Behavior monitoring and risk early warning system for obstetrical patient

Abstract

The invention relates to the technical field of deep learning, and discloses a behavior monitoring and risk early warning system for obstetrical patients, which comprises a data sensing acquisition module, a data analysis processing module, a human body area modeling module, a behavior early warning module, an edge calculation module and a comprehensive judgment module. According to the system, pixel-level segmentation and key point regression are performed on patient image data by constructing a mask key point-continuous time attention fusion network, a continuous time attention mechanism inspired by a neuron loop is introduced, dynamic modeling and weight distribution are performed on continuous frame attitude features, and a continuous frame attitude fusion model is constructed. Therefore, the continuous time evolution trend of the posture change of the obstetrical patient is described. And in combination with human body area modeling and behavior early warning strategies, intelligent identification and early warning of risk behaviors such as postpartum dysphoria, abnormal turning over and falling down are realized. The accuracy and stability of behavior monitoring are improved, and the method is suitable for intelligent safety management in an obstetrical monitoring scene.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a behavior monitoring and risk warning system for obstetric patients. Background Technology

[0002] With the deepening application of artificial intelligence and intelligent computing technologies in the field of medical monitoring, patient behavior analysis and risk warning technologies based on computational models have gradually become research hotspots. Existing intelligent technologies mostly employ convolutional neural networks (CNNs), pose keypoint detection models, recurrent neural networks (RNNs), or attention-based temporal models to automatically analyze patient images or sensor data to identify abnormal behaviors or potential risks. However, these technologies still have significant limitations in actual obstetric monitoring scenarios: on the one hand, most existing intelligent methods focus on feature modeling within discrete time steps or fixed time windows, and their temporal dependencies are usually characterized through discrete recursion or static attention weights, making it difficult to effectively describe the cumulative effect and evolutionary trend of obstetric patient posture changes over continuous time scales; on the other hand, existing attention mechanisms generally adopt static weight calculation methods based on dot products or linear mappings, lacking the ability to jointly model time constants, adaptive decay, and target driving strength, which can easily lead to misjudgments when faced with complex behavioral patterns such as postpartum weakness, intermittent agitation, and progressive abnormal movements. Furthermore, while some intelligent solutions incorporate multimodal sensing data, their model structures often employ simple splicing or parallel fusion methods, failing to achieve unified modeling of posture characteristics, temporal dependencies, and behavioral risk contributions at the computational model level. This limits the system's ability to represent complex ward environments and continuous behavioral evolution processes. Therefore, there is an urgent need for an intelligent behavior monitoring and risk warning technology based on advanced computational models, capable of dynamically modeling the posture characteristics of obstetric patients within a continuous time frame and accurately depicting the evolutionary trends of abnormal behaviors to meet the practical needs of obstetric patient safety monitoring. Summary of the Invention

[0003] This invention addresses the shortcomings of existing intelligent behavior monitoring technologies in continuous-time modeling capabilities, dynamic modulation of attention weights, and characterization of complex behavioral evolution. It proposes a behavior monitoring and risk warning system for obstetric patients. The core innovation of this invention lies in elevating the patient posture behavior modeling problem from a traditional discrete-time or static attention framework to a computational modeling problem of continuous-time state evolution. By introducing a dual-gated continuous-time attention mechanism inspired by neuronal circuit dynamics, the cumulative effect, decay characteristics, and abnormal driving intensity of posture features over time are uniformly represented as a learnable continuous-time dynamic process. Specifically, this invention no longer employs fixed or instantaneous attention weight calculation methods. Instead, it models the attention state corresponding to the query-key pair as a first-order continuous-time dynamic system, allowing the attention weights to adaptively evolve with pseudo-time, thereby achieving a fine characterization of posture change trends at the computational model level. Simultaneously, through a sparse connection topology and group modulation structure inspired by neuronal circuits, a dual-gated mechanism of content target gating and time constant gating is constructed, enabling the model to learn the driving contribution of posture features to abnormal behavior and their accumulation and decay patterns over continuous time scales. Based on the above computational model, this invention achieves high-precision modeling and risk warning of complex and progressive abnormal behavior patterns of obstetric patients while maintaining model interpretability and edge computing deployability. This significantly improves the expressive power and predictive reliability of the intelligent behavior monitoring system in continuous time scenarios.

