Safety supervision method and system for internet rehabilitation nursing

Abstract

The invention discloses a safety supervision method and system for Internet rehabilitation nursing, and relates to the technical field of Internet medical treatment, and the method comprises the steps: building a dynamic space-time causal graph model based on collected data, defining a propagation path between nodes according to a biomechanical causal chain, and limiting a feature propagation time sequence through a causal time sequence verification layer; dynamically generating a gradient mask matrix based on the physiological rhythm represented by the heart rate variability data, and adjusting the gradient weight of the reinforcement learning strategy network by using the gradient mask matrix; in combination with electroencephalogram alertness wave signals and equipment torque parameters, performing cross-modal anomaly collaborative judgment through the adjusted reinforcement learning strategy network by utilizing a manifold envelope dynamic contraction mechanism; if it is judged that the cross-modal abnormity occurs, a dynamic safety threshold value is generated based on the electromyographic signals and patient individualized parameters, and switching of control modes of the rehabilitation equipment is driven according to the dynamic safety threshold value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of internet medical treatment, in particular to a safety supervision method and system for internet rehabilitation nursing. BACKGROUND

[0002] In the field of internet rehabilitation nursing, remote data acquisition technology is used to monitor patients and control rehabilitation equipment, which has improved rehabilitation efficiency and reduced manpower to a certain extent. However, the existing system has obvious shortcomings.

[0003] On the one hand, many current solutions only focus on single type data, such as judging risks based on single modal information such as device torque or joint angle, which makes the identification of abnormal state not timely enough and often lags behind. On the other hand, most safety thresholds use fixed preset values, which makes it difficult for device control strategies to match the actual tolerance of patients and accurately meet the changing needs of different patients.

[0004] At present, there is no effective solution to the above problems. SUMMARY

[0005] The embodiments of the present application provide a safety supervision method and system for internet rehabilitation nursing to solve the above technical problems.

[0006] The present application provides a safety supervision method for internet rehabilitation nursing, comprising:

[0007] The joint motion trajectory of the patient, the device torque parameter, the electroencephalogram alertness wave signal, the heart rate variability data and the electromyographic signal are synchronously collected by the biosensor, the rehabilitation device actuator and the environmental sensor;

[0008] A dynamic spatiotemporal causal graph model is constructed based on the collected data, the propagation path between nodes is defined according to the biomechanical causal chain, and the feature propagation time sequence is limited by the causal time sequence verification layer;

[0009] A gradient mask matrix is dynamically generated based on the physiological rhythm represented by the heart rate variability data, and the gradient weight of the reinforcement learning strategy network is adjusted using the gradient mask matrix, wherein the reinforcement learning strategy network is constructed based on the dynamic spatiotemporal causal graph model;

[0010] The electroencephalogram alertness wave signal and the device torque parameter are combined, and the cross-modal abnormality collaborative judgment is performed by the adjusted reinforcement learning strategy network using the manifold envelope dynamic shrinkage mechanism;

[0011] If it is determined that a cross-modal anomaly occurs, a dynamic safety threshold is generated based on the electromyographic signal and patient individualized parameters, and the control mode of the rehabilitation device is switched according to the dynamic safety threshold.

[0012] Further, the construction and training of the dynamic spatio-temporal causal graph model comprises:

[0013] Based on the causal chain of joint stress and muscle injury in the rehabilitation medical knowledge base, a one-way propagation path from the device node to the patient node is predefined, and the reverse propagation weight of the patient node to the device node is set to zero;

[0014] In the graph convolution layer of the spatio-temporal graph neural network, a timestamp verification module is embedded, the difference between the device node timestamp and the patient node timestamp is calculated, if the device node feature timestamp is later than the patient node feature timestamp, a decay penalty term is applied to the weight of the edge, and the decay coefficient is the product of the inverse of the time difference and a preset time-sensitive factor;

[0015] In the multi-head attention mechanism, a hard causal mask matrix is introduced, which allows the device node historical time feature to participate in the attention weight calculation of the patient node, and shields the attention propagation path of future time data;

[0016] In the backward propagation phase, according to the confidence value of the causal chain in the rehabilitation medical knowledge base, the gradient path from the device node to the patient node is dynamically truncated, and the truncation strength is inversely proportional to the confidence value.

[0017] Further, the implementation of cross-modal anomaly collaborative determination comprises:

[0018] The alpha wave energy in the preset frequency band of the electroencephalogram alert wave signal is extracted by multi-scale wavelet transform, when the alpha wave energy is detected to drop more than a first threshold value within a set time window, an electroencephalogram alert trigger signal is generated and the device parameter analysis mode is activated;

[0019] The device torque parameters, joint angles and current parameters are mapped to the latent space using a variational autoencoder, the probability density distribution of normal operation data is constructed, and the device parameter manifold deviation degree is calculated based on the dynamic Mahalanobis distance;

[0020] When the electroencephalogram alert trigger signal is activated, the device parameter deviation degree determination threshold is switched from a first static threshold to a second dynamic threshold, and the second dynamic threshold is a preset proportion of the first static threshold;

[0021] A time window synchronization verification mechanism is started, when the device parameter manifold deviation degree continuously exceeds the second dynamic threshold and the time overlap area with the electroencephalogram alert trigger signal exceeds a set proportion, it is determined that a cross-modal collaborative anomaly occurs;

[0022] The time-space superimposed visualization report of the device parameter manifold projection map and the electroencephalogram time-frequency map can mark the coupling area of the abnormal event in the time and space dimensions, and provide device control parameter modification suggestions.

