Method and device for assessing risk of brain abnormality based on transfer learning and pose estimation

Abstract

This invention provides a method and device for assessing the risk of craniocerebral abnormalities based on transfer learning and posture estimation. The method includes acquiring millimeter-wave radar point cloud data of the monitored object; performing topological structure analysis on the millimeter-wave radar point cloud data to extract geodesic distance-invariant feature vectors (GDIs); inputting the GDIs into a pre-constructed craniocerebral abnormality risk assessment model; and obtaining the craniocerebral abnormality risk assessment result of the monitored object based on the abnormality risk probability output by the craniocerebral abnormality risk assessment model. This invention can improve the accuracy of early craniocerebral abnormality risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of medical and health monitoring and artificial intelligence technology, specifically relating to a method and device for assessing the risk of cranial abnormalities based on transfer learning and posture estimation. Background Technology

[0002] Stroke (such as cerebral infarction) is one of the leading causes of disability and death worldwide. The earlier it is detected and intervened, the better the prognosis for patients. Traditional stroke risk assessment mainly relies on periodic examinations conducted by medical institutions (such as blood indicators and imaging examinations), which have the problems of being static and lagging, and cannot achieve continuous and dynamic early warning.

[0003] With the development of sensing technology and artificial intelligence, vision-based human posture estimation and millimeter-wave radar-based vital sign monitoring have made home health monitoring possible. However, existing technologies face many challenges when applied to home-based stroke early warning scenarios: Accuracy versus privacy: While continuous posture monitoring based on cameras (RGB or depth) can provide rich spatial information, it poses a serious risk of privacy leakage and is susceptible to changes in lighting and occlusion. On the other hand, while using millimeter-wave radar can protect privacy and enable all-weather detection, it provides sparse spatial semantic information, making it difficult to accurately capture subtle postural abnormalities in the precursors of stroke (such as slight drooping of the mouth or mild weakness in one side of the body).

[0004] Insufficient scene adaptability and robustness: Home environments present interference such as furniture occlusion, loose clothing, and complex lighting. Traditional pose estimation methods rely on absolute spatial coordinates or general image features, which are prone to feature drift in such complex scenes, leading to a high false alarm rate.

[0005] Insufficient utilization of data value: When using federated learning for collaborative training of multi-user data to improve model generalization, existing methods (such as FedAvg) typically perform simple weighted aggregation based solely on the amount of client data, ignoring the quality differences, individual physiological differences, and high false positive costs inherent in medical data. This results in low-quality data contaminating the model, while high-value data representing high-risk groups or unique patterns are buried, making it impossible to achieve safe, effective, and equitable personalized model optimization.

[0006] Therefore, there is an urgent need for a method that can adapt to complex home environments while protecting user privacy, and fully and securely utilize the value of distributed data to achieve high-precision, low-false-report early risk assessment of cranial abnormalities. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method and device for assessing the risk of craniocerebral abnormalities based on transfer learning and posture estimation, so as to improve the accuracy of early craniocerebral abnormality risk assessment.

[0008] In a first aspect, the present invention provides a method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation, the method comprising the following steps: Acquire millimeter-wave radar point cloud data of the monitored object; Topological structure analysis is performed on millimeter-wave radar point cloud data to extract geodesic distance invariant feature vectors; geodesic distance invariant feature vectors characterize the relative topological relationships and dynamic micro-motion information between key parts of the human body surface; The geodesic distance invariant feature vector is input into the pre-constructed craniocerebral abnormality risk assessment model. The craniocerebral abnormality risk assessment model is constructed based on an asymmetric cross-modal attention mechanism. During the training phase, the model establishes a mapping relationship from millimeter-wave radar temporal vital signs to video spatial attitude features, while in the current inference phase, it only performs calculations based on millimeter-wave radar data. Based on the abnormal risk probability output by the aforementioned intracranial abnormality risk assessment model, the intracranial abnormality risk assessment result of the monitored object is obtained; the intracranial abnormality risk assessment result includes the warning level.

