Multi-dimensional feature fused prospective risk early warning and response system for tooth trauma of children

Abstract

The invention relates to the technical field of medical health monitoring, and discloses a multi-dimensional feature fusion child tooth trauma prospective risk early warning and response system, which comprises a data acquisition module for acquiring multi-source heterogeneous data such as child behaviors, positions, environments and health records; the behavior and scene recognition module is used for accurately recognizing instant behaviors and scenes of children by adopting a deep learning model; the core of the risk prediction and scoring module is to construct an age segmentation sub-model, perform multi-factor dynamic risk calculation and predict a time window trend; the nursing suggestion generating and pushing module is used for automatically pushing targeted suggestions to multiple terminals when the risk score exceeds a threshold value; and the data closed-loop and model updating module is used for safely and continuously optimizing the model by utilizing multi-node feedback data under a federated learning framework. According to the method, the age segmentation sub-model and the space-time dynamic association model are constructed, and the federal learning framework is combined, so that accurate, dynamic and highly personalized prospective early warning of the child tooth trauma risk is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical health monitoring, in particular to a multi-dimensional feature fusion child dental trauma prospective risk early warning and response system. BACKGROUND

[0002] Dental trauma is one of the common oral emergencies in children and adolescents, especially in preschool and school-age children. Because children of this age are lively and active, their balance and coordination have not yet fully developed, and they lack awareness of danger, so they are prone to dental damage due to accidents such as falling, colliding, etc. during daily play, running and sports activities. Dental trauma not only causes immediate pain and dysfunction in children, but also can cause long-term complications such as pulp necrosis, root resorption, and abnormal tooth replacement, which has a long-term negative impact on children's oral health, maxillofacial development, and even mental health, accompanied by a complex treatment process and significant economic burden. Therefore, developing a technical solution that can effectively prevent or reduce the risk of dental trauma in children has important social value and practical significance.

[0003] In order to protect the safety of children, existing technologies mainly intervene through two types of solutions. One is the general safety monitoring technology based on wearable devices, such as the commonly seen smart watches for children or anti-lost locators on the market. These devices usually integrate acceleration sensors and global positioning system (GPS) modules, and their core function is to achieve location tracking, set up a safe area (electronic fence), and send an alarm to the guardian when a severe impact (usually interpreted as a fall) is detected. The other is environmental safety protection measures, mainly in places where children frequently move around, such as kindergartens, playgrounds, etc., laying soft ground mats, installing anti-collision strips at the sharp corners of tables and chairs, etc., to reduce the severity of injuries by modifying the physical environment.

[0004] Although existing technologies have improved the level of child safety monitoring and the passive protection capability of the environment to some extent, there are still some deficiencies: the existing technology based on wearable devices has significant lag and non-specificity in preventing dental trauma. The fundamental reason is that the algorithm model used by these devices is essentially event-driven rather than risk prediction. They can only trigger an alarm at the moment or after a high-impact event such as a fall or collision has occurred, at which point the injury may have already been caused, and they cannot achieve true prevention beforehand. In addition, their risk assessment model is too simplified, usually relying only on a single acceleration threshold, and cannot distinguish between benign large-scale movements (such as jumping) and specific dangerous behaviors with high dental trauma risk (such as face-down forward falls), resulting in a large number of invalid alarms or key risk omissions.

[0005] Moreover, the prior art lacks comprehensive and personalized consideration of risk factors. The occurrence of dental trauma is a complex process involving individual factors, behavioral factors and environmental factors. The reason why the prior art cannot provide accurate risk assessment is that its technical architecture cannot effectively integrate these three types of key information. For example, it cannot dynamically correlate real-time behavior of a child (such as running at the stairwell) with the child's specific physiological condition (such as the incisors being in the deciduous stage or having a protruding front tooth) and the specific risks of the environment (such as a hard floor). This fragmentation of data and lack of models in the prior art solutions can only stay at a general and low-dimensional level of safety monitoring, and cannot provide precise and forward-looking guidance and intervention for the prevention of dental trauma, which is a specific and complex type of injury. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a multi-dimensional feature fusion child dental trauma forward-looking risk early warning and response system, which solves the problem that the prior art cannot integrate individual, behavioral and environmental multi-dimensional data, and thus cannot accurately and forwardly predict the risk of dental trauma.

[0007] To achieve the above-mentioned purpose, the present application is implemented by the following technical solutions: The multi-dimensional feature fusion child dental trauma forward-looking risk early warning and response system comprises a data acquisition module, a behavior and scene recognition module, a risk prediction and scoring module, a nursing suggestion generation and pushing module, and a data closed loop and model updating module.

[0008] The data acquisition module is configured to acquire multi-source raw data of a child. In a specific embodiment, the module comprises: A dynamic behavior data acquisition unit, which is implemented by a wearable terminal integrated with a three-axis accelerometer, a gyroscope and an inertial measurement unit (IMU). The terminal is configured to acquire time-series dynamic behavior parameters of a child in different activity states, including motion posture, acceleration, angular velocity and displacement path.

[0009] An environmental perception data acquisition unit, which is configured to acquire physical environmental characteristics of an activity space in which a child is located, including illumination, air pressure fluctuation, background sound, temperature change and contact impact, by means of an ambient light sensor, an air pressure sensor, a sound sensor, a temperature and humidity sensor and a collision sensing element arranged on or around the terminal.

[0010] A position data acquisition unit, which employs a multi-mode fusion positioning technology that combines global positioning system (GPS), wireless local area network (Wi-Fi) positioning and Bluetooth low energy (BLE) beacon technology. Through trajectory recognition and path matching algorithms, the unit realizes real-time acquisition of a child's activity trajectory, path modeling and high-risk area marking.

[0011] an individual health information input interface, which imports structured health information data of the child individual through user input or in a manner synchronized with a medical information system, and the data specifically includes: age and developmental stage; tooth replacement status, tooth mobility level; presence or absence of oral structure features such as early loss of deciduous teeth, difficulty in eruption of permanent teeth, sparse or dense dentition, etc.; past dental trauma records, including trauma time, damaged part, severity and treatment method; and behavior characteristic labels reported by guardians.

[0012] The data acquisition module can fuse the data from different sources mentioned above and unify them into a structured data template, and support data desensitization processing and time synchronization.