[0023] Further, the implementation mode of the decay penalty term strength of the timestamp verification module dynamically associated with the rehabilitation stage is:

[0024] In the acute phase within 3 days after the operation, a first-order decay function of the time difference value is used, and the decay coefficient is the reciprocal of the time difference value multiplied by the acute phase sensitivity factor;

[0025] In the intermediate rehabilitation period, a second-order decay function of the time difference value is used, and the decay coefficient is the square reciprocal of the time difference value multiplied by the intermediate period sensitivity factor;

[0026] In the pre-discharge evaluation period, a fixed decay coefficient is used, and its value is dynamically adjusted by the risk level in the patient rehabilitation evaluation report.

[0027] Further, the implementation of the time window synchronization verification mechanism includes:

[0028] When the duration of the electroencephalogram alert trigger signal exceeds the set proportion threshold of the device parameter deviation duration, a secondary verification process is started;

[0029] In the secondary verification process, the time sequence relationship between the timestamps of the device control instruction sequence and the electroencephalogram alert trigger signal generated by the electroencephalogram alert wave signal is analyzed to detect whether there is an abnormal instruction mutation synchronized with the electroencephalogram signal in time;

[0030] If the abnormal instruction mutation amount exceeds the instruction mutation amount threshold, a cross-modal abnormality alarm is triggered, and a device parameter modification instruction set is generated;

[0031] The device parameter modification instruction set includes reducing the upper limit value of the device torque parameter, switching the compliant control mode, and pushing an emergency intervention notification to the rehabilitation therapist.

[0032] Further, the generation of the gradient mask matrix includes:

[0033] The sympathetic nerve activity index of the patient is extracted through the heart rate variability data analysis, and a dynamic ratio of the activity to a preset activity threshold is calculated;

[0034] When the dynamic ratio exceeds a first set threshold, the gradient weight of the device parameter modification action is increased to a preset multiple of the basic weight;

[0035] When the dynamic ratio is lower than a second set threshold, the gradient weight of the non-emergency operation is attenuated to a preset proportion of the basic weight;

[0036] During the night rest period, the gradient weight of the non-emergency operation is locked to a preset minimum value.

[0037] Further, when the reinforcement learning policy network performs gradient aggregation, the implementation of the knowledge graph constraint gradient weight includes:

[0038] For disease feature gradients marked as strong associations in the knowledge graph, set a minimum gradient weight threshold, and the minimum gradient weight threshold increases dynamically with the rehabilitation stage;

[0039] For medium association feature gradients, grade attenuation is performed according to their importance level in the current rehabilitation stage, and the attenuation ratio is inversely proportional to the grade;

[0040] For irrelevant feature gradients, apply a gradient weight upper limit, and the upper limit value is dynamically calculated by the medical relevance score of the feature in the knowledge graph.

[0041] Further, the generation of the dynamic safety threshold includes:

[0042] Calculate the real-time fatigue coefficient according to the electromyographic signal amplitude decay rate, and generate an age attenuation factor in combination with the patient's age;

[0043] Multiply the real-time fatigue coefficient, the age attenuation factor, and the medical history weight to generate a dynamically adjusted device force threshold;

[0044] When the real-time fatigue coefficient exceeds the preset fatigue threshold, activate the device compliance control mode, and dynamically adjust the motor stiffness value based on the fatigue coefficient;

[0045] If the fatigue coefficient continuously exceeds the preset fatigue threshold for a set period of time, an emergency braking instruction of the device is triggered.

[0046] Further, the assignment rule of the medical history weight is:

[0047] For patients within 3 days after surgery, superimpose a postoperative protection weight factor, and the value of the postoperative protection weight factor is dynamically calculated by the type of surgery and intraoperative complication records;

[0048] For patients with a history of cardiovascular disease, superimpose a cardiovascular risk weight factor, and its value is dynamically adjusted by the severity of the medical history and recent physiological indicators;

[0049] For other patients, use a basic weight factor, and the basic value is generated by comprehensive evaluation of the patient's age and underlying disease history.