[0009] Optionally, topological analysis is performed on the millimeter-wave radar point cloud data to extract geodesic range-invariant feature vectors, including: Extract human body point sets from millimeter-wave radar point clouds, map the human body point sets to a spherical coordinate system centered on the radar, and generate a body surface depth raster map. The surface depth raster map is treated as a graph structure. The raster is used as a node. The edge weights are calculated based on the angle and depth differences between the nodes. The local geodesic path map of the body surface is constructed based on the edge weights. Key points are selected in the local geodesic path map of the body surface based on curvature, density and short-term stability indices; key points correspond to key parts of human anatomy. The shortest path algorithm is used to calculate the geodesic distance between all key plotted point pairs, and a geodesic distance matrix is ​​constructed and normalized. The left and right difference features and short-time window micro-motion features are calculated based on the normalized geodesic distance matrix, and then spliced ​​to generate a geodesic distance invariant feature vector.

[0010] Optionally, edge weights can be calculated based on the angle and depth differences between nodes, including: Through calculation formula

[0011] Obtaining border rights ;in, and Representing adjacent nodes respectively Angular difference in spherical coordinates This represents the depth value corresponding to the node. This represents the deformation sensitivity coefficient, used to adjust the weight of the impact of depth changes on distance calculation. This coefficient is dynamically adjusted based on the frequency of environmental occlusion.

[0012] Optional asymmetric cross-modal attention mechanisms include: The craniocerebral abnormality risk assessment model includes a video feature extraction branch and a radar feature extraction branch; the video feature extraction branch is used to extract video features from video data; the radar feature extraction branch is used to extract radar features from millimeter-wave radar point cloud data. During the model training phase, cross-modal feature fusion is performed using the RV-Transformer model; video features are mapped to query vectors, and radar features are mapped to key vectors and value vectors. By forcing the cranial abnormality risk assessment model to learn the radar micro-motion feature pattern in the key vector when the spatial pose corresponding to the query vector is abnormal, an attention mechanism is used to establish a mapping relationship from radar feature abnormalities to video feature abnormalities, thereby encoding medical visual semantics into the radar feature processing logic.

[0013] Optionally, the parameter updates of the intracranial abnormality risk assessment model employ a federated learning aggregation mechanism, which includes: For each household edge device, a contribution score is calculated from three dimensions: data reliability, task performance, and coverage generalization. Hard gating mechanisms are applied to filter data sources that do not meet preset quality thresholds, while soft gating mechanisms are applied to assign protective weights to data from home edge devices identified as high-risk groups. Based on contribution scores and gating results, the aggregate weights of each household edge device are dynamically calculated and weighted averaged to generate a new generation of global model.

[0014] Optionally, the hard gating mechanism includes: when the false alarm rate of the local cranial abnormality risk assessment model of the home terminal device is higher than the preset false alarm rate tolerance threshold, or the data signal-to-noise ratio is lower than the preset signal-to-noise ratio qualified threshold, the aggregation weight of the home terminal device is reset to zero. The soft gating mechanism includes: identifying whether a home device belongs to a high-risk group; if it belongs to a high-risk group and the model recall rate is lower than the preset value, a protective coefficient is introduced to increase the aggregation weight of the home device.

[0015] Optionally, contribution scores can be calculated from three dimensions: data reliability, task performance, and coverage generalization, including: Through calculation formula

[0016] Get home devices Data reliability score The data reliability score is used to assess the cleanliness and reliability of home-based devices. Indicates the weighting coefficient. This represents the Sigmoid activation function. Indicates home terminal device The amount of data, Indicates home terminal device signal-to-noise ratio, Indicates home terminal device Data missing rate, Indicates the device number for the home device; Through calculation formula

[0017] Get home devices Task performance score ;in, Indicates the weighting coefficient. Indicates home terminal device The AUC value on the validation set, Indicates home terminal device Recall rate on the validation set, Indicates home terminal device False positive rate on the validation set; Through calculation formula

[0018] Get home devices Coverage generalization score ;in, Indicates the weighting coefficient. Indicates home terminal device The diversity of coverage scenarios, Indicates home terminal device The population coverage scarcity indicator corresponding to the monitored subjects. This is used to penalize home devices that deviate too much from the global data distribution, preventing the negative impact of extreme outliers; Home devices Contribution score Data reliability score Task performance score and coverage generalization score The weighted average is obtained.

[0019] Optionally, the selection strategy for key anchor points is as follows: Through calculation formula

[0020] Obtain key points ;in, Indicates a predefined search area. Represents a node curvature, Represents a node Point density, Indicates short-term stability index, This represents the weighting coefficient of each indicator.