[0013] The behavior and scene recognition module is configured to receive and process the multi-source raw data acquired by the data acquisition module. In a specific embodiment, the module adopts a multi-modal deep learning structure, and the specific implementation is as follows: a behavior recognition submodule, which adopts a deep learning model based on the combination of a convolutional neural network (CNN) and a long short-term memory network (LSTM). This model models the time series data such as acceleration, angular velocity, and attitude angle provided by an inertial measurement unit (IMU) to extract a behavior feature vector, thereby realizing the identification and classification of high-dynamic actions related to the risk of dental trauma such as running, sprinting, repeated jumping, sudden falling, collision, and sliding.

[0014] a scene recognition submodule, which is based on an image recognition model (such as MobileNet or ResNet) and an audio convolution analysis network, and combines the location information acquired by the location data acquisition unit to perform multi-modal recognition and scene classification on the activity space currently occupied by the child. The identifiable scene types include the table corner area in the indoor environment of the family, the stairway in the school environment, and the slide exit and cement ground in the outdoor activity area.

[0015] a behavior-scene cross-fusion modeling unit, which cross-couples the behavior types identified by the behavior recognition submodule and the scene types identified by the scene recognition submodule, and constructs a behavior-scene risk matrix. This matrix assigns weights to specific combinations, for example, the combination of running and stairway is assigned a higher risk weight. This matrix serves as the input of the risk prediction and scoring module.

[0016] The risk prediction and scoring module is configured to construct a multi-factor risk prediction model by integrating the identification results of the behavior and scene recognition module and the data imported by the individual health information input interface. In a specific embodiment, the module includes: The risk scoring model construction unit takes the current behavior recognition result and its corresponding impact magnitude, the risk coefficient of the current activity scenario, the individual's oral development status, and the frequency of historical dental trauma records as input factors, and calculates the dental trauma risk score at the current moment through a weighted scoring function. The specific form of this function is: ; In the formula, For at any time The calculated risk score for dental trauma; The current moment; The activation function is used to map the weighted sum to the interval [0,1], making it a normalized score; The total number of risk factors input into the model; This serves as an index for risk factors, with values ​​ranging from 1 to... ; For the first The adjustable weight coefficients of each risk factor are dynamically adjusted by a personalized risk weight adjustment unit. For the first Each risk factor at time The unit quantifies the risk score by combining it with a preset threshold, classifying the risk score into four risk levels: low, medium, high, and extremely high.

[0017] A time window trend prediction mechanism, used to predict a future time window. The mechanism tracks the historical sequence of risk scores, analyzing their changing trends. In one specific embodiment, the mechanism employs a gated recurrent unit (GRU) or a long short-term memory network (LSTM) to analyze the historical sequence of risk scores. Modeling to predict future moments Risk score : ; In the formula, For future moments Predicted values ​​of dental trauma risk scores; For a trained gated recurrent unit neural network prediction model; The historical risk score time series is input into the prediction model; The current moment; The predicted time step, its value within a preset time window. Within the specified range. The mechanism's output includes the risk score's upward slope, the predicted peak time point, and a threshold exceeding warning indication.

[0018] A personalized risk weight adjustment unit adjusts the weights in the weighted scoring function. Dynamic adjustments will be made. These adjustments will be based on factors including the child's age group, activity level assessments derived from analyzing historical behavioral data, and the presence of oral structural vulnerabilities such as protruding anterior teeth.

[0019] The nursing suggestion generation and push module is configured to perform an operation when the dental trauma risk score exceeds a preset threshold. In one specific embodiment, the module includes: The suggestion generation unit automatically invokes an intervention suggestion library stored locally or in the cloud when it detects that the risk score has reached or exceeded a preset threshold. This unit combines the currently identified high-risk behavior types and activity scenarios, the child's mixed dentition stage, and historical dental trauma events to generate personalized care suggestions, such as wearing a mouthguard, suspending strenuous activities, or staying away from dangerous areas.

[0020] The multi-terminal push unit transmits generated nursing advice to various user terminals in real time, including parents' mobile applications, children's smartwatches, and the school's management platform. This unit is equipped with a risk level response mapping mechanism, dynamically matching the alert format according to the risk level: medium risk triggers a light vibration alert, high risk triggers a voice broadcast, and extremely high risk triggers a multi-channel strong alert and can be linked to campus terminal broadcasts.

[0021] The visual feedback module presents care suggestions in the form of images, animations, or cartoon voices, and allows children to interact with feedback via buttons.

[0022] The data closure and model update module is configured to optimize the system model after collecting feedback data on dental trauma events. In a specific embodiment, this module includes: The dental trauma event recording interface is used by users to upload event feedback data after a dental trauma event actually occurs. The data includes the timestamp of the event, location, trauma site, cause of trauma, and accompanying photos or diagnostic reports.

[0023] The model error analysis unit analyzes missed (damage occurred without warning) or false (warning occurred but no damage occurred) samples that appear during system operation. By tracing back the data window before the event occurred, the error type is classified and attributed, thereby constructing a blind zone behavior sample library containing samples with low recognition rates.

[0024] The model optimization module continuously optimizes the multi-factor risk prediction model based on historical operating data and user feedback data. It employs a few-shot fine-tuning mechanism, treating error samples as high-value samples to adjust local weights in the model. This module also features a personalized model fine-tuning mechanism. When a specific child has sufficient historical data or belongs to a highly sensitive population, a unique local model copy is generated for that child. Based on the child's behavioral habits and response performance, the risk score threshold and risk factor weights of this local model copy are periodically updated.

[0025] Furthermore, the system described in this invention supports a cloud-edge collaborative deployment method. In this deployment method, a lightweight model version is deployed on edge devices such as wearable terminals to perform real-time behavior recognition and basic scoring tasks, reducing response latency; a complete intelligent algorithm model is deployed on a cloud server for centralized management, complex model training, and cross-user learning; through parameter synchronization and hierarchical model deployment mechanisms, a collaborative working mode is achieved where cloud training optimization and terminals periodically obtain simplified model copies, ensuring data processing efficiency and data privacy compliance.