[0004] This invention provides a behavior monitoring and risk warning system for obstetric patients. The system includes a data sensing and acquisition module, a data analysis and processing module, a human body region modeling module, a behavior warning module, an edge computing module, and a comprehensive judgment module.

[0005] The data sensing and acquisition module acquires continuous and stable patient image data, forming multimodal sensing data.

[0006] The data analysis and processing module establishes a masked keypoint-continuous temporal attention fusion network and processes patient image data through this network. During processing, a multi-task joint loss function is introduced to jointly optimize the model parameters of the masked keypoint-continuous temporal attention fusion network, enabling the network to simultaneously complete pixel-level segmentation and keypoint regression tasks. It outputs the coordinate information of 17 key points corresponding to the obstetric patient's body and generates pose anomaly judgment information based on the continuous temporal attention modeling results. The masked keypoint-continuous temporal attention fusion network includes a convolutional neural network (CNN), a region proposal network (RPN), a masking segmentation branch, a keypoint detection branch, and a continuous temporal attention mechanism inspired by neuronal circuits.

[0007] The human body region modeling module defines a semantic rule correspondence table, and performs key point grouping and region mapping modeling processing according to the key point-region correspondence relationship defined in the semantic rule correspondence table, mapping and dividing 6 human body regions.

[0008] Behavior warning module: Calculates the spatial displacement of key points in each of the six human body regions using the Chebyshev distance calculation formula, determines abnormal movement states, and generates behavior warning information;

[0009] Edge computing module: Uses TinyML to perform fusion and inference on multimodal perception data at the bedside, and outputs a three-level behavior level;

[0010] The comprehensive judgment module combines posture anomaly judgment information, behavior warning information, and three-level behavior level to make a comprehensive judgment, generate a comprehensive warning result, and output it to the execution end to trigger prompts, records, and constrain execution actions.

[0011] Furthermore, the data analysis and processing module processes the patient image data through a masked keypoint-continuous-time attention fusion network, outputting the coordinate information of 17 key points to generate pose anomaly determination information. This process specifically includes the following steps:

[0012] Step S1: Input the patient image data into the CNN to obtain the patient environment depth feature map;

[0013] Step S2: Input the patient environment depth feature map into the RPN, and use the anchor box generation mechanism, sliding window search and non-maximum suppression algorithm to filter and obtain a set of candidate regions for body parts;

[0014] Step S3: Perform spatial scale alignment on the candidate region set of body parts to obtain a candidate region set of uniform size;

[0015] Step S4: Input the uniform-sized candidate region set into the mask segmentation branch and the key point detection branch; the mask segmentation branch performs pixel-level segmentation on the candidate regions and outputs a pixel-level segmentation mask of the patient's body, which is used to extract the global contour change features of the patient's body. The global contour change features include the overall contour shape of the patient, spatial position distribution, and contour change information over time, thereby achieving accurate positioning of the patient's body; the key point detection branch performs key point regression on the candidate regions and outputs the coordinate information of 17 key points corresponding to the patient's body, and constructs the kinematic features of the skeleton structure based on the coordinate information of the 17 key points, which is used to characterize the spatial positional relationship between the joints of the obstetric patient's limbs and their movement changes over time;

[0016] Step S5: Based on the global contour change features and the kinematic features of the skeleton structure, construct a continuous frame temporal feature sequence in chronological order. Use a continuous temporal attention mechanism inspired by neuronal circuits to process the continuous frame temporal feature sequence. Perform continuous temporal dependency modeling and dynamic weight allocation on the posture features at different times to obtain a continuous temporal context feature representation that reflects the trend of posture change. Based on the continuous temporal context feature representation, conduct a comprehensive analysis of the temporal consistency of the patient's posture changes, motion accumulation features, and abnormal evolution trend. When the analysis results show that there is a continuous or gradually increasing abnormal movement trend, generate posture abnormality judgment information.