[0021] Optionally, the initialization and training process of the craniocerebral abnormality risk assessment model includes: A medical attention weight matrix was constructed, and based on prior knowledge of clinical medicine, higher initial weights were assigned to feature channels that were strongly correlated with cranial abnormalities, and lower initial weights were assigned to feature channels that were weakly correlated. A dynamic risk-sensitive penalty term is introduced into the objective function of model training to enhance the attention to high-risk samples during backpropagation; the coefficient of the penalty term is positively correlated with the model's misclassification rate for samples of different risk levels on the validation set.

[0022] Secondly, the present invention provides a brain abnormality risk assessment device, comprising: Millimeter-wave radar modules are used to collect radar point cloud data of the monitored objects; Memory, used to store computer programs and brain abnormality risk assessment models; The processor, electrically connected to the memory and millimeter-wave radar module, is used to execute computer programs to implement the above-described method steps; The early warning module is used to output early warning information when the risk assessment results of craniocerebral abnormalities meet preset triggering conditions.

[0023] The present invention has at least the following beneficial effects: Through an innovative asymmetric cross-modal attention mechanism, the model utilizes the high spatial resolution of video data to verify and enhance the association between radar features and pathological postures during the offline training phase, while only requiring millimeter-wave radar single-mode operation during the online inference phase. This fundamentally solves the privacy leakage problem caused by continuous video surveillance. At the same time, the strong mapping relationship established in the early stage ensures that high-precision anomaly identification can be achieved using only radar data. By extracting geodesic distance-invariant features, the focus is shifted from the position and orientation of the human body in absolute space to the inherent relative topological relationship between key parts of the human body. This feature is naturally robust to occlusion, clothing, lighting changes, and changes in sensor distance and angle, effectively overcoming interference from the complex home environment and significantly reducing the false alarm rate. Through a three-dimensional contribution-gated weighted aggregation mechanism, the value of the home edge is comprehensively evaluated from three dimensions: data reliability, task performance, and coverage generalization. Hard gating is introduced to filter low-quality data, and soft gating is introduced to protect high-risk groups. This makes the evolution of the global model closely align with the clinical needs of high-precision and low-false-alarm stroke early warning, and avoids ignoring valuable but potentially poorly performing high-risk user data, achieving safer and more effective personalized model optimization. Attached Figure Description

[0024] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0025] Figure 1 This is a flowchart of a brain abnormality risk assessment method based on transfer learning and posture estimation in one embodiment of this application; Figure 2 A flowchart of the federated learning aggregation mechanism used for parameter updating of the craniocerebral abnormality risk assessment model in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a brain abnormality risk assessment device according to one embodiment of this application. Detailed Implementation

[0026] The technical solution of the present invention will now be described in detail and completely with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0027] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0028] Example 1 See Figure 1 This embodiment provides a method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation. This method can be executed collaboratively by a home edge device (e.g., integrated millimeter-wave radar) deployed in a home environment and a cloud server, specifically including steps 11 to 14.

[0029] Step 11: Obtain millimeter-wave radar point cloud data of the monitored object.

[0030] Specifically, millimeter-wave radar is used to continuously or periodically scan and monitor objects (such as elderly users) to acquire their millimeter-wave radar point cloud data. In one feasible implementation, the millimeter-wave radar operates at a frequency of 60 GHz, with a detection range of 0.5 to 5 meters and a sampling rate of 10 Hz, enabling the acquisition of point cloud data including target distance, azimuth angle, radial velocity, and reflection intensity. This data acquisition process is non-contact and seamless, fully protecting user privacy (during the online inference phase, integrated millimeter-wave radar collects vital signs and micro-movement data, while the camera only briefly activates for verification when a suspected anomaly is detected, thus protecting privacy).

[0031] In practical implementation, to ensure the fusion of multimodal data, radar data must undergo strict alignment processing during acquisition: hardware synchronization between the radar and camera is achieved through the GPIO interface, with latency controlled to ≤50ms. (or hardware timestamp synchronization accuracy ≤ 1ms); establish a mapping relationship between "radar voxel → video pixel" by using camera focal length, principal point coordinates and radar installation angle (e.g., 0° horizontal, 30° vertical), and project the radar's 3D world coordinates onto the 2D pixel coordinates of the video frame.

[0032] To improve data quality, after obtaining the raw point cloud, preprocessing is required to extract valid data: Kalman filtering is used for noise reduction, missing values ​​are filled using the mean, and standardization is performed to eliminate individual differences.