[0026] This invention provides a multi-dimensional feature fusion-based proactive risk warning and response system for pediatric dental trauma. It offers the following advantages: 1. This invention comprehensively integrates multi-dimensional characteristics such as children's dynamic behavior, surrounding environment, and individual oral health status through a data acquisition module, constructing a three-dimensional risk profile. More importantly, through a time window trend prediction mechanism (such as a GRU neural network) within the risk prediction and scoring module, the system can not only assess the risk at the current moment but also predict the risk change trend in the near future, thereby achieving early risk prediction and intervention and overcoming the limitation of traditional monitoring methods that can only respond after danger occurs.

[0027] 2. This invention abandons the one-size-fits-all early warning model. First, by constructing age-segmented sub-models, the system can be specifically optimized for the typical behaviors and risk patterns of children in different age groups (such as 3-6 years old and 7-12 years old). Second, the core spatiotemporal behavior and individual dynamic correlation model enables risk assessment to have context-aware capabilities. For example, when a child enters the stairwell from the playground, the risk weight of running behavior can be dynamically increased; when a face-down fall is identified and combined with the individual characteristic of the child having loose primary teeth, the risk coefficient can be dynamically amplified. This dynamic correlation mechanism makes the early warning more in line with real-world scenarios, effectively reducing false alarms and missed alarms.

[0028] 3. This invention, by introducing a federated learning framework, allows multiple institutions (such as different kindergartens and hospitals) to collaboratively train a global model with stronger generalization capabilities without sharing original sensitive data, effectively solving the data silo problem. Simultaneously, by utilizing group feature transfer technology, the knowledge of the global model can be used to quickly initialize personalized models for new users, solving the problem of model optimization lag caused by insufficient individual data. Furthermore, the active learning and edge case mining units can proactively screen the most uncertain critical samples of the model and request manual annotation, using limited supervision resources effectively, thereby accelerating model iteration and optimization at a lower cost and higher efficiency. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the data acquisition module of the present invention; Figure 4 This is a schematic diagram of the behavior and scene recognition module of the present invention; Figure 5 This is a schematic diagram of the risk prediction and scoring module of the present invention; Figure 6 This is a schematic diagram of the nursing suggestion generation and push module of the present invention. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see the appendix Figure 1 , Figure 1 This is a schematic diagram of a system structure according to an embodiment of the present invention. The present invention provides a multi-dimensional feature fusion-based proactive risk warning and response system for pediatric dental trauma, which may include: Data acquisition module 100 is used to acquire multi-source raw data of children; The behavior and scene recognition module 200 is used to receive and process multi-source raw data to identify the child's current behavioral state and the activity scene in which the child is active; The risk prediction and scoring module 300 is used to integrate the identification results with individual health information, calculate a quantitative risk score, and predict risk trends. The nursing suggestion generation and push module 400 is used to generate and push targeted nursing suggestions when the risk score exceeds a preset threshold. The data closure and model update module 500 is used to optimize the parameters of the prediction model based on event feedback data.

[0032] Please refer to the appendix. Figure 2 , Figure 2 This is a schematic flowchart of a method according to an embodiment of the present invention. The present invention provides a method for predicting and responding to dental trauma behavior in children, which may include the following steps: Step S101: Perform data collection to obtain multi-source raw data of the child, including dynamic behavior data, location information, environmental parameter data, individual oral health information and historical dental trauma records.

[0033] Step S102: Perform behavior and scene recognition, process the collected multi-source raw data, and identify the child's current behavioral state and activity scene based on a deep learning model.

[0034] Step S103: Perform risk prediction and scoring. Combine the behavioral scene recognition results with individual health information to construct and apply a multi-factor risk prediction model, calculate a quantified dental trauma risk score, and predict the trend of the score within a preset time window.

[0035] Step S104: Generate and push nursing suggestions. Determine whether the risk score exceeds a preset threshold. If it does, automatically generate and push targeted nursing suggestions to one or more terminals.

[0036] Step S105: Perform data closure and model update. After collecting feedback data on dental trauma events, analyze false alarm and missed alarm samples, and optimize the parameters of the multi-factor risk prediction model based on the analysis results.

[0037] In one specific embodiment, the overall operation of the system begins with the data acquisition module 100. This data acquisition module 100 continuously collects dynamic behavioral data, environmental parameters, and location information through hardware units deployed in the child's wearable device or its activity environment, and receives individual health information input by the user through a software interface to form a multi-dimensional raw database, thus completing step S101.

[0038] Next, the behavior and scene recognition module 200 receives the data stream from the data acquisition module 100. The deep learning model within this module analyzes the time-series dynamic behavior data to classify specific actions such as running, jumping, and sudden falls. Simultaneously, the module combines location data and environmental perception data to classify the physical space in which the child is located. This process executes step S102, ultimately outputting structured behavior and scene labels.

[0039] Subsequently, the risk prediction and scoring module 300 receives the tags output by the behavior and scene recognition module 200 and integrates them with the individual health data of the child obtained from the data acquisition module 100. This risk prediction and scoring module 300 applies a multi-factor risk prediction model to quantify and weight the input factors from multiple dimensions to calculate the quantified risk score value at the current moment. Simultaneously, based on historical risk score sequences, the risk prediction and scoring module 300 also predicts the risk change trend within a future time window, completing step S103.

[0040] When the current risk score or predicted risk score calculated by the risk prediction and scoring module 300 exceeds one or more preset thresholds, the nursing suggestion generation and push module 400 is triggered, executing step S104. Based on the current risk level, behavior type, and scenario information, the nursing suggestion generation and push module 400 retrieves and generates one or more nursing suggestions from a preset intervention suggestion library, and pushes the suggestions to designated terminal devices via network communication protocols.

[0041] Finally, the system performs step S105 through the data closure and model update module 500 to achieve self-optimization. When a dental trauma event actually occurs, the caregiver can upload relevant information about the event through the event recording interface. The model update module 500 compares this feedback with the system's own early warning records. If a missed report is found, the data before the event is marked as a high-value sample, which is used to fine-tune and update the model parameters in the risk prediction and scoring module 300. This process constitutes the system's closed-loop learning and iterative optimization path.