[0050] The present application provides a safety supervision system for internet rehabilitation nursing, comprising:

[0051] A data acquisition module for synchronously acquiring patient joint motion trajectories, device torque parameters, electroencephalogram alert wave signals, heart rate variability data, and electromyographic signals through biological sensors, rehabilitation device actuators, and environmental sensors;

[0052]

[0052] a spatiotemporal causal graph model construction module, configured to construct a dynamic spatiotemporal causal graph model based on the collected data, define a propagation path between nodes according to a biomechanical causal chain, and limit a feature propagation time sequence through a causal time sequence verification layer;

[0053] a gradient mask matrix generation module, configured to dynamically generate a gradient mask matrix based on a physiological rhythm represented by the heart rate variability data, and adjust a gradient weight of a reinforcement learning strategy network by using the gradient mask matrix, wherein the reinforcement learning strategy network is constructed based on the dynamic spatiotemporal causal graph model;

[0054] an abnormality cooperative judgment module, configured to combine the electroencephalogram alert wave signal and the device torque parameter, and perform cross-modal abnormality cooperative judgment through the adjusted reinforcement learning strategy network by using a manifold envelope dynamic contraction mechanism;

[0055] a control mode switching module, configured to, if it is determined that a cross-modal abnormality occurs, generate a dynamic safety threshold based on the electromyogram signal and the patient individualized parameter, and drive switching of a control mode of the rehabilitation device according to the dynamic safety threshold.

[0056] Based on the embodiments provided in the present application, joint motion, device torque, electroencephalogram, heart rate, and electromyogram data are synchronously collected, a dynamic causal model is constructed, coupling analysis of device abnormality and patient physiological state is realized, and the problem of early warning lag caused by a single data source in a traditional scheme is solved. A device force upper limit is dynamically calculated based on a real-time electromyogram attenuation rate, age, and medical history, so that rehabilitation intensity is accurately matched with patient tolerance, and excessive force or insufficient training caused by a fixed threshold is avoided. Through the spatiotemporal causal graph model, device abnormality is forced to be earlier than patient abnormality, and a misjudgment scene (e.g., accidental synchronization of device vibration and patient action) in which “data co-occurrence but no causality” is excluded in traditional correlation analysis. BRIEF DESCRIPTION OF DRAWINGS

[0057] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the present application and, do not limit the present application in any way. In the drawings:

[0058] Figure 1 a flowchart of an optional safety supervision method for internet rehabilitation nursing according to an embodiment of the present application;

[0059] Figure 2 a flowchart of another optional safety supervision method for internet rehabilitation nursing according to an embodiment of the present application;

[0060] Figure 3 a structural diagram of an optional safety supervision system for internet rehabilitation nursing according to an embodiment of the present application.

[0061] The objectives, functional characteristics and advantages of the present application will be further described with reference to the embodiments in combination with the accompanying drawings. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without any creative work fall within the protection scope of the present application.

[0063] Optionally, as shown in the present application, a safety supervision method for internet rehabilitation nursing is provided, comprising: Figure 1

[0064] S101, synchronously collecting joint motion trajectory of a patient, device torque parameter, electroencephalogram vigilance wave signal, heart rate variability data and electromyogram signal through a biological sensor, a rehabilitation device actuator and an environmental sensor;

[0065] S102, constructing a dynamic space-time causal graph model based on the collected data, defining a propagation path between nodes according to a biomechanical causal chain, and limiting a feature propagation time sequence through a causal time sequence verification layer;

[0066] S103, dynamically generating a gradient mask matrix based on physiological rhythm represented by the heart rate variability data, and adjusting the gradient weight of a reinforcement learning strategy network by using the gradient mask matrix; wherein the reinforcement learning strategy network is constructed based on the dynamic space-time causal graph model;

[0067] S104, combining the electroencephalogram vigilance wave signal and the device torque parameter, and utilizing a manifold envelope dynamic contraction mechanism to perform cross-modal abnormality collaborative judgment through the adjusted reinforcement learning strategy network;

[0068] S105, if it is judged that cross-modal abnormality occurs, generating a dynamic safety threshold based on the electromyogram signal and individualized parameters of the patient, and driving the switching of the control mode of the rehabilitation device according to the dynamic safety threshold.

[0069] ​Based on the embodiments provided in the present application, by synchronously collecting joint motion, device torque, electroencephalogram, heart rate, electromyogram data, a dynamic causal model is constructed, coupling analysis of device abnormalities and patient physiological state is realized, and the early warning lag problem caused by a single data source in the traditional scheme is solved. Based on the real-time electromyogram attenuation rate, age, and disease history, the upper limit of device force is dynamically calculated, so that the rehabilitation intensity is accurately matched with the patient's tolerance, and excessive force or insufficient training caused by fixed threshold is avoided. Through the spatiotemporal causal graph model, device abnormalities are forced to be earlier than patient abnormalities, and the misjudgment scene of "data co-occurrence but no causality" in traditional correlation analysis (such as accidental synchronization of device vibration and patient action) is excluded.

[0070] Further, as shown in Figure 2 The construction and training of the dynamic spatiotemporal causal graph model include:

[0071] S201, based on the causal chain of joint stress and muscle injury in the rehabilitation medical knowledge base, a one-way propagation path from the device node to the patient node is predefined, and the reverse propagation weight of the patient node to the device node is set to zero;

[0072] S202, a timestamp verification module is embedded in the graph convolution layer of the spatiotemporal graph neural network, the difference between the device node timestamp and the patient node timestamp is calculated, if the device node feature timestamp is later than the patient node feature timestamp, a decay penalty term is applied to the weight of the edge, and the decay coefficient is the product of the inverse of the time difference and a preset time sensitive factor;

[0073] Optionally, in the present embodiment, the preset time sensitive factor can be set to 0.8, 0.7, etc.