[0033] The above methods can obtain high-precision, low-interference, and pathologically relevant millimeter-wave radar point cloud data, providing accurate input for subsequent cranial abnormality risk assessment models.

[0034] Step 12: Perform topological structure analysis on the millimeter-wave radar point cloud data and extract geodesic distance invariant feature vectors.

[0035] The aforementioned geodesic distance invariant eigenvectors characterize the relative topological relationships and dynamic micro-motion information between key parts of the human body surface.

[0036] Specifically, step 12 includes steps 12.1 to 12.5.

[0037] Step 12.1: Extract the human body point set from the millimeter-wave radar point cloud, map the human body point set to a spherical coordinate system centered on the radar, and generate a body surface depth raster map.

[0038] Specifically, the DBSCAN algorithm (with a neighborhood radius of 0.1m and a minimum number of points of 5) or the threshold connectivity algorithm can be used to extract a set of candidate points belonging to the human body from the sparse point cloud. The extracted point set is transferred from the Cartesian coordinate system. Transform to a radar-centric spherical coordinate system ; Create a surface depth raster map with a resolution of 64×64×32. Surface depth raster map The value of each raster cell is calculated using median filtering: This step transforms the discrete point cloud into a continuous surface representation.

[0039] Step 12.2: Treat the surface depth raster map as a graph structure, use the raster as nodes, calculate the edge weights based on the angle and depth differences between nodes, and construct a local geodesic path map of the surface based on the edge weights.

[0040] Specifically, the depth raster map is viewed as a network graph. A grid is a node, and connecting edges are established between adjacent grids.

[0041] The edge weights are calculated based on the angle and depth differences between nodes, including: Through calculation formula

[0042] Obtaining border rights ;in, and Representing adjacent nodes respectively Angular difference in spherical coordinates This represents the depth value corresponding to the node. This represents the deformation sensitivity coefficient (initial value can be set to 0.8), used to adjust the weight of depth changes on distance calculation. This coefficient is dynamically adjusted based on the frequency of environmental occlusion. For example, if the frequency of environmental occlusion is greater than 5 times per day, then... Automatically adjusted to 1.5. In this embodiment of the invention, the edge weight represents the "fold distance" of the body surface, laying the foundation for subsequent calculation of geodesic distance (the shortest path along the body surface).

[0043] Step 12.3: Select key plotting points on the local geodesic path map of the body surface based on curvature, density, and short-term stability indices.

[0044] In this embodiment of the invention, key anchor points correspond to key anatomical locations in the human body. In practice, instead of selecting traditional skeletal points, multiple key anchor points are automatically selected based on stability, density, and medical relevance. .

[0045] In one feasible implementation, screening is performed using three-dimensional indicators: curvature (curv≥0.7), density (dens≥8 / cm³), and short-term stability (stable≥0.85). For example, two on the face (corresponding to the corners of the mouth), two on each shoulder, one on each elbow, wrist, hip, and knee, and one on the center of the torso.

[0046] The selection strategy is to select each predefined anatomical region Internally, through calculation formula

[0047] Obtain key points ;in, Indicates a predefined search area. Represents a node The curvature (calculated by the change in neighborhood depth). Represents a node Point density, This represents a short-term stability index, used to measure the stability of a node's position within a short time window. This represents the weighting coefficient of each indicator.

[0048] Step 12.4: Calculate the geodesic distance between all key plotted point pairs using the shortest path algorithm, construct the geodesic distance matrix, and perform normalization processing.

[0049] Specifically, Dijkstra's algorithm is used to calculate the shortest path length between all key point pairs. ,generate 3D geodesic distance matrix .

[0050] To eliminate the influence of individual height, body shape, and distance from the radar, normalization is performed: The normalized geodesic distance matrix is ​​obtained. Normalized geodesic distance matrix It reflects the proportional relationship of the human body itself, achieving scale invariance (regardless of a person's height, weight, or build, the relative topological features remain stable).

[0051] Step 12.5: Calculate the left and right difference features and short-time window micro-motion features based on the normalized geodesic distance matrix, and concatenate them to generate a geodesic distance invariant feature vector.

[0052] Specifically, the normalized geodesic distance matrix The upper triangular part is unfolded into a one-dimensional vector, representing the static topological structure of the human body surface, thus obtaining geometric features. ; Calculate the difference in distance from key points on the same side to the central axis to quantify the degree of facial asymmetry or limb asymmetry (values ​​within 10 ... (between), to obtain left and right difference features Within a 5-second sliding window (1-second step), the standard deviation and coefficient of variation of geodesic distance are calculated to capture dynamic micro-motion information such as torso swaying and trembling, thus obtaining short-time window micro-motion characteristics. .