[0042] Please see the appendix Figure 3 , Figure 3 This is a schematic diagram of the structure of a data acquisition module 100 according to an embodiment of the present invention. The data acquisition module 100 serves as the perception foundation of the system and is configured to acquire raw data related to the risk of dental trauma in children from multiple dimensions. In a specific embodiment, the data acquisition module 100 includes a dynamic behavior data acquisition unit 110, an environmental perception data acquisition unit 120, a location data acquisition unit 130, and an individual health information input interface 140.

[0043] The dynamic behavior data acquisition unit 110 is physically housed as an inertial measurement unit (IMU) integrated into a child-wearable device (such as a smartwatch or a dedicated sensor pendant). This IMU integrates a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer. The three-axis accelerometer measures the linear acceleration of the device along the three orthogonal axes (X, Y, and Z) to reflect the child's translational motion and impacts. The three-axis gyroscope measures the angular velocity of the device around the three axes to reflect the child's rotation and attitude changes. The three-axis magnetometer senses the Earth's magnetic field to assist in attitude calculation and eliminate accumulated errors. The unit 110 continuously acquires raw sensor data at a preset sampling frequency (e.g., 50Hz or 100Hz) and outputs a sequence of dynamic behavior parameters, including acceleration vectors, angular velocity vectors, and attitude angles (such as roll, pitch, and yaw), through a built-in attitude calculation algorithm (e.g., complementary filtering or Kalman filtering).

[0044] The environmental perception data acquisition unit 120 includes a series of sensor elements deployed on or around the wearable device. Specifically, it includes an ambient light sensor for acquiring ambient light illuminance values ​​to help determine whether the child is outdoors in bright light or indoors or in a dimly lit hallway; a barometric pressure sensor for measuring atmospheric pressure, the rapid change of which can indicate that the child is undergoing rapid vertical displacement (such as climbing stairs) or experiencing strenuous activity; and one or more impact sensing elements, such as piezoelectric ceramic sensors, which are positioned at specific locations on the wearable device's casing. When the device is subjected to a direct impact, this element generates a voltage signal proportional to the impact force, thereby enabling direct quantitative detection of the impact event.

[0045] Location data acquisition unit 130 employs a fusion positioning technology to acquire a child's geographic location information in different environments. In open outdoor areas, unit 130 prioritizes using a Global Positioning System (GPS) module to obtain latitude and longitude coordinates. When a child enters indoor areas or areas where GPS signals are blocked, the system automatically switches to Wi-Fi positioning mode. It estimates the current location by scanning the Signal Strength Indication (RSSI) values ​​of surrounding Wi-Fi access points (APs) and matching them with a pre-built Wi-Fi fingerprint database. To achieve higher accuracy in specific high-risk areas (such as slide exits, stairwells, and hard surfaces on the playground), Bluetooth Low Energy (BLE) beacons are pre-deployed in these areas. When a child enters the signal coverage area of ​​a BLE beacon, location data acquisition unit 130 receives and analyzes the BLE broadcast signal to achieve centimeter-level near-field positioning and area identification.

[0046] The Individual Health Information Input Interface 140 is a software interface used to receive and store structured individual health information of children, input by guardians through a companion mobile application (APP) or imported through integration with the Hospital Information System (HIS). This information is organized into an electronic profile, with data fields including: the child's age and specific developmental stage; oral health status, specifically the status of mixed dentition (e.g., deciduous / permanent / mixed dentition), the presence and grade of specific tooth mobility, and the presence of malocclusions such as anterior crossbite or anterior protrusion; and records of past dental trauma, including the time of each trauma, the location of the damaged tooth, the type of trauma (e.g., tooth concussion, dislocation, fracture), and the treatment method. This static or semi-static individual information provides foundational data support for subsequent personalized risk assessments.

[0047] Please see the appendix Figure 4 , Figure 4 This is a schematic diagram of the behavior and scene recognition module 200 according to an embodiment of the present invention. This module 200 is the core hub connecting raw data and risk assessment; it is not only responsible for analyzing the child's current state, but also introduces a forward-looking model optimization mechanism.

[0048] In a preferred embodiment, the module 200 includes a behavior recognition submodule 210, a scene recognition submodule 220, a behavior-scene cross-fusion modeling unit 230, and an active learning and edge case mining unit 240.

[0049] The behavior recognition submodule 210 employs a deep learning model combining a convolutional neural network (CNN) and a long short-term memory network (LSTM). This model is trained to process high-dimensional temporal data output by the dynamic behavior data acquisition unit 110. Through one-dimensional convolutional layers, the model can automatically extract local temporal features from the motion data, such as the frequency of steps and the peak value of impacts, as well as other micro-movement patterns.

[0050] Subsequently, the Long Short-Term Memory (LSTM) network layer models the temporal dependencies of these feature sequences, thereby accurately learning and recognizing highly dynamic compound actions with significant risks of dental trauma, such as running, jumping, sudden falls, and violent collisions. While outputting the final classification result, this submodule 210 also outputs a confidence score vector representing the reliability of the classification result through a Softmax layer.

[0051] The scene recognition submodule 220 employs a multimodal data fusion strategy to achieve accurate environmental perception. It not only utilizes geographic coordinate information from the location data acquisition unit 130 for coarse-grained scene judgment (such as an outdoor playground or an indoor building), but also further integrates keyframe images from the wearable device's camera and audio signals from the sound sensor.

[0052] For example, by analyzing textures and objects in images (such as slides and stair steps) and combining them with features in audio signals (such as hard ground echoes and crowd noise), the system can accurately classify specific micro-scenes directly related to the risk of dental trauma, such as slide exit areas, stairwells, and hard surfaces on playgrounds.

[0053] The behavior-scenario cross-integration modeling unit 230 serves as a preprocessing step for risk quantification. This unit internally constructs and stores a multi-dimensional, dynamically updatable behavior-scenario risk matrix. The matrix's index includes not only behavior and scenario categories but can also incorporate other dimensions, such as time (e.g., break times). Each cell stores a baseline risk weight value derived from expert knowledge and historical data statistics.

[0054] Upon receiving behavior and scenario labels, the behavior-scenario cross-fusion modeling unit 230 outputs a risk coefficient value that integrates the coupling effect of the current specific behavior and the specific scenario through table lookup and interpolation calculation, providing key input for subsequent risk scoring.