[0074] S203, a hard causal mask matrix is introduced in the multi-head attention mechanism, allowing the device node historical time feature to participate in the attention weight calculation of the patient node, and shielding the attention propagation path of future time data;

[0075] S204, in the reverse propagation stage, according to the confidence value of the causal chain in the rehabilitation medical knowledge base, the gradient path from the device node to the patient node is dynamically truncated, and the truncation strength is inversely proportional to the confidence value.

[0076] Based on the embodiments provided in the present application, the time difference verification of the device and the patient node is embedded in the model training, ensuring that the device abnormal event is strictly earlier than the patient abnormal event on the time axis, and eliminating the reverse causal interference from the algorithm level. The gradient is dynamically truncated according to the confidence of the causal chain during reverse propagation, preventing the model from learning false correlations that violate the laws of biomechanics, such as reverse correction of device parameters by patient electromyogram abnormalities.

[0077] Further, the implementation of cross-modal abnormality collaborative judgment includes:

[0078] The alpha wave energy of a preset frequency band in the electroencephalogram vigilance wave signal is extracted by multi-scale wavelet transform, and when the alpha wave energy is detected to drop by more than a first threshold value within a set time window, an electroencephalogram vigilance trigger signal is generated and a device parameter analysis mode is activated;

[0079] Optionally, in the embodiment, the preset frequency band can include but is not limited to 8-13 Hz; the set time window can include but is not limited to 10 minutes, 15 minutes, etc.; and the first threshold value can include but is not limited to 30%, 20%, etc.

[0080] The device torque parameter, joint angle and current parameter are mapped to a latent space using a variational autoencoder, a probability density distribution of normal operation data is constructed, and a device parameter manifold deviation degree is calculated based on a dynamic Mahalanobis distance;

[0081] When the electroencephalogram vigilance trigger signal is activated, the device parameter deviation degree threshold value is switched from a first static threshold value to a second dynamic threshold value, and the second dynamic threshold value is a preset proportion value of the first static threshold value;

[0082] Optionally, in the embodiment, the first static threshold value includes but is not limited to 0.7, 0.6, etc.; and the second dynamic threshold value includes but is not limited to 80%, 70% of the first static threshold value, etc.

[0083] A time window synchronization verification mechanism is started, and when the device parameter manifold deviation degree continuously exceeds the second dynamic threshold value and the time overlap area with the electroencephalogram vigilance trigger signal exceeds a set proportion, it is determined that there is a cross-modal collaborative anomaly;

[0084] Optionally, in the embodiment, the set proportion includes but is not limited to 50%, 60%, etc.

[0085] A spatiotemporal superposition visualization report of the device parameter manifold projection map and the electroencephalogram time-frequency map is generated, the coupling area of the abnormal event in the time and space dimensions is marked, and a device control parameter correction suggestion is provided.

[0086] In a specific embodiment, when the device torque parameter exceeds the cross-modal dynamic threshold value (calculated based on the electroencephalogram vigilance wave), an abnormal marker is activated. After the abnormal marker is activated, if the mutation quantity of the device control instruction sequence exceeds an instruction mutation quantity threshold value, it is determined that there is a cross-modal collaborative anomaly.

[0087] The cross-modal dynamic threshold value is used to determine whether the device torque parameter deviates from the normal range, and the value lies in realizing the dynamic coupling of the device control parameter and the patient's physiological state. When the patient's electroencephalogram vigilance wave decreases, the cross-modal dynamic threshold value decreases accordingly, thereby improving the detection sensitivity of the device torque anomaly and ensuring that the device parameter is adjusted in time when the patient has a stress reaction.

[0088]

[0089] wherein, is a cross-modal dynamic threshold (unit: N·m) for device torque parameter anomaly determination; is a static determination threshold (unit: N·m) preset according to the rated torque of the device; is an electroencephalogram-device coupling coefficient, the value range is 0.5 to 0.8, and the default is 0.6; is an alpha wave energy baseline value (unit: μV²), which is taken from the mean value under the patient's calm state; is a real-time alpha wave energy value (unit: μV²). Based on the embodiments provided in the present application, the device parameter determination threshold is determined in real time by the alpha wave energy drop, the sensitivity of the patient's stress reaction driving device is improved, and the potential risk of false negatives is avoided. The spatiotemporal coupling report of the device parameters and the electroencephalogram signal is generated, providing a traceable abnormal event evidence chain for the rehabilitation therapist, and improving the reliability of clinical decision-making.

[0090] Further, the implementation mode of the decay penalty term strength of the time stamp verification module dynamically associated with the rehabilitation stage is:

[0091] In the acute stage within 3 days after the operation, a first-order decay function of the time difference value is used, and the decay coefficient is the reciprocal of the time difference value multiplied by the acute stage sensitive factor;

[0092] In the intermediate rehabilitation stage, a second-order decay function of the time difference value is used, and the decay coefficient is the square reciprocal of the time difference value multiplied by the intermediate stage sensitive factor;

[0093] In the pre-discharge evaluation stage, a fixed decay coefficient is used, and the value is dynamically adjusted by the risk level in the patient rehabilitation evaluation report.