[0053] Ultimately, , as well as By splicing, a highly robust geodesic distance-invariant eigenvector is formed. .

[0054] The final geodesic distance invariant eigenvector Based entirely on the relative topological relationships and dynamic changes of the human body itself, it exhibits strong robustness to common interferences in home settings such as occlusion, distance changes, orientation changes, and loose clothing, providing a stable and reliable input for subsequent accurate assessment.

[0055] Step 13: Input the geodesic distance invariant feature vector into the pre-constructed cranial abnormality risk assessment model.

[0056] In this embodiment of the invention, the cranial abnormality risk assessment model is constructed based on an asymmetric cross-modal attention mechanism. During the training phase, the model establishes a mapping relationship from millimeter-wave radar temporal characteristics to video spatial attitude characteristics, while during the current inference phase, it only performs calculations based on millimeter-wave radar data.

[0057] In one feasible implementation, the craniocerebral abnormality risk assessment model includes a video feature extraction branch and a radar feature extraction branch, which are independent of each other but interact deeply in the middle. Specifically, the video feature extraction branch is used to extract video features from video data, and the radar feature extraction branch is used to extract radar features from millimeter-wave radar point cloud data.

[0058] In practical implementation, the radar feature extraction branch employs a "1D-CNN + BiLSTM" network for millimeter-wave radar time-series data. The 1D-CNN extracts local anomaly segments, while the BiLSTM captures long-term temporal dependencies, ultimately outputting a 256-dimensional radar time-series feature vector. Simultaneously, the point cloud is transformed into a dense feature map using Voxel Grid, and 3D spatial features are extracted via VoxelNet (3DCNN) and compressed into a 2D feature map that matches the video features. The video feature extraction branch utilizes ResNet-50 (2D CNN) to extract spatial pose features from the video frames. .

[0059] To achieve "asymmetric" mapping, this invention uses the RV-Transformer model to perform cross-modal feature fusion during the model training phase to establish cross-modal semantic mapping.

[0060] Specifically, video features are mapped to query vectors, and radar features are mapped to key vectors and value vectors.

[0061] Finally, by using an attention mechanism to force the cranial abnormality risk assessment model to learn the radar micro-motion feature pattern corresponding to the key vector when the spatial pose corresponding to the query vector is abnormal, a mapping relationship between radar feature abnormalities and video feature abnormalities is established, thereby encoding medical visual semantics into the radar feature processing logic.

[0062] In one feasible implementation, attention is calculated as follows:

[0063] in, Represents the query vector. These represent the key vector and the value vector, respectively.

[0064] This design forces the model to use physiological micro-motion anomalies captured by radar (as a "evidence" library, K / V) as a benchmark during training to "question" and activate corresponding spatial pose anomaly hypotheses (Q) in the video. Only when a suspected anomaly detected in the video (such as mouth corner deviation) can find corresponding radar evidence (such as facial micro-motion anomalies) will the video feature be enhanced, thereby effectively suppressing misjudgments caused by changes in video lighting.

[0065] In this embodiment of the invention, during model initialization, a medical attention weight matrix is ​​constructed based on clinical knowledge (e.g., facial asymmetry is a stronger indicator of cerebral infarction). Differentiated initial weights and regularization constraints are assigned to different feature channels to guide the model to focus on pathology-related features. Specifically, feature channels strongly correlated with cranial abnormalities are assigned higher initial weights, while weakly correlated feature channels are assigned lower initial weights.

[0066] In practice, the home-based device loads a pre-trained brain abnormality risk assessment model. At this point, the brain abnormality risk assessment model only uses the radar feature input branch. The geodesic distance-invariant feature vector extracted in step 12 is then used... Input the pre-trained brain abnormality risk assessment model. Since the model has established a strong mapping from radar features to fused features through asymmetric attention during training, the corresponding pathological pattern recognition neurons in the model can be activated solely based on radar features, thus outputting the abnormality risk probability without any video data. The camera remains off, only briefly activating for verification in rare cases where the model triggers a high-susceptibility alarm (e.g., the radar shows an abnormality probability greater than 30% for three consecutive frames) and with user authorization. The camera immediately shuts down after verification, fundamentally balancing accuracy and privacy.