[0017] Furthermore, step S5 specifically includes the following:

[0018] Step S51: Map the continuous frame temporal feature sequence into query feature vector, key feature vector, and value feature vector; for each query feature, calculate its correlation measure with each key feature within the time window, and based on the correlation measure, use the Top-K sparse filtering strategy to select the K key features with the highest correlation for each query feature to obtain the filtered key features; concatenate the filtered key features with the corresponding query features to form sparse query-key joint features;

[0019] Step S52: Construct a neuron-loop-inspired dual-gating mechanism. The sparse query-key joint features are input into the neuron-loop-inspired dual-gating mechanism. Based on a neuron-loop-inspired sparse connection structure and a grouping modulation strategy, the mechanism performs nonlinear mapping on the sparse query-key joint features, generating content target gating parameters and time constant gating parameters for continuous-time attention modeling. The construction method of the neuron-loop-inspired dual-gating mechanism is as follows: Based on a gated neural network, by introducing a neuron-loop-inspired sparse connection topology and a grouping modulation strategy, the modeling method in the gated neural network, which only statically adjusts feature weights, is structurally improved. Based on this structural improvement, dual-gating branches are constructed to generate content target gating parameters and time constant gating parameters, respectively. This allows the gating mechanism to not only adjust the expression intensity of pose features but also control the accumulation rate and decay characteristics of pose features in the continuous-time dimension, thus constructing the neuron-loop-inspired dual-gating mechanism.

[0020] Step S53: Treat the attention state of each query-key pair corresponding to the sparse query-key joint features as a state variable that evolves continuously over time. Introduce a first-order continuous-time dynamic model to model the change process of the attention state over time. Through content target gating parameters and time constant gating parameters, make the attention state exhibit dynamic evolution characteristics that are jointly modulated by target-driven and time decay in the time dimension. Construct a continuous-time attention state evolution model to characterize the cumulative change trend of obstetric patients' postures on a continuous time scale.

[0021] Step S54: Preset a pseudo-time interval, and numerically solve the continuous-time attention state evolution model based on the pseudo-time interval to obtain the attention state value corresponding to each time step; normalize the attention state value to generate continuous-time attention weights, which are used to represent the relative importance of pose features in different time frames in the overall pose evolution process.

[0022] Step S55: The posture features of the corresponding time step are weighted and fused according to the continuous time attention weight to obtain the fused continuous time posture context feature representation; wherein, the weighted fusion process comprehensively considers the cumulative contribution of posture features in the time dimension, rather than relying solely on instantaneous posture changes; based on the continuous time posture context feature representation, the temporal consistency, motion accumulation characteristics and abnormal evolution trend of the posture changes of obstetric patients are comprehensively analyzed; when the analysis results meet the preset abnormal judgment conditions, posture abnormality judgment information is generated.

[0023] Furthermore, the construction method of the multi-task joint loss function is as follows: introduce the classification loss function, the bounding box regression loss function, and the mask loss function, and integrate the above three types of loss functions to construct the multi-task joint loss function.

[0024] Furthermore, the behavior warning module generates behavior warning information by: calculating the spatial displacement of key points in each of the six human body regions using the Chebyshev distance calculation formula, obtaining the change in displacement distance and duration of key points; setting movement distance thresholds and duration thresholds; when the change in displacement distance of all key points in any human body region is greater than the movement distance threshold and the duration of key points exceeds the duration threshold, it is determined that the human body region is in an abnormal movement state, and behavior warning information is generated.

[0025] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0026] This invention introduces a continuous-time attention calculation model inspired by neural circuits, enabling dynamic modeling of obstetric patients' postural behavior on a continuous time scale. This enhances the overall perception capability of the behavior monitoring system regarding postural change trends, solving the problem that existing intelligent technologies rely solely on discrete time steps or fixed time windows for analysis, making it difficult to characterize the cumulative evolution of behavior. This improves the accuracy and stability of early warning in complex behavioral scenarios such as postpartum restlessness, progressive abnormal turning over, and potential fall risks. By treating the attention weights of postural features as state variables evolving continuously over time, the system can identify the essential differences between short-term micro-movements and sustained abnormal behavior, thus providing a more reliable basis for judging behavioral trends in the safety monitoring of obstetric patients.