[0094] In a specific embodiment, the dynamic decay of the spatiotemporal causal edge weight can be calculated based on the following formula:

[0095]

[0096] wherein, is the adjusted edge weight; is the initial edge weight; is the time stamp difference value (unit: seconds) between the device node and the patient node, if the device anomaly is later than the patient anomaly, then ; is the index of different types of rehabilitation stages, is the acute stage (0-3 days after the operation); is the intermediate stage (4-14 days after the operation); is the evaluation stage (>14 days after the operation); is a rehabilitation stage time sensitive factor, specifically, the acute stage: ; intermediate phase: ; evaluation phase: ; is a phase penalty coefficient, specifically, acute phase: ; intermediate phase: ; evaluation phase: .

[0097] Based on the above formula, through the exponential decay mechanism, it is ensured that the device abnormal event must be earlier than the physiological abnormality of the patient (such as the mutation of the electromyographic signal), and the pseudo-causal relationship of “patient abnormality affecting the device” is excluded from the model structure level. The acute phase adopts a more stringent time-sensitive factor and a penalty coefficient =, which fits the high-risk characteristics of the patient in the early postoperative period and avoids misjudgment caused by delayed response of the device.

[0098] Optionally, in the embodiment, the fixed attenuation coefficient can be determined according to the risk level, specifically, the low risk level is 0.5, the medium risk level is 0.3, and the high risk level is 0.1.

[0099] Based on the embodiments provided in the present application, the acute phase adopts a first-order decay function to strengthen the penalty, and the evaluation phase weakens the penalty, so that the model weight distribution fits the characteristics of the patient's recovery period.

[0100] Further, the implementation of the time window synchronization verification mechanism comprises:

[0101] When the duration of the electroencephalogram alert trigger signal exceeds the set proportion threshold of the device parameter deviation duration, a secondary verification process is started;

[0102] Optionally, in the embodiment, the set proportion threshold includes but is not limited to 60%, 70%, etc.

[0103] In the secondary verification process, the time sequence relationship between the time stamp of the device control instruction sequence and the electroencephalogram alert trigger signal generated by the electroencephalogram alert wave signal is analyzed, and it is detected whether there is an abnormal instruction mutation synchronized with the electroencephalogram signal; if the abnormal instruction mutation exceeds the instruction mutation threshold, a cross-modal abnormality alarm is triggered, and a device parameter correction instruction set is generated; wherein the instruction mutation threshold is used to determine whether the mutation amplitude of the device control instruction sequence is abnormal, which is related to the time sequence change of the device control instruction sequence.

[0104] Wherein, the abnormal instruction mutation is obtained by calculating the difference between the standard deviation of the device control instruction sequence in the set time window and the baseline value. The specific formula is: ; wherein, is the abnormal instruction mutation; is the value of the control instruction; is the baseline average value; ​To set the number of samples within a time window.

[0105] The device parameter correction instruction set includes lowering the upper limit value of the device torque parameter, switching the compliant control mode, and pushing an emergency intervention notification to the therapist, which is consistent with the cross-modal abnormality collaborative determination logic based on the reinforcement learning strategy network.

[0106] Based on the embodiments provided in the present application, the time sequence relationship between the device control instruction and the abnormality of the brain electrical signal is analyzed, and accidental abnormality labels caused by sensor noise or communication delay are excluded.

[0107] Further, the generation of the gradient mask matrix comprises:

[0108] The heart rate variability data analysis extracts a patient sympathetic nerve activity index, and calculates a dynamic ratio of the activity to a preset activity threshold value;

[0109] Optionally, in the present embodiment, the preset activity threshold value includes but is not limited to 70%, 60%, etc.

[0110] When the dynamic ratio exceeds a first set threshold value, the gradient weight of the device parameter modification action is increased to a preset multiple of the basic weight;

[0111] Optionally, in the present embodiment, the first set threshold value includes but is not limited to 70%, 75%, etc.; and the preset multiple can include but is not limited to 2 times, 1.5 times, etc.

[0112] When the dynamic ratio is lower than a second set threshold value, the gradient weight of the non-emergency operation is attenuated to a preset proportion of the basic weight;

[0113] Optionally, in the present embodiment, the second set threshold value includes but is not limited to 30%, 40%, etc.; and the preset proportion includes but is not limited to 50%, 60%, etc.

[0114] During the night rest period, the gradient weight of the non-emergency operation is locked to a preset minimum value, and the adjustment of the gradient weight is consistent with the reinforcement learning strategy network and the patient physiological rhythm logic.

[0115] Based on the embodiments provided in the present application, the gradient weight is dynamically adjusted through the sympathetic nerve activity, and the non-emergency operation is forcibly locked at night, avoiding the interference of the therapist's misoperation with the patient's rest.

[0116] Further, when the reinforcement learning strategy network performs gradient aggregation, the implementation of the knowledge graph to constrain the gradient weight comprises:

[0117] The disease feature gradient marked as a strong correlation in the knowledge graph is set with a minimum gradient weight threshold value, and the minimum gradient weight threshold value is dynamically increased with the rehabilitation stage, for example, increased by 0.1 every week.