[0067] In this embodiment of the invention, a dynamic risk-sensitive penalty term is introduced into the objective function of the brain abnormality risk assessment model training to enhance the attention to high-risk samples during backpropagation.

[0068] Specifically, the coefficient of the penalty term is positively correlated with the model's misclassification rate for samples of different risk levels on the validation set. If the model misses a high-risk sample (a real precursor to stroke), the system automatically increases the penalty weight for that category. Through this performance feedback loop, the model is forced to prioritize recall of high-risk abnormal patterns even when only radar input is available.

[0069] The training process of the craniocerebral abnormality risk assessment model in this embodiment of the invention is described below, specifically including offline training and online learning.

[0070] During offline training, multiple pre-stroke posture samples are simultaneously collected using millimeter-wave radar and cameras. Hardware alignment is ensured through the GPIO interface, and end-to-end training is performed on the server side to establish a strong mapping relationship from "radar temporal micro-motion features" to "video spatial posture features". This enables the model to learn to accurately reconstruct abnormal human postures even without video input, based solely on minute changes in body surface folding detected by radar.

[0071] During online learning, after the model is distributed to the home, incremental training is performed to adapt to individual differences. Specifically, this includes employing Elastic Weight Consolidation (EWC) and knowledge distillation techniques. When new valid data is added locally (meeting a coverage rate of ≥90% for 12 types of stroke-related features), the EWC penalty coefficient is used. Limit the fluctuation of the weights of core medical features (fluctuations need to be limited). The home edge devices only upload encrypted parameter differences. The server uses a three-dimensional contribution gating mechanism (data reliability, task performance, coverage and generalization) to weight and aggregate the models from each home edge, and then updates the global model.

[0072] Step 14: Based on the abnormal risk probability output by the intracranial abnormality risk assessment model, obtain the intracranial abnormality risk assessment result of the monitored subject.

[0073] In one feasible implementation, the risk assessment results for intracranial abnormalities include a warning level. The model ultimately outputs a risk probability value and classifies it into low, medium, and high risk warning levels based on preset thresholds. In practice, the home-based device performs different operations according to the level: for low-risk levels, it pushes text summaries and data charts to associated terminals, presenting the health status of the monitored object in a concise and clear manner. For example, it sends a daily health summary at a fixed time each day, including the normal range of core parameters, explanations of the absence of abnormal signs, and comparisons with historical data, making it convenient for associated terminals to view without pop-up notifications or other intrusive reminders.

[0074] When a medium-risk level is reached, a ringtone and pop-up notification will be triggered to alert the associated terminal, along with a screenshot of the abnormal data. The guardian's app will receive a pop-up and ringtone alert, displaying the specific risk situation and an attached screenshot of the abnormal data; the local terminal at home will display the cause of the risk and a manual feedback button on the touchscreen through a flashing yellow LED, intermittent buzzer beeps, and the touchscreen, allowing the associated terminal to be promptly informed and take initial action based on the situation.

[0075] If the risk level is high, a mandatory pop-up and continuous ringing will be sent to the associated terminals. The guardian's terminal will receive the mandatory pop-up and continuous ringing, displaying high-risk emergency alarm information, including specific abnormal parameters and risk probability. After the mandatory pop-up and continuous ringing are triggered at a high-risk level, the associated terminal needs to respond. If the guardian's terminal does not respond, the community medical terminal will receive a platform pop-up and voice broadcast, simultaneously pushing a real-time stream of abnormal parameters. The home terminal will rapidly flash a red LED and continuously sound an alarm (the sound can be manually turned off), and automatically dial the guardian's phone. If no one answers within 30 seconds, it will transfer to the community medical duty phone, and the touchscreen will display a prompt to wait for assistance. These measures ensure that the monitored individuals can receive timely emergency medical assistance in high-risk situations, forming a complete closed loop from risk identification to emergency response. This ensures that monitored individuals receive corresponding attention and treatment at different levels of stroke risk, effectively improving the timeliness and effectiveness of stroke risk monitoring and response.

[0076] Example 2 This embodiment focuses on describing the federated learning aggregation mechanism used for parameter updates in the described intracranial abnormality risk assessment model. This mechanism runs on a cloud server and aims to securely optimize the global model using data from multiple home devices.

[0077] refer to Figure 2 The federated learning aggregation mechanism includes steps 21 to 23.