[0027] This invention, through the construction of a neuronal circuit-inspired dual-gating mechanism, achieves collaborative modeling of the driving strength and temporal accumulation characteristics of abnormal posture features. This enhances the ability to distinguish the contribution of posture features in the behavioral risk assessment process, and solves the problem of static weight calculation in existing attention mechanisms, which makes it difficult to reflect the actual influence of different posture features in the formation of abnormal behavior. This strengthens the invention's ability to perform refined analysis on unplanned turning over, abnormal limb exertion, and potential extubation behaviors in obstetric patients. By jointly modulating content-target gating and time constant gating, this invention can adaptively adjust the abnormality assessment results according to the persistence and rate of change of posture features, avoiding false alarms caused by brief interference or physiological micro-movements.

[0028] This invention deeply integrates a continuous-time attention computation model with mask segmentation, key point detection, and multimodal perception results, achieving unified modeling of posture spatial structure information and temporal evolution information. This enhances the robustness and generalization ability of the obstetric patient behavior monitoring system in complex ward environments, solves the problem of unstable recognition performance under conditions of occlusion, lighting changes, and environmental interference using single features or simple fusion methods, and strengthens the real-time operation capability and engineering applicability of this invention in bedside edge computing scenarios. Thus, while ensuring the safety of obstetric patients, it effectively improves the practical value and clinical application feasibility of the behavior monitoring and risk warning system. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of a module for a behavior monitoring and risk warning system for obstetric patients provided by the present invention;

[0030] Figure 2 This is a flowchart illustrating the mask key point-continuous time attention fusion network proposed in Example 2. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] Example 1, according to Figure 1 This invention provides a behavior monitoring and risk warning system for obstetric patients, applicable to obstetric intensive care units, postpartum recovery wards, or medical facilities with continuous monitoring capabilities. It is used for continuous behavior monitoring and risk warning intervention for obstetric patients at risk of unplanned extubation, postpartum agitation, confusion, limited ability to turn over independently, or postpartum weakness leading to falls. The system includes a data sensing and acquisition module, a data analysis and processing module, a human body region modeling module, a behavior warning module, an edge computing module, and a comprehensive judgment module.

[0033] The data sensing and acquisition module, deployed at the bedside and above the maternity ward, is used for multi-source, non-contact, and low-invasive sensing and acquisition of the behavioral status of maternity patients. Specifically, it includes:

[0034] Cameras deployed at the bedside and above the maternity ward continuously capture video of maternity patients at 20 frames per second and 640×480 resolution, obtaining continuous and stable patient image data of the patient's torso, limbs and bedside environment.

[0035] The patient's conscious EEG characteristics and finger flexion electromyography data are detected by a bioelectric sensor array. The pressure change data of the magnitude, direction and duration of the force applied by the patient's limbs are detected by a micro-motion pressure sensor network constructed by flexible piezoresistive fibers embedded in the restraint band.

[0036] A 60 GHz millimeter-wave radar, placed next to the bed, is used to monitor the movement trajectory, range of motion, and micro-motion characteristics of the patient's large limbs and upper limbs in real time, so as to enhance the ability to perceive the behavior of obstetric patients at night, under obstructed or insufficient light conditions.

[0037] Through the above-mentioned multi-source sensing methods, multimodal sensing data containing video, EEG, EMG, pressure, and millimeter-wave motion information is formed;

[0038] The data analysis and processing module establishes a masked keypoint-continuous temporal attention fusion network and processes patient image data through this network. During processing, a multi-task joint loss function is introduced to jointly optimize the model parameters of the masked keypoint-continuous temporal attention fusion network, enabling the network to simultaneously complete pixel-level segmentation and keypoint regression tasks. It outputs the coordinate information of 17 keypoints corresponding to the obstetric patient's body and generates posture anomaly judgment information reflecting the trend of posture changes in obstetric patients based on the continuous temporal attention modeling results. The masked keypoint-continuous temporal attention fusion network includes a convolutional neural network (CNN), a region proposal network (RPN), a masking segmentation branch, a keypoint detection branch, and a continuous temporal attention mechanism inspired by neuronal circuits.