[0118] Optionally, in the embodiment, the minimum gradient weight threshold includes but is not limited to 0.8, 0.7, etc.

[0119] For the medium correlation feature gradient, the importance level thereof in the current rehabilitation stage is graded and decayed, and the decay ratio is inversely proportional to the level;

[0120] Optionally, in the embodiment, the importance level can include first level, second level, and third level; and the decay ratio being inversely proportional to the level includes but is not limited to first level decay of 20%, second level decay of 40%, and third level decay of 60%.

[0121] For the irrelevant feature gradient, a gradient weight upper limit is applied, and the upper limit value is dynamically calculated according to the medical correlation score of the feature in the knowledge graph. The constraint of the knowledge graph on the gradient weight is consistent with the model constructed based on the collected data and the logic of the patient's rehabilitation situation.

[0122] Optionally, in the embodiment, the gradient weight upper limit can include but is not limited to an upper limit value of 0.3, 0.4, etc.

[0123] Based on the embodiments provided in the present application, a gradient weight lower limit is set for the strong correlation feature such as the electromyographic signal of the spinal cord injury patient, to ensure that the model preferentially captures the core medical indicators.

[0124] Further, the generation of the dynamic safety threshold comprises:

[0125] The real-time fatigue coefficient is calculated according to the electromyographic signal amplitude decay rate, and the age decay factor is generated in combination with the patient's age;

[0126] The real-time fatigue coefficient, the age decay factor, and the medical history weight are multiplied to generate a dynamically adjusted device force threshold;

[0127] When the real-time fatigue coefficient exceeds the preset fatigue threshold, the device compliance control mode is activated, and the motor stiffness value is dynamically adjusted based on the fatigue coefficient;

[0128] If the fatigue coefficient continuously exceeds the preset fatigue threshold for a set time length, an emergency braking instruction of the device is triggered, and the generation of the dynamic safety threshold and the related operations are consistent with the cross-modal abnormality judgment result and the logic of the patient's physical condition.

[0129] Optionally, in the embodiment, the set time length can include but is not limited to 5 minutes and 6 minutes.

[0130] Based on the embodiments provided in the present application, the electromyographic decay rate is converted into a real-time fatigue coefficient to drive the device to switch to a compliance control mode, thereby reducing secondary injury caused by high-load training.

[0131] Further, the assignment rule of the medical history weight is:

[0132] For patients within 3 days after surgery, a postoperative protection weight factor is superimposed, and the value of the postoperative protection weight factor is dynamically calculated according to the type of surgery and intraoperative complications;

[0133] Optionally, in this embodiment, the postoperative protection weight factor includes but is not limited to 1.5 for abdominal surgery and 2 for joint replacement surgery.

[0134] For patients with a history of cardiovascular disease, a cardiovascular risk weight factor is superimposed, and its value is dynamically adjusted according to the severity of the medical history and recent physiological indicators.

[0135] Optionally, in this embodiment, the cardiovascular risk weight factor can include but is not limited to 1.2 for mild history, 1.5 for moderate history, and 1.8 for severe history.

[0136] For other patients, a basic weight factor is used, and the basic value is generated by comprehensive evaluation of patient age and medical history.

[0137] In a specific embodiment, the dynamic safety threshold can be calculated based on the following formula:

[0138]

[0139] wherein, is the dynamic safety threshold (unit: N·m); is the basic safety threshold (unit: N·m), which is preset according to the device model; is the electromyographic signal decay rate (unit: % / s), , is the baseline amplitude, is the real-time amplitude, is the monitoring time window; is the age attenuation factor, is the patient age (unit: years); is the medical history weight factor, is the postoperative day, is the systolic blood pressure (unit: mmHg), for example, for postoperative patients: for patients with a history of cardiovascular disease: and for other patients: . .

[0140] Based on the above formula, the electromyographic decay rate, age, medical history and other multi-dimensional parameters are integrated to dynamically adjust the upper limit of the device force, avoiding excessive training or insufficient rehabilitation caused by fixed threshold. When the electromyographic decay rate continues to exceed the limit, a soft control mode is triggered, such as reducing the motor stiffness to reduce the risk of muscle strain, especially for patients in the early postoperative period.

[0141] Based on the embodiments provided in the application, the safety threshold is automatically reduced within 3 days after the operation, and the weight is dynamically calculated in combination with the intraoperative complication record to meet the conservative needs of early rehabilitation.

[0142] Optionally, as shown in the application, a safety supervision system for internet rehabilitation nursing is provided, comprising: Figure 3

[0143] The data acquisition module 301 is configured to synchronously acquire joint movement trajectories of a patient, device torque parameters, electroencephalogram vigilance wave signals, heart rate variability data and electromyogram signals through biosensors, rehabilitation device actuators and environmental sensors.

[0144] The spatiotemporal causal graph model construction module 302 is configured to construct a dynamic spatiotemporal causal graph model based on the acquired data, define a propagation path between nodes according to a biomechanical causal chain, and limit a feature propagation time sequence through a causal time sequence verification layer.