[0078] Step 21: For each household edge device, calculate the contribution score from three dimensions: data reliability, task performance, and coverage generalization.

[0079] Specifically, through the calculation formula

[0080] Get home devices Data reliability score The data reliability score is used to assess the cleanliness and reliability of home-based devices. Indicates the weighting coefficient. This represents the Sigmoid activation function. Indicates home terminal device The amount of data, Indicates home terminal device signal-to-noise ratio, Indicates home terminal device Data missing rate, This indicates the device number for the home device; generally, data with a high signal-to-noise ratio and low missing rate is more reliable.

[0081] Through calculation formula

[0082] Get home devices Task performance score ;in, Indicates the weighting coefficient. Indicates home terminal device The AUC value on the validation set, Indicates home terminal device Recall rate on the validation set, Indicates home terminal device False positive rate on the validation set; in practical implementation The setting is set to a larger value (e.g., 2.0), which severely punishes high false alarm rates, meeting the extremely high security requirements of medical scenarios.

[0083] Through calculation formula

[0084] Get home devices Coverage generalization score ;in, Indicates the weighting coefficient. Indicates home terminal device The diversity of coverage scenarios, Indicates home terminal device The population coverage scarcity indicator corresponding to the monitored subjects. This is used to penalize home devices that deviate too much from the global data distribution, preventing the negative impact of extreme outliers.

[0085] Home devices Contribution score Data reliability score Task performance score and coverage generalization score The weighted average is obtained.

[0086] Step 22: Apply a hard gating mechanism to filter data sources that do not meet the preset quality threshold, and apply a soft gating mechanism to assign protective weights to the data of home edge devices identified as high-risk groups.

[0087] In this embodiment of the invention, the hard gating adopts a "one-vote veto system" to completely exclude "bad ends" with extremely poor data quality or extremely unstable model performance from aggregation, thereby preventing the global model from being polluted.

[0088] In practical implementation, the hard gating mechanism includes: when the false alarm rate of the local intracranial abnormality risk assessment model on the home terminal device exceeds a preset false alarm rate tolerance threshold, or the data signal-to-noise ratio is lower than a preset acceptable signal-to-noise ratio threshold, the aggregation weight of the home terminal device is reset to zero. Once the hard gating is triggered, the aggregation weight of the home terminal device will be directly set to zero, and its model parameters will not be included in the global model update.

[0089] The soft gating mechanism includes: identifying whether a home device belongs to a high-risk group; if it belongs to a high-risk group and the model recall rate is lower than the preset value, a protective coefficient is introduced to increase the aggregation weight of the home device.

[0090] In practice, before implementing soft gating, the system first calculates a high-risk index using basic tags from the home end. This determines whether the household belongs to a high-risk group. One feasible implementation method is... .when At that time, the household was identified as a high-risk group. The soft gating proposed in this invention aims to protect the data value of high-risk groups. For these groups, even if the performance of their local models fluctuates in the short term (such as low recall), a "protective coefficient" will be used to ensure that their features are not overwhelmed by large amounts of data.

[0091] In this embodiment of the invention, the value of the protection coefficient is positively correlated with the proportion of high-risk individuals in the currently participating households. In one feasible implementation, the protection coefficient... This ensures that the system can automatically increase attention when there are many "difficult samples".

[0092] Ultimately, the overall gating value at each home device end. Through calculation formula This is obtained. Specifically, when the false alarm rate of the local intracranial abnormality risk assessment model on the home terminal device exceeds the preset false alarm rate tolerance threshold, or the data signal-to-noise ratio is lower than the preset acceptable signal-to-noise ratio threshold, ;otherwise, When the home device is identified as belonging to a high-risk group, and the model recall rate is lower than the preset value, ;otherwise, .

[0093] Step 23: Based on the contribution score and gating results, dynamically calculate the aggregate weight of each household edge device and perform a weighted average to generate a new generation global model.

[0094] Specifically, through the calculation formula Obtain home edge devices Aggregate weight . Indicates home terminal device The contribution score.

[0095] The server is based on weight. The model parameters of each household edge device are weighted and averaged to generate a new generation of global model, which is then distributed to each household edge device to update its local model.

[0096] This aggregation mechanism enables a shift from "data volume-driven" to "data value and risk-driven." Hard gating ensures the model's safety baseline, soft gating maintains medical equity, and 3D assessment comprehensively and meticulously mines the value of distributed data, allowing the global model to continuously and stably improve the early warning accuracy for various population groups (especially high-risk groups) while controlling false alarms.