[0039] The human body region modeling module defines a semantic rule correspondence table. It assigns the coordinates of 17 key points to the corresponding regions as defined in the table, performing key point grouping and region mapping modeling. This divides the obstetric patient's body into six regions: head and neck, trunk, left upper limb, right upper limb, left lower limb, and right lower limb, supporting region-level behavioral analysis and risk assessment. The semantic rule correspondence is shown in Table 1.

[0040] Table 1

[0041] ;

[0042] Behavioral warning module: Calculates the spatial displacement of key points in each of the six human body regions using the Chebyshev distance formula, determines abnormal movement states, and generates behavioral warning information to alert obstetric patients to potential risks of turning over, falling, or unplanned behavior. This behavioral warning information indicates potential abnormal activity risks in obstetric patients, including excessive agitation, abnormal turning over, involuntary vigorous limb movements, and potential unplanned contact with medical tubing, supporting subsequent risk assessment, tiered treatment, and warning response.

[0043] Edge computing module: Using TinyML, multimodal perception data is fused and inferred at the bedside to output a three-level behavioral level for obstetric patients, where: Level 1: Involuntary tremor behavior, manifested as micro-vibrations with a frequency higher than 5 Hz, usually corresponding to physiological tremors or slight involuntary movements; Level 2: Purposeful movement behavior, manifested as sustained force greater than 15 N on one or both limbs, usually corresponding to attempts to turn over, raise hands, get up, or touch bedside instruments; Level 3: Dangerous behavior, manifested as hand movement trajectories less than 10 cm away from infusion lines, monitoring lines, or bedside medical devices, with a high risk of falls or unplanned extubation;

[0044] The comprehensive judgment module combines abnormal posture judgment information, behavior warning information, and three-level behavior level to make a comprehensive judgment, generate a comprehensive warning result, and output it to the execution end to trigger prompts, records, and constrain execution actions, thereby improving the accuracy and reliability of behavior monitoring and risk warning while ensuring the safety of obstetric patients.

[0045] Example 2, according to Figure 2 This embodiment is based on Embodiment 1. In this embodiment, the data analysis and processing module processes the patient image data through a masked keypoint-continuous-time attention fusion network, outputs the coordinate information of 17 key points, and generates posture abnormality judgment information. The specific steps include:

[0046] Step S1: Image feature extraction: Input the patient image data into the CNN, and obtain the patient environment depth feature map through multi-layer convolution, pooling and non-linear activation operations; the patient environment depth feature map includes patient body information as well as the surrounding environment information of bedside equipment, restraint straps and tubes, which is used to support subsequent human body region localization and pose analysis.

[0047] Step S2: Candidate Region Generation: Input the patient environment depth feature map into the RPN, and use the anchor box generation mechanism, sliding window search and non-maximum suppression (NMS) algorithm to filter and obtain a set of candidate regions for body parts;

[0048] Step S3: Region Scale Alignment: Perform spatial scale alignment on the set of candidate regions for body parts to obtain a set of candidate regions of uniform size;

[0049] Step S4: Masking and Keypoint Modeling: Input a set of candidate regions of uniform size into the masking segmentation branch and the keypoint detection branch; the masking segmentation branch performs pixel-level segmentation on the candidate regions and outputs a pixel-level segmentation mask of the patient's body, which is used to extract the global contour change features of the patient's body. The global contour change features include the overall contour shape of the patient, the spatial distribution of the contour, and the information on the change of the contour over time, thereby achieving accurate positioning of the patient's body; the keypoint detection branch performs keypoint regression on the candidate regions and outputs the coordinate information of 17 key points corresponding to the patient's body, and constructs the kinematic features of the skeleton structure based on the coordinate information of the 17 key points, which are used to characterize the spatial positional relationship between the joints of the obstetric patient's limbs and their movement changes over time;

[0050] Step S5: Continuous Temporal Pose Analysis: Based on the global contour change features and the kinematic features of the skeleton structure, a continuous frame temporal feature sequence is constructed in chronological order. The continuous frame temporal feature sequence is processed using a continuous temporal attention mechanism inspired by neural circuits. The pose features at different times are modeled with continuous temporal dependence and dynamically weighted to obtain a continuous temporal context feature representation that reflects the trend of pose change. Based on the continuous temporal context feature representation, the temporal consistency of the patient's pose changes, motion accumulation features, and abnormal evolution trends are comprehensively analyzed. When the analysis results show that there is a continuous or gradually increasing abnormal movement trend, pose abnormality judgment information is generated.