[0055] The active learning and edge case mining unit 240 is a key innovation of this invention, enabling rapid and efficient model iteration. It constitutes a human-machine collaborative intelligent data annotation and model optimization triggering mechanism. This mechanism transforms the model's learning process from a passive mode that relies entirely on past injury events to an active mode that actively mines and learns from high-information samples. In a specific embodiment, the operating logic of unit 240 is as follows: Precisely defined trigger conditions: This unit continuously monitors two key indicators.

[0056] The first is the confidence score of behavior recognition. When the highest confidence score output by the behavior recognition submodule 210 is lower than a dynamically adjustable threshold (e.g., initially set to 0.7), it indicates that the model's judgment of the current behavior is ambiguous, such as the difficulty in clearly distinguishing between a child's fast walking and slow running at the feature level.

[0057] The second is the critical state of risk scoring. When the risk score calculated by the risk prediction and scoring module 300 falls into a narrow range around the preset risk level threshold (for example, the threshold for defining medium risk and high risk is 0.6, so the critical range can be set to [0.55, 0.65]), it indicates that the current scenario is at the boundary of the risk state, which is the area where the model is most likely to misjudge.

[0058] Context-aware annotation request: When any triggering condition is met, the unit 240 does not simply request annotation, but automatically generates a context-aware, structured annotation request. This request utilizes natural language processing templates, integrates currently identified behavior and scene information, and pushes it to the guardian's terminal as an easy-to-understand interactive question.

[0059] For example, the request could be: "The system has noticed a child moving quickly near the stairwell. Can you confirm whether the child is running or walking briskly?" This approach not only improves the accuracy of the annotations but also optimizes the user experience. To prevent user fatigue, the unit 240 also incorporates a request throttling mechanism, limiting the maximum number of requests per unit of time.

[0060] Generation and Prioritization of High-Value Samples: After the guardian makes a simple selection or confirmation through the terminal interface, the manually labeled result is returned. Unit 240 immediately binds this authoritative manually labeled result with the original data fragment that triggered the request (including IMU sequences, locations, image keyframes, etc.) to form a verified, high-value edge case sample. This sample, after being specially labeled, will be prioritized and sent to the training sample pool of the data closure and model update module 500.

[0061] During the next fine-tuning training of the model, these labeled edge case samples will be given higher sampling weights, thereby guiding the model to focus on learning and correcting its judgment ability in fuzzy regions and decision boundaries, accelerating the model's convergence speed and improving its recognition accuracy in critical risk scenarios.

[0062] Please see the appendix Figure 5 , Figure 5 This is a schematic diagram of the risk prediction and scoring module 300 according to an embodiment of the present invention. The risk scoring model construction unit 310 is based on an advanced algorithm framework using a spatiotemporal behavior and individual dynamic correlation model. This framework replaces the traditional static risk matrix weighting method and can dynamically assess risk based on real-time changing context. In a specific embodiment, the operating logic of unit 310 is as follows: First, unit 310 internally constructs and stores a set of risk scoring sub-models specific to children of different age groups. Upon receiving input data, unit 310 automatically selects the corresponding sub-model for risk calculation based on the child's age provided by the individual health information input interface 140. For example, the system can divide children into two main age groups: Sub-model for children aged 3-6: This sub-model is specifically configured to identify high-risk behavioral patterns in preschool children. Its algorithm and weight design focus on actions such as aimless running, loss of balance when climbing to heights, and falls, and increase the weight of risk factors related to indoor environments, low steps, table corners, and other scenarios.

[0063] Sub-model for children aged 7-12: This sub-model is optimized for the behavioral characteristics of school-aged children. Its algorithm focuses more on recognizing behaviors such as collisions in sports, falls while cycling, and collisions during vigorous play, and increases the weight of risk factors related to scenarios such as hard playgrounds, gymnasiums, and bike paths.

[0064] Secondly, within the selected sub-model, the model is capable of performing temporally dynamic correlation reasoning. Its key innovation lies in: Dynamic adjustment of risk weights based on scene transitions: This model is trained to identify state transitions in scene sequences. For example, when the system detects that a child's scene label sequence changes from a flat playground to a stairwell, the model will recognize this as a spatial transition event. For such events, the model will automatically and instantaneously increase the risk weight coefficients associated with highly dynamic behaviors such as running and jumping. The underlying logic is that sudden changes in the spatial environment require children to make more complex balance adjustments, and the risk of instability during vigorous movement is higher than in a single stable scene.

[0065] Dynamic amplification of risk coefficients based on behavior-individual feature coupling: This model can dynamically couple specific high-risk behaviors with individual physiological vulnerability characteristics to achieve non-linear amplification of risk. For example, when the behavior recognition submodule 210 identifies a face-down forward fall as a high-risk behavior, the model will immediately query the individual health information input interface 140 for information about the child's oral cavity status.

[0066] If the query reveals that the child is in the mixed dentition stage with loose primary teeth, or has vulnerable characteristics such as protruding front teeth, the model will apply a significant amplification factor (e.g., multiply the risk value by 1.5 or higher) to the risk score of this fall. This mechanism makes risk assessment no longer a simple summation of factors, but rather an accurate capture of the resonance effect caused by a specific combination of risk factors, thereby improving the sensitivity of early warning for high-risk events.

[0067] Finally, unit 310 integrates the dynamically adjusted and amplified risk parameters to calculate a highly contextualized and personalized dental trauma risk score for the current moment. Unit 310 receives input factors from multiple sources, including: behavior-scenario risk coefficients output by the behavior-scenario cross-fusion modeling unit 230; impact quantification values ​​directly acquired by the dynamic behavior data acquisition unit 110; and quantitative indicators characterizing individual vulnerability imported from the individual health information input interface 140. These input factors are uniformly represented as a risk factor vector. Unit 310 calculates the current dental trauma risk score using the following weighted scoring function. : ; In the formula, For at any time The calculated risk score for dental trauma; The current moment; The activation function is used to map the weighted sum to the interval [0,1], making it a normalized score; The total number of risk factors input into the model; This serves as an index for risk factors, with values ​​ranging from 1 to... ; For the first The adjustable weight coefficients of each risk factor are dynamically adjusted by a personalized risk weight adjustment unit. For the first Each risk factor at time The quantized value.

