the risks of traction, bending, entanglement, or insufficient margin that may occur in the corresponding pipe when the patient transitions from one training node to another.

4. The method according to claim 1, characterized in that, In step S3, the AI ​​nursing assessment model integrates basic nursing assessment data, historical training records, real-time vital signs, stage status diagrams, and multi-channel constraint diagrams, and outputs the admission results for each training node in the preparation period, bedside period, and bed exit period, which are allowed to enter, restricted to enter, suspended for observation, or prohibited from entering. And / or, in step S3, the assist limiting parameters include back lifting speed, upper limit of back lifting angle, hip assist intensity, knee support limit status, foot support height, handrail locking status, and upper limit of standing holding time; the pipeline constraint results include at least one of allowing passage, reducing speed to pass, pausing to observe passage, and prohibiting passage.

5. The method according to claim 1, characterized in that, In step S4, when the pipeline constraint result corresponding to a certain training node is "no passage", the training node is deleted or downgraded to the previous stage training node; when the pipeline constraint result corresponding to a certain training node is "pause observation pass", a pause observation node is inserted before the training node, and the pipeline suspension status and patient posture are re-identified in the pause observation node. And / or, in step S5, the abnormality types include abnormal tubing traction, abnormal tubing bending, abnormal tubing entanglement, abnormal posture imbalance, abnormal insufficient active weight-bearing, and abnormal vital signs; when the abnormality type is abnormal tubing traction or abnormal tubing bending, priority is given to switching to a safe retreat node that reduces the amplitude of body position changes; when the abnormality type is abnormal posture imbalance or abnormal insufficient active weight-bearing, priority is given to switching to the previous training node or the bedside training node; when the abnormality type is abnormal vital signs, priority is given to switching to the preparation training node or the training termination node.

6. The method according to claim 1, characterized in that, In step S6, if a node is deleted, downgraded, or rolled back due to pipeline constraints during the training, the risk weight of the corresponding pipeline will be increased in the next training path, and the maximum allowable level of the training node related to that pipeline will be reduced. If the training continuously completes a preset number of training nodes without any abnormalities, the maximum allowable node in the next training path will be increased or the retention time of the corresponding node will be extended. And / or, the method is used to control the mechanical execution actions and safety protection actions of a bedside training device, the bedside training device including at least one of a back lifting mechanism, a hip assist mechanism, a knee support limiting mechanism, a foot pedal support mechanism, a handrail locking mechanism, and a tubing follow-up management mechanism.

7. A system for implementing the method according to any one of claims 1-6, characterized in that, It includes a data acquisition module, a visual recognition module, a stage state diagram generation module, a multi-pipeline constraint diagram generation module, an AI nursing assessment module, a coupled path planning module, an execution control module, and a training and update module; The data acquisition module is used to collect patients' basic nursing care assessment data, historical training records, and real-time vital signs. The visual recognition module is used to identify the patient's posture, the entry point of the tube on the body surface, the fixation point of the tube, and the suspension status of the tube based on the camera image. The stage state diagram generation module is used to generate a stage state diagram that includes training nodes for the preparation period, bedside period, and bed exit period. The multi-pipe constraint diagram generation module is used to generate a multi-pipe constraint diagram based on the pipe inlet point, pipe fixing point, and pipe suspension state on the body surface. The AI ​​nursing assessment module is used to output the admission results, assistance limitation parameters and pipeline constraint results of each training node based on basic nursing assessment data, historical training records, real-time vital signs, stage status diagram and multi-circuit constraint diagram. The coupled path planning module is used to retain, delete, downgrade, or insert pause observation nodes for training nodes in the stage state diagram according to the admission results and pipeline constraint results, and generate the current training path from supine to standing position. The execution control module is used to control the bedside training device to execute the training path, and to switch to the safe rollback node corresponding to the type of abnormality in case of an abnormality. The training update module is used to update the node order, maximum allowed nodes, and assist limit parameters of the next training path based on the actual completed nodes, deleted nodes, downgraded nodes, paused observation nodes, rollback nodes, and their abnormal reasons.

8. The system according to claim 7, characterized in that, The visual recognition module includes a human key point recognition unit, a body posture discrimination unit, a pipeline inlet point recognition unit, and a pipeline suspension state recognition unit; the human key point recognition unit is used to recognize key points of the head and neck, shoulders, hips, knees, ankles, and feet, and the pipeline suspension state recognition unit is used to recognize the pipeline reserved length, bending state, cross-entanglement state, and free swinging state; And / or, the multi-pipe constraint diagram generation module is configured to take the pipe inlet point, pipe fixing point and pipe overhang as constraint objects, and generate the corresponding pipe traction risk, bending risk, entanglement risk and insufficient margin risk according to the patient's body position change when switching from one training node to another training node; And / or, the coupled path planning module includes a node filtering unit, a node degradation unit, a pause observation node insertion unit, and a safety rollback node configuration unit; the node filtering unit is used to delete training nodes that do not meet the admission results or pipeline constraint results; the node degradation unit is used to downgrade the off-bed training node to the bedside training node or the preparation training node; the pause observation node insertion unit is used to insert a pause observation node before the risk training node; and the safety rollback node configuration unit is used to configure corresponding safety rollback nodes for different anomaly types. And / or, the execution control module is communicatively connected to the back lifting mechanism, hip assist mechanism, knee support limiting mechanism, foot support mechanism, handrail locking mechanism and pipeline follow-up management mechanism of the bedside training equipment, so as to control the patient to complete at least some of the training actions in the following order: lying up, sitting up at the bedside, partial weight-bearing and assisted standing, according to the current training path from lying down to standing. And / or, the training update module is configured to, after a training session, write the actual completed nodes, abnormal nodes, rollback nodes, pipeline constraint results, vital sign changes, posture stability changes, and assist limit parameters into the historical training record, and use this as input for the AI ​​nursing assessment module to output the admission results, assist limit parameters, and pipeline constraint results for the next time.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 6.

Images