[0051] Example 3, this example is based on Example 2. In this example, step S5 specifically includes the following:

[0052] Step S51: Map the continuous frame temporal feature sequence into query feature vector, key feature vector, and value feature vector; for each query feature, calculate its correlation measure with each key feature within the time window, and based on the correlation measure, use the Top-K sparse filtering strategy to select the K key features with the highest correlation for each query feature to obtain the filtered key features; concatenate the filtered key features with the corresponding query features to form sparse query-key joint features;

[0053] Step S52: Construct a neuron-circuit-inspired dual-gating mechanism. The sparse query-key joint features are input into this mechanism. Based on a neuron-circuit-inspired sparse connection structure and grouping modulation strategy, the mechanism performs nonlinear mapping on the sparse query-key joint features, generating content-target gating parameters and time-constant gating parameters for continuous-time attention modeling. The content-target gating parameter characterizes the driving strength of obstetric patient posture features in the abnormality detection process, indicating the contribution of the current posture features to potential abnormal behavior. The time-constant gating parameter characterizes the accumulation rate and decay characteristics of posture features over time. This allows for the regulation of the evolution speed and stability of posture anomaly information over continuous time scales. The construction method of the neuron circuit-inspired dual-gating mechanism is as follows: Based on the gated neural network, a sparse connection topology and group modulation strategy inspired by neuron circuits are introduced to structurally improve the modeling method of the gated neural network that only statically adjusts the feature weights. On the basis of the structural improvement, dual-gating branches are constructed for generating content target gating parameters and time constant gating parameters, respectively. This enables the gating mechanism to not only regulate the expression intensity of posture features but also control the accumulation rate and decay characteristics of posture features over continuous time, thus constructing the neuron circuit-inspired dual-gating mechanism.

[0054] Step S53: Treat the attention state of each query-key pair corresponding to the sparse query-key joint features as a state variable that evolves continuously over time. Introduce a first-order continuous-time dynamic model to model the change process of the attention state over time. Through content target gating parameters and time constant gating parameters, make the attention state exhibit dynamic evolution characteristics that are jointly modulated by target-driven and time decay in the time dimension. Construct a continuous-time attention state evolution model to characterize the cumulative change trend of obstetric patients' postures on a continuous time scale.

[0055] The formula used in the continuous-time attention state evolution model:

[0056] ;

[0057] in, Represents a continuous-time variable. Indicates time Attention state at all times Indicates attention state Instantaneous rate of change over time; The input joint features corresponding to this attention state are obtained from the "sparse query-key joint features" constructed in step S51; This indicates that the gating is related to the "time constant / time scale"; Represents the time constant gating parameter. Indicates joint features from input The time constant gating parameters generated by the driver; Indicates joint features from input The content target gating parameters generated by the driver;

[0058] Step S54: Preset a pseudo-time interval, and numerically solve the continuous-time attention state evolution model based on the pseudo-time interval to obtain the attention state value corresponding to each time step; normalize the attention state value to generate continuous-time attention weights, which are used to represent the relative importance of pose features in different time frames in the overall pose evolution process.

[0059] Formula for steady-state solution of continuous-time attention:

[0060] , ;

[0061] in, This represents the initial attention state value. This represents the steady-state attention value. This represents the exponential decay factor, used to describe the initial deviation. The process of gradually disappearing as pseudo-time progresses;

[0062] Step S55: The posture features of the corresponding time step are weighted and fused according to the continuous-time attention weight to obtain the fused continuous-time posture context feature representation. The weighted fusion process comprehensively considers the cumulative contribution of posture features in the time dimension, rather than relying solely on instantaneous posture changes. Based on the continuous-time posture context feature representation, the temporal consistency, motion accumulation characteristics, and abnormal evolution trend of obstetric patients' posture changes are comprehensively analyzed. When the analysis results meet the preset abnormal judgment conditions, posture abnormality judgment information is generated to support obstetric patients' behavior monitoring and risk warning.