[0068] Calculate the risk score Then, the unit 310 will, according to a set of preset thresholds (e.g., The rating value is divided into low ,middle ,high Extremely high Four risk levels.

[0069] The time window trend prediction mechanism 320 employs a short-time series prediction algorithm to perform rolling predictions on historical dental trauma risk score sequences. In a specific embodiment, mechanism 320 uses a gated recurrent unit (GRU) neural network model. This GRU model takes the risk score sequence within the most recent time window as input. By learning from this sequence, the model can capture the changing patterns and dynamic trends of the risk scores. The output of mechanism 320 is a predicted value for the risk score at one or more future time steps, the calculation process of which can be expressed as follows: ; In the formula, For future moments Predicted values ​​of dental trauma risk scores; For a trained gated recurrent unit neural network prediction model; The time series consisting of a series of past risk scores is input into the prediction model; The current moment; The predicted time step is within a preset time window range. This represents the length of the time window used for prediction.

[0070] By analyzing the first derivative of the predicted sequence, the upward slope of the risk score can be obtained; by finding the local maximum of the predicted sequence, the possible time point of the risk peak can be predicted; by comparing the predicted value with the risk level threshold, early warning of threshold overshoot can be achieved.

[0071] The personalized risk weight adjustment unit 330 has its functions enhanced in this embodiment. It is not only used for initial weight setting, but more importantly, it serves as a dynamic parameter server to provide the latest individual vulnerability parameters (such as tooth loosening level, sensitivity of historical injury sites, etc.) in real time for the spatiotemporal behavior and individual dynamic correlation model, so as to support the dynamic amplification calculation of risk coefficients.

[0072] During system initialization or when updating a child's individual health information, the unit 330 will select or generate an initial set of weight vectors from a preset weight configuration library based on the child's age group, the activity level assessment obtained through historical data analysis (e.g., divided into quiet, normal, and active types), and whether there are oral structural vulnerabilities such as protruding anterior teeth or severely loose teeth.

[0073] For example, for a child assessed as active and with protruding front teeth, Unit 330 would increase the weight of risk factors associated with high-dynamic behaviors such as running and colliding, as well as the weight of risk factors associated with frontal impacts, thereby making the risk scoring model more sensitive to that particular child.

[0074] Please see the appendix Figure 6 , Figure 6This is a schematic diagram of a nursing suggestion generation and push module 400 according to an embodiment of the present invention. The nursing suggestion generation and push module 400 is configured to execute a series of response operations when the dental trauma risk score output by the risk prediction and scoring module 300 exceeds a preset threshold. Its core function is to transform abstract risk level signals into specific, executable intervention instructions and effectively communicate them to relevant parties. In a specific embodiment, the nursing suggestion generation and push module 400 includes a suggestion generation unit 410, a multi-terminal push unit 420, and a visualization feedback module 430.

[0075] The suggestion generation unit 410 integrates a structured intervention suggestion library. This library is a knowledge base storing a large number of nursing suggestion entries. Each entry is bound to one or more trigger conditions, including: risk level (low, medium, high, very high), specific behavioral tags (such as running, jumping), specific scene tags (such as stairwell, slide), the child's mixed dentition stage (such as primary dentition, mixed dentition), and the existence of a specific historical dental trauma record. When the suggestion generation unit 410 receives a trigger signal (containing the current risk level, behavior, and scene information) from the risk prediction and scoring module 300, it searches the intervention suggestion library using a preset rule matching engine or a decision tree model.

[0076] The process matches all suggestions that meet the current triggering conditions and selects or combines them according to priority rules to generate the most suitable care suggestion for the current situation. For example, when high risk, running behavior, stairwell scene, and the child is in the mixed dentition stage are detected, the generated suggestion might be: Please stop running near the stairs.

[0077] The multi-terminal push unit 420 is responsible for reliably and in real-time distributing the care recommendations generated by the recommendation generation unit 410 to one or more pre-bound user terminals via network communication protocols (such as TCP / IP or MQTT). These terminals include mobile applications on smartphones held by guardians, smartwatches worn by children, and campus security management platforms in specific deployment environments (such as schools or kindergartens). The unit 420 is configured with a risk level response mapping mechanism that maps different risk levels to different alert formats and intensities. For example: Medium risk level: Triggers a slight, short vibration from a smartwatch worn by a child, accompanied by a gentle alert sound.

[0078] High-risk level: Triggers the smartwatch to vibrate multiple times in a rhythmic manner and broadcast specific suggestions in the form of a cartoon voice. At the same time, it pushes a text notification containing risk details and suggestions to the guardian's mobile application.

[0079] Extremely high risk level: Triggers the smartwatch to emit continuous strong vibrations and alarm sounds, while simultaneously pushing the highest priority alarm notification to the guardian's mobile application. The alarm information can also be linked to the campus terminal broadcasting system or directly notified to relevant management personnel by calling the API interface.

[0080] The visual feedback module 430 is primarily deployed on child-friendly terminal devices such as smartwatches worn by children. The goal of this module 430 is to convert care suggestions in text or voice format into visual forms that are easier for children to understand and accept. For example, for the suggestion to slow down, an animation of a small person changing from running quickly to walking slowly will be displayed on the screen; for the suggestion to stay away from the table corner, a table corner icon will be displayed on the screen along with a red prohibition symbol.

[0081] Furthermore, module 430 also supports simple interactive feedback from children. For example, after receiving a reminder, a child can confirm receipt by touching the "OK" or "I understand" button on the screen. This feedback is recorded and transmitted back to the system, which can be used to confirm the effectiveness of information delivery and also serve as one of the inputs to the data loop and model update module 500 to evaluate the effectiveness of different reminder methods for specific children.

[0082] The data closure and model update module 500 is a key component in realizing the adaptive and self-evolving capabilities of the system of the present invention. It is configured to analyze and optimize the system's predictive model after collecting real feedback data on dental trauma events. In a specific embodiment, the model update module 500 includes a dental trauma event recording interface 510, a model error analysis unit 520, and a model optimization module 530.