[0145] The gradient mask matrix generation module 303 is configured to dynamically generate a gradient mask matrix based on physiological rhythm represented by the heart rate variability data, and adjust gradient weights of a reinforcement learning strategy network using the gradient mask matrix, wherein the reinforcement learning strategy network is constructed based on the dynamic spatiotemporal causal graph model.

[0146] The abnormality cooperative judgment module 304 is configured to combine the electroencephalogram vigilance wave signals and the device torque parameters, utilize a manifold envelope dynamic shrinkage mechanism, and perform cross-modal abnormality cooperative judgment through the adjusted reinforcement learning strategy network.

[0147] The control mode switching module 305 is configured to generate a dynamic safety threshold based on the electromyogram signals and patient individualized parameters if cross-modal abnormalities are determined, and drive switching of a control mode of a rehabilitation device according to the dynamic safety threshold.

[0148] It should be noted that in the present application, the embodiments implemented by the safety supervision system for internet rehabilitation nursing can be mutually referenced with the embodiments implemented by the safety supervision method for internet rehabilitation nursing, and the present application will not be repeated here.

[0149] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.​

Claims

1. A safety supervision method for Internet rehabilitation nursing, characterized in that: include: Through biosensors, rehabilitation equipment actuators and environmental sensors, the patient's joint motion trajectory, equipment torque parameters, EEG alertness wave signals, heart rate variability data and electromyographic signals are synchronously collected; Based on the collected data, a dynamic spatiotemporal causal graph model is constructed, the propagation paths between nodes are defined according to the biomechanical causal chain, and the feature propagation timing is restricted through the causal timing verification layer; Dynamically generating a gradient mask matrix based on the physiological rhythm represented by the heart rate variability data, and using the gradient mask matrix to adjust the gradient weight of a reinforcement learning strategy network; wherein the reinforcement learning strategy network is constructed based on the dynamic spatiotemporal causal graph model; Combining the EEG alertness wave signal with the device torque parameter, utilizing the manifold envelope dynamic contraction mechanism, and performing cross-modal anomaly collaborative judgment through an adjusted reinforcement learning strategy network; If a cross-modal abnormality is determined to occur, a dynamic safety threshold is generated based on the electromyographic signal and the patient's individualized parameters, and the control mode of the rehabilitation device is driven to switch according to the dynamic safety threshold.

2. The safety supervision method for Internet rehabilitation nursing according to claim 1 is characterized in that: The construction and training of the dynamic spatiotemporal causal graph model includes: Based on the causal chain of joint force and muscle injury in the rehabilitation medicine knowledge base, a one-way propagation path from the device node to the patient node is predefined, and the backpropagation weight of the patient node to the device node is set to zero; A timestamp verification module is embedded in the graph convolution layer of the spatiotemporal graph neural network to calculate the difference between the device node timestamp and the patient node timestamp. If the device node feature timestamp is later than the patient node feature timestamp, an attenuation penalty term is applied to the weight of the edge. The attenuation coefficient is the product of the inverse of the time difference and a preset time sensitivity factor. A hard causal mask matrix is ​​introduced into the multi-head attention mechanism, allowing the historical features of the device node to participate in the attention weight calculation of the patient node and shielding the attention propagation path of future data; In the back-propagation stage, the gradient path from the device node to the patient node is dynamically truncated according to the confidence value of the causal chain in the rehabilitation medicine knowledge base, and the truncation strength is inversely proportional to the confidence value.

3. The safety supervision method for Internet rehabilitation nursing according to claim 1 is characterized in that: The implementation of cross-modal anomaly collaborative judgment includes: Extracting the alpha wave energy of a preset frequency band from the EEG alertness wave signal through multi-scale wavelet transform. When it is detected that the alpha wave energy drops below a first threshold within a set time window, an EEG alertness trigger signal is generated and the device parameter analysis mode is activated. Using a variational autoencoder to perform latent space mapping on the device torque parameters, joint angles, and current parameters, constructing a probability density distribution of normal operation data, and calculating the device parameter manifold deviation based on dynamic Mahalanobis distance; When the EEG alertness trigger signal is activated, the device parameter deviation determination threshold is switched from the first static threshold to the second dynamic threshold, where the second dynamic threshold is a preset proportional value of the first static threshold; Initiate a time window synchronization verification mechanism. When the device parameter manifold deviation continuously exceeds the second dynamic threshold and the time overlap area with the EEG alertness trigger signal exceeds a set ratio, it is determined to be a cross-modal collaboration anomaly. Generate a spatiotemporal visualization report of the device parameter manifold projection diagram and the EEG time-frequency diagram, mark the coupling areas of abnormal events in the time and space dimensions, and provide correction suggestions for device control parameters.

4. The safety supervision method for Internet rehabilitation nursing according to claim 2 is characterized in that: The implementation method of dynamically associating the intensity of the attenuation penalty term of the timestamp verification module with the recovery stage is as follows: During the acute phase within 3 days after surgery, the first-order attenuation function of the time difference is used, and the attenuation coefficient is the inverse of the time difference multiplied by the acute phase sensitivity factor; In the middle stage of rehabilitation, the second-order attenuation function of the time difference is used, and the attenuation coefficient is the inverse of the square of the time difference multiplied by the middle stage sensitivity factor; A fixed attenuation coefficient is used during the pre-discharge assessment period, and its value is dynamically adjusted according to the risk level in the patient's rehabilitation assessment report.