[0063] Example 4, based on Example 3, describes the construction of the multi-task joint loss function as follows: A classification loss function, a bounding box regression loss function, and a mask loss function are introduced, and these three loss functions are integrated to construct the multi-task joint loss function. The formula used is as follows:

[0064] Classification loss function:

[0065] ;

[0066] in, Category labels representing real objects, The model represents the categories of real objects. The predicted probability; Minimize the probability of the negative logarithm; This represents the predicted probability distribution vector output by the model; Represents the classification loss function;

[0067] Bounding box regression loss function:

[0068] ;

[0069] in, Represents the true bounding box regression parameters. This represents the regression parameters of the predicted bounding box. This represents the bounding box regression loss function; These represent the coordinates of the center point. Center point coordinates ,Width ,high ; Indicates the index of the four parametric components of the bounding box; Indicates the category predicted by the model. The corresponding bounding box is in the 1st position. The values ​​of each parameter component (regression results) The true bounding box represents the first position. The values ​​of each parameter component; This represents the smoothed L1 loss function;

[0070] Mask loss function:

[0071] ;

[0072] in, This represents the mask loss function. Indicates the mask resolution. It is the total number of pixels in the mask; Represents the row and column coordinates of a pixel; Indicates the actual mask in pixels The value of the position, Indicates category Predicted mask at pixels The predicted probability of location;

[0073] Multi-task joint loss function formula:

[0074] ;

[0075] in, This represents the joint loss function for multiple tasks.

[0076] Example 5, based on Example 4, in this example, the process of generating behavior warning information by the behavior warning module specifically includes: calculating the spatial displacement of key points in each of the six human body regions using the Euclidean distance calculation formula, obtaining the change in displacement distance and duration of key points; setting a movement distance threshold and a duration threshold, when it is detected that the change in displacement distance of all key points in any human body region is greater than the movement distance threshold and the duration of key points exceeds the duration threshold, it is determined that the human body region is in an abnormal movement state, and behavior warning information is generated.

[0077] Example 6, based on Example 4, in this example, the process of generating behavior warning information by the behavior warning module specifically includes: calculating the spatial displacement of key points in each of the six human body regions using the Chebyshev distance calculation formula, obtaining the change in displacement distance and duration of key points; setting a movement distance threshold and a duration threshold, when it is detected that the change in displacement distance of all key points in any human body region is greater than the movement distance threshold and the duration of key points exceeds the duration threshold, it is determined that the human body region is in an abnormal movement state, and behavior warning information is generated.

[0078] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A behavioral monitoring and risk alert system for obstetric patients, characterized in that: The system comprises; Data perception acquisition module, acquiring patient image data; forming multi-modal perception data; Data analysis processing module, establishing a mask key point-continuous time attention fusion network, processing patient image data through the mask key point-continuous time attention fusion network; in the processing process, a multi-task joint loss function is introduced to jointly optimize the model parameters of the mask key point-continuous time attention fusion network, so that the network can complete the pixel-level segmentation and key point regression tasks at the same time, output 17 key point coordinate information, and generate posture abnormality judgment information; Human region modeling module, defining a semantic rule corresponding table, grouping and region mapping modeling processing 17 key point coordinate information according to the corresponding relationship of the semantic rule corresponding table, and mapping and dividing 6 human body regions; Behavior warning module: calculate the spatial displacement of the key points in each human body region in the 6 human body regions through the Chebyshev distance calculation formula, determine the abnormal motion state, and generate behavior warning information; Edge computing module: using TinyML to fuse and infer multi-modal perception data, outputting three-level behavior grades; Comprehensive judgment module, combining posture abnormality judgment information, behavior warning information and three-level behavior grades for comprehensive judgment, generating comprehensive warning results, triggering prompt, recording and constraint execution actions.