[0083] The Dental Trauma Event Recording Interface 510 is a software interface deployed within a guardian's mobile application. When a child experiences a dental trauma event, the guardian can upload relevant feedback data through this interface. The interface guides users to submit information in a structured manner, with data fields including: the precise timestamp of the event; the geographical location or scene description of the event; the specific location of the damaged tooth; the type of trauma (e.g., tooth concussion, tooth dislocation, crown fracture, root fracture, etc.); the user's self-assessed possible contributing factors; and support for uploading on-site photos or hospital diagnostic reports as supplementary evidence. This real-world event data, proactively provided by users, forms the cornerstone of model optimization.

[0084] The model error analysis unit 520 is configured to evaluate the system's predictive performance and attribute prediction errors. This unit 520 continuously compares the system's own generated warning logs with real event reports received by the dental trauma event recording interface 510. When a real event report is received, the unit 520 backtracks through the system's warning logs for a preset time window (e.g., the previous 5 minutes) prior to the event's occurrence. If no warnings of excessively high or extremely high risk levels are found within this time window, the event is considered a missed report.

[0085] For missed detection samples, unit 520 automatically extracts all multi-source raw data within the time window before the event occurs and marks it as a high-value, difficult sample that the model needs to focus on learning, storing it in a blind zone behavior sample library. In addition, unit 520 can also analyze false alarms, that is, feedback where the system issues a warning but no actual harm or dangerous behavior avoidance occurs within a long period of time.

[0086] The model optimization module 530, based on the output of the model error analysis unit 520, periodically or event-triggeredly updates and optimizes the model in the risk prediction and scoring module 300. This module 530 employs two collaborative optimization strategies: The first method is the few-shot fine-tuning mechanism. This mechanism adds the hard sample data corresponding to the missed events collected by the model error analysis unit 520 to the existing training dataset. Then, a small learning rate is used to adjust the model parameters (such as the weights of the CNN-LSTM or the weights of the weighted scoring function) in the risk scoring model building unit 310. Retraining is then performed. This fine-tuning method allows the model to focus on learning and correcting the blind spots that cause false negatives without compromising its recognition ability in existing scenarios, thereby improving its generalization ability.

[0087] The second mechanism is a personalized model fine-tuning mechanism. This mechanism generates a unique predictive model copy for children who meet specific conditions. Triggering conditions include: the child's accumulated historical data reaching a preset threshold, or the child being identified as a high-risk individual with a specific oral condition. Once triggered, the system creates an independent local model copy based on the current master predictive model.

[0088] Subsequently, the model optimization module 530 uses only the child's specific historical behavioral data, event feedback data, and intervention response data to fine-tune this local model copy. This process not only optimizes the model weights but also adjusts the child-specific risk level classification thresholds, thereby enabling the entire system to achieve a higher fit and predictive accuracy for the child's behavioral patterns and risk characteristics.

[0089] In one specific embodiment, the data acquisition module not only serves a single child but also acts as a data node in a distributed data network. Multiple data acquisition modules deployed in different institutions (such as different kindergartens, health check centers, or hospitals) collectively constitute a multi-center data source. Under this architecture, the raw data collected by each node is processed and used for model training only locally, adhering to privacy protection principles, and no raw data is uploaded or exchanged.

[0090] The model optimization module 530, based on the output of the model error analysis unit 520, periodically or event-triggeredly updates and optimizes the model in the risk prediction and scoring module 300. In a specific embodiment, this module 530 employs a collaborative optimization strategy based on a combination of federated learning and population feature transfer.

[0091] The first stage of this collaborative optimization strategy is to build a general risk prediction model across populations. This process is carried out within a federated learning framework. Systems deployed at multiple data nodes in different kindergartens, hospitals, etc., act as participants, each training its risk prediction model locally using locally collected and labeled error samples (hard samples).

[0092] After each training round, each node uploads only the model's parameter updates (such as gradient or weight changes), not the original data, to a centralized aggregation server. The aggregation server securely aggregates the parameter updates uploaded by all participants (e.g., using a federated averaging algorithm), generates an updated global model parameter set, and sends it back to each data node to begin the next round of local training. Through multiple iterations, the system trains a cross-population general model with strong generalization capabilities that integrates data from multiple centers and scenarios, without compromising the privacy of any party's original data.

[0093] The second stage of this collaborative optimization strategy is to achieve efficient fine-tuning of personalized models. This process utilizes group feature transfer technology. When a new child user starts using the system, or when an existing user's personalized model needs to be updated, the system no longer trains from scratch. Instead, it first uses the cross-group general model trained in the previous stage as the initialization basis for a local model copy specific to that child.

[0094] Furthermore, the model optimization module 530 selectively transfers risk feature patterns most relevant to the child's group from the feature space learned by the general model, based on the child's individual characteristics (such as age, mixed dentition stage, and activity level classification). For example, for a 6-year-old child in the mixed dentition stage, the model will focus on transferring feature weights learned from the general model regarding common high-risk micro-movements in running by 6-year-old children in the mixed dentition stage. This approach greatly reduces the dependence of personalized models on the amount of data available for each individual, solves the cold start problem in model optimization, and accelerates model convergence speed and the accuracy of fitting individual risks.