5. The safety supervision method for Internet rehabilitation nursing according to claim 3 is characterized in that: The implementation of the time window synchronization verification mechanism includes: When the duration of the EEG alertness trigger signal exceeds the set ratio threshold of the device parameter deviation duration, the secondary verification process is initiated; In the secondary verification process, the time sequence relationship between the timestamp of the device control instruction sequence and the EEG alertness trigger signal generated by the EEG alertness wave signal is analyzed to detect whether there is an abnormal instruction mutation synchronized with the EEG signal time; If the abnormal instruction mutation exceeds the instruction mutation threshold, a cross-modal abnormality alarm is triggered and a device parameter correction instruction set is generated; The device parameter correction instruction set includes lowering the upper limit of the device torque parameter, switching the compliant control mode, and pushing an emergency intervention notification to the rehabilitation therapist.

6. The safety supervision method for Internet rehabilitation nursing according to claim 1 is characterized in that: The generation of the gradient mask matrix includes: Extracting the patient's sympathetic nerve activity index through the heart rate variability data analysis, and calculating the dynamic ratio of the activity to a preset activity threshold; When the dynamic ratio exceeds the first set threshold, the gradient weight of the device parameter modification action is increased to a preset multiple of the basic weight; When the dynamic ratio is lower than a second set threshold, the gradient weight of the non-emergency operation is decayed to a preset ratio of the basic weight; During the nighttime rest period, the gradient weights for non-urgent operations are locked to a preset minimum value.

7. The safety supervision method for Internet rehabilitation nursing according to claim 1 is characterized in that: When performing gradient aggregation in the reinforcement learning policy network, the implementation of knowledge graph constraint gradient weights includes: For the disease feature gradients marked as strongly associated in the knowledge graph, a minimum gradient weight threshold is set, and the minimum gradient weight threshold increases dynamically with the recovery stage; For medium-correlated feature gradients, graded attenuation is performed according to their importance level in the current rehabilitation stage, and the attenuation ratio is inversely proportional to the level; For the gradient of unrelated features, an upper limit on the gradient weight is imposed, and the upper limit value is dynamically calculated by the medical relevance score of the feature in the knowledge graph.

8. The security supervision method for Internet rehabilitation nursing according to claim 1 is characterized in that: The generation of the dynamic security threshold includes: The real-time fatigue coefficient is calculated based on the attenuation rate of the electromyographic signal amplitude, and the age attenuation factor is generated in combination with the patient's age; Multiply the real-time fatigue coefficient, age attenuation factor and medical history weight to generate a dynamically adjusted device force threshold; When the real-time fatigue coefficient exceeds the preset fatigue threshold, the device's compliant control mode is activated, and the motor stiffness value is dynamically adjusted based on the fatigue coefficient; If the fatigue coefficient continues to exceed the preset fatigue threshold for a set period of time, the equipment emergency braking command is triggered.

9. The safety supervision method for Internet rehabilitation nursing according to claim 8, characterized in that: The assignment rule of the medical history weight is: For patients within 3 days after surgery, a postoperative protection weight factor is added, and the value of the postoperative protection weight factor is dynamically calculated based on the type of surgery and intraoperative complication records; For patients with a history of cardiovascular disease, a cardiovascular risk weighting factor is added, and its value is dynamically adjusted according to the severity of the medical history and recent physiological indicators; For other patients, a basic weight factor is used, and the basic value is generated by a comprehensive assessment of the patient's age and underlying disease history.

10. A security supervision system for Internet rehabilitation nursing, characterized in that: include: The data acquisition module is used to synchronously collect the patient's joint motion trajectory, equipment torque parameters, EEG alertness wave signals, heart rate variability data and electromyographic signals through biosensors, rehabilitation equipment actuators and environmental sensors; The spatiotemporal causal graph model construction module is used to construct a dynamic spatiotemporal causal graph model based on the collected data, define the propagation path between nodes according to the biomechanical causal chain, and limit the feature propagation timing through the causal timing verification layer; a gradient mask matrix generation module, configured to dynamically generate a gradient mask matrix based on the physiological rhythm represented by the heart rate variability data, and to use the gradient mask matrix to adjust the gradient weights of a reinforcement learning strategy network, wherein the reinforcement learning strategy network is constructed based on the dynamic spatiotemporal causal graph model; An abnormality collaborative judgment module, used to combine the EEG alertness wave signal and the device torque parameter, use the manifold envelope dynamic contraction mechanism, and perform cross-modal abnormality collaborative judgment through an adjusted reinforcement learning strategy network; The control mode switching module is used to generate a dynamic safety threshold based on the electromyographic signal and the patient's individualized parameters if a cross-modal abnormality is determined to have occurred, and to drive the switching of the control mode of the rehabilitation device according to the dynamic safety threshold.

Images