2. A behavior monitoring and risk alert system for obstetric patients as claimed in claim 1 wherein: The mask key point-continuous time attention fusion network comprises CNN, RPN, mask segmentation branch, key point detection branch and continuous time attention mechanism based on neuron circuit inspiration.

3. A behavior monitoring and risk alerting system for obstetric patients as claimed in claim 2 wherein: The data analysis processing module processes the patient image data through the mask key point-continuous time attention fusion network, outputs 17 key point coordinate information, and generates posture abnormality judgment information, which specifically includes the following steps: Step S1: input the patient image data into the CNN to obtain patient environment depth feature maps; Step S2: input the patient environment depth feature maps into the RPN, use the anchor box generation mechanism, sliding window search and non-maximum suppression algorithm to screen to obtain a body part candidate region set; Step S3: perform spatial scale alignment operation on the body part candidate region set to obtain a uniform size candidate region set; Step S4: input the uniform size candidate region set into the mask segmentation branch and the key point detection branch; the mask segmentation branch performs pixel-level segmentation on the candidate region to extract global contour change features; the key point detection branch performs key point regression on the candidate region to output 17 key point coordinate information, and constructs skeleton structure kinematics features based on the 17 key point coordinate information; Step S5: according to the global contour change features and the skeleton structure kinematics features, construct a continuous frame time sequence feature sequence in time sequence, process the continuous frame time sequence feature sequence by using the continuous time attention mechanism based on neuron circuit inspiration, and generate posture abnormality judgment information.

4. A behavior monitoring and risk alerting system for obstetric patients as claimed in claim 3 wherein: Step S5, specifically includes the following contents: Step S51: Map the sequence of continuous frame timing features into a query feature vector, a key feature vector, and a value feature vector, calculate a correlation measure for each query feature, and based on the correlation measure result, use a Top-K sparse screening strategy to obtain screened key features; Splice the screened key features with the corresponding query features to form a sparse query-key joint feature; Step S52: Construct a neuron circuit inspired double-gated mechanism, input the sparse query-key joint feature into the neuron circuit inspired double-gated mechanism, and based on the neuron circuit inspired sparse connection structure and grouping modulation strategy, perform nonlinear mapping processing on the sparse query-key joint feature to generate content target gating parameters and time constant gating parameters, respectively; Step S53: Treat the attention state of each query-key pair corresponding to the sparse query-key joint feature as a state variable evolving in continuous time, introduce a first-order continuous time dynamic model to model the change process of the attention state over time; Through the content target gating parameters and the time constant gating parameters, the attention state presents a dynamic evolution characteristic in the time dimension that is jointly modulated by target driving and time decay, and a continuous time attention state evolution model is constructed; Step S54: Predefine a pseudo-time interval, numerically solve the continuous time attention state evolution model based on the pseudo-time interval to obtain the attention state value, and normalize the attention state value to generate a continuous time attention weight; Step S55: Perform weighted fusion according to the continuous time attention weight to obtain a continuous time posture context feature representation; Based on the continuous time posture context feature representation, the time consistency, motion accumulation characteristics, and abnormal evolution trend of the obstetric patient posture change are comprehensively analyzed to generate posture abnormality determination information.

5. A behavior monitoring and risk alerting system for obstetric patients as claimed in claim 4 wherein: The construction method of the neuron circuit inspired double-gated mechanism is: based on the gated neural network, by introducing the neuron circuit inspired sparse connection topology and grouping modulation strategy, the modeling method of the gated structure in the gated neural network which only statically adjusts the feature weight is structurally improved; based on the structural improvement, a double-gated branch for generating content target gating parameters and time constant gating parameters is constructed to obtain the neuron circuit inspired double-gated mechanism.

6. A behavior monitoring and risk alerting system for obstetric patients as claimed in claim 1, wherein: The construction method of the multi-task joint loss function is: introduce a classification loss function, a bounding box regression loss function, and a mask loss function, and integrate the three types of loss functions to construct a multi-task joint loss function.

Images