[0095] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional feature fusion child dental trauma prospective risk early warning and response system, characterized in that, The system comprises: a data acquisition module for acquiring multi-source raw data of children, including dynamic behavior data, location information, and environmental parameter data collected in real time through wearable devices or mobile terminals, and individual oral health information and historical dental trauma records input through user input or a hospital interface; a behavior and scene recognition module configured to receive and process the multi-source raw data acquired by the data acquisition module, identify the current behavior state and the activity scene of the child based on a deep learning model; a risk prediction and scoring module configured to select and apply an exclusive age segment sub-model according to the age segment of the child, integrate the identification results of the behavior and scene recognition module, the individual oral health information, and the historical dental trauma records, construct a multi-factor risk prediction model, and calculate a quantitative dental trauma risk score value and support trend prediction within a time window; a nursing suggestion generation and pushing module configured to automatically generate and push targeted nursing suggestions to multiple terminals when the risk score value exceeds a preset threshold; a data closed loop and model updating module configured to analyze false positive and false negative samples based on dental trauma event feedback data collected from the multiple data nodes under a federated learning framework, and optimize the parameters of the multi-factor risk prediction model based on the analysis results. 2.The multi-dimensional feature fusion-based child dental trauma pre-impact risk warning and response system according to claim 1, characterized in that, The data acquisition module comprises: a dynamic behavior data acquisition unit integrated with a wearable terminal of a three-axis accelerometer, a gyroscope, and an inertial measurement unit, for real-time acquisition of dynamic behavior parameters of the child's motion posture, acceleration, angular velocity, and displacement path; an environmental perception data acquisition unit including an ambient light sensor, a barometric pressure sensor, and a collision sensing element, for acquiring physical environmental characteristics of the activity space where the child is located; a location data acquisition unit based on a fusion positioning technology including a global positioning system, a wireless local area network positioning, and a Bluetooth low energy beacon, for acquiring the activity trajectory of the child; an individual health information input interface for importing structured health information data of the child, including tooth replacement status, tooth loosening level, and previous dental trauma records. 3.The multi-dimensional feature fusion-based children's dental trauma pre-impact risk warning and response system according to claim 1, characterized in that, The behavior and scene recognition module comprises: a behavior recognition sub-module that uses a deep learning model combining a convolutional neural network and a long short-term memory network to model the dynamic behavior parameters to identify high dynamic actions such as running, jumping, and sudden falling; a scene recognition sub-module based on an image recognition model or an audio convolution analysis network, and combining the location data, to classify the activity space where the child is currently located, including a slide exit, a stair exit, and a playground hard ground; a behavior-scene cross-fusion modeling unit configured to cross-couple model the identified high dynamic actions and scene classification, construct a behavior-scene risk matrix, and use it as input for subsequent risk scoring. An active learning and edge case mining unit configured to actively generate and push a labeling request to a guardian terminal to obtain real-time manual labeling of a current edge case when a confidence level of an identification result output by the behavior identification submodule is lower than a preset value or a risk score value calculated by the risk prediction and scoring module is in a critical interval near a preset risk threshold.

4. The multi-dimensional feature fusion-based prospective risk warning and response system for children's dental trauma according to claim 1, characterized in that, The risk prediction and scoring module comprises: A risk score model construction unit taking the output of the behavior and scene identification module, the individual oral health data, and the historical dental trauma records as input factors, and selecting and applying an age-segmented sub-model according to the age segmentation of children, the age-segmented sub-model internally constructed with a spatiotemporal behavior and individual dynamic correlation model, the dynamic correlation model used to dynamically adjust the risk weight of a corresponding behavior when a spatial transition of a child is detected, and dynamically amplify the risk coefficient when a specific high-risk behavior and individual physiological vulnerability feature are identified, to calculate a dental trauma risk score value at the current time, and divide the risk score value into four risk levels in combination with a preset threshold; A time window trend prediction mechanism that uses an algorithm based on short sequence prediction to rollingly predict the sequence of the dental trauma risk score value to analyze the trend of the risk score in a future set time window; A personalized risk weight adjustment unit configured to provide individual parameters required for dynamic adjustment of the spatiotemporal behavior and individual dynamic correlation model according to the age group, activity level assessment, and oral structure vulnerability of children.

5. The multi-dimensional feature fusion-based prospective risk warning and response system for children's dental trauma according to claim 4, characterized in that, The time window trend prediction mechanism uses a gated recurrent unit neural network to model the risk score curve to predict the rising slope of the risk score, the peak prediction time point, and the threshold out-of-bound early warning indication, thereby realizing prospective prediction.

6. The multi-dimensional feature fusion-based prospective risk warning and response system for children's dental trauma according to claim 1, characterized in that, The care suggestion generation and pushing module comprises: A suggestion generation unit configured to automatically generate personalized care suggestions from an intervention suggestion library after detecting that the dental trauma risk score value reaches or exceeds a preset threshold, in combination with the currently identified high-risk behavior type and activity scene, the suggestions including wearing a mouthguard, pausing intense activities, and moving away from dangerous areas; A multi-terminal pushing unit for real-time delivery of the generated care suggestion content to receiving terminals of the parent end, the child end, and the campus end; A visual feedback module for presenting the care suggestions in the form of images, animations, or cartoon voices to adapt to the cognitive ability and interaction habits of children.

7. The multi-dimensional feature fusion-based prospective risk warning and response system for children's dental trauma according to claim 6, characterized in that, The multi-terminal pushing unit is configured with a risk level response mapping mechanism for dividing the reminder intensity into multiple levels according to the risk level output by the risk prediction and scoring module, and dynamically matching different reminder forms, including but not limited to: a light vibration reminder triggered by a medium risk, a voice broadcast triggered by a high risk, and a multi-channel strong reminder triggered by an extremely high risk and linked to a campus terminal broadcast.

8. The multi-dimensional feature fusion-based prospective risk alert and response system for children's dental trauma according to claim 1, wherein, The data closed loop and model updating module comprises: A dental trauma event record interface is configured to upload feedback data including event timestamp, trauma site, trauma cause and related image data by a user after the actual occurrence of a dental trauma event; A model error analysis unit is configured to backtrack the data window before the event occurs when a false negative or false positive occurs, classify and attribute the error type to build a blind area behavior sample library; A model optimization module is configured to use the error samples as high-value samples, continuously optimize the multi-factor risk prediction model under the federal learning framework by performing local model training at each data node and aggregating model parameters to generate a universal model across populations. 9.The multi-dimensional feature fusion-based child tooth trauma pre-impact risk warning and response system according to claim 8, characterized in that, The model optimization module has a model individualization fine-tuning mechanism, which includes: When the specific child's historical usage data exceeds a preset threshold or belongs to a high-risk population, a local model copy is generated for the specific child; The cross-population universal model is used for initialization, and the high-risk behavior patterns of populations similar to the child's characteristics are transferred to the local model copy through population characteristic transfer The risk score threshold and risk factor weight of the local model copy are periodically updated according to the child's behavior habits and response performance.

10. The multi-dimensional feature fusion-based prospective risk warning and response system for children's dental trauma according to claim 1, characterized in that, The system supports a cloud-edge collaborative deployment method, including deploying a lightweight model version on a wearable terminal or mobile device to perform real-time behavior recognition and basic scoring tasks; deploying a complete intelligent algorithm model on a cloud server for centralized management, complex model training and cross-user learning; And through parameter synchronization and model hierarchical deployment, a collaborative working mode of cloud training optimization and terminal regular acquisition of simplified model copies is realized.

Images