[0097] It should be noted that traditional average aggregation (FedAvg) often masks the characteristics of a few high-risk patients with normal data from a large number of healthy individuals. Soft gating, by introducing a protective coefficient, improves the recall rate for identifying prodromal symptoms of stroke in high-risk populations to 91.3%, significantly higher than the 72% of previous methods. If the global model accuracy improvement after aggregation is less than expected (less than 3%), it can be configured to trigger a rollback and automatically adjust the gating parameters for re-aggregation.

[0098] Example 3 refer to Figure 3 This embodiment provides a brain abnormality risk assessment device 300 that implements the above method. This device can be deployed in a home setting and includes: The millimeter-wave radar module 310 is used to collect radar point cloud data of the monitored object.

[0099] The memory 320 is used to store computer programs and a trained and updated risk assessment model for cranial abnormalities as described in Example 1.

[0100] The processor 330 is electrically connected to the memory 320 and the millimeter-wave radar module 310. When the processor 330 executes the computer program, it extracts geodesic distance-invariant feature vectors, loads and runs the cranial abnormality risk assessment model for inference, and controls the early warning module 340 based on the results.

[0101] The early warning module 340 is used to output early warning information when the risk assessment results of craniocerebral abnormalities meet preset triggering conditions.

[0102] Optionally, the device may also include a communication module 350 for interacting with a cloud server to complete model download and parameter upload.

[0103] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that, for ease of description and brevity, the division of the above-mentioned functional units and modules is only used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0104] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0105] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation, characterized in that, The method includes: Acquire millimeter-wave radar point cloud data of the monitored object; Topological structure analysis is performed on the millimeter-wave radar point cloud data to extract geodesic distance invariant feature vectors; the geodesic distance invariant feature vectors characterize the relative topological relationships and dynamic micro-motion information between key parts of the human body surface; The geodesic distance invariant feature vector is input into a pre-constructed craniocerebral abnormality risk assessment model. The craniocerebral abnormality risk assessment model is constructed based on an asymmetric cross-modal attention mechanism. During the training phase, the model establishes a mapping relationship from millimeter-wave radar temporal vital signs to video spatial attitude features, while during the current inference phase, it only performs calculations based on millimeter-wave radar data. Based on the abnormal risk probability output by the cranial abnormality risk assessment model, the cranial abnormality risk assessment result of the monitored object is obtained; the cranial abnormality risk assessment result includes the warning level.

2. The method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation according to claim 1, characterized in that, The topological analysis of the millimeter-wave radar point cloud data, and the extraction of geodesic distance-invariant feature vectors, includes: Human body point set is extracted from the millimeter-wave radar point cloud, and the human body point set is mapped to a spherical coordinate system centered on the radar to generate a body surface depth raster map. The surface depth raster map is regarded as a graph structure. The raster is used as a node. The edge weights are calculated based on the angle difference and depth difference between the nodes. The local geodesic path map of the body surface is constructed based on the edge weights. Key points are selected on the local geodesic path map of the body surface based on curvature, density, and short-term stability indices; these key points correspond to key anatomical locations of the human body. The shortest path algorithm is used to calculate the geodesic distance between all key plotted point pairs, and a geodesic distance matrix is ​​constructed and normalized. The left and right difference features and short-time window micro-motion features are calculated based on the normalized geodesic distance matrix, and then concatenated to generate the geodesic distance invariant feature vector.

3. The method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation according to claim 2, characterized in that, The calculation of edge weights based on the angle and depth differences between nodes includes: Through calculation formula Obtaining border rights ;in, and Representing adjacent nodes respectively Angular difference in spherical coordinates This represents the depth value corresponding to the node. This represents the deformation sensitivity coefficient, used to adjust the weight of the impact of depth changes on distance calculation. This coefficient is dynamically adjusted based on the frequency of environmental occlusion.

4. The method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation according to claim 1, characterized in that, The asymmetric cross-modal attention mechanism includes: The brain abnormality risk assessment model includes a video feature extraction branch and a radar feature extraction branch; the video feature extraction branch is used to extract video features from video data; the radar feature extraction branch is used to extract radar features from the millimeter-wave radar point cloud data. During the model training phase, cross-modal feature fusion is performed using the RV-Transformer model; wherein, the video features are mapped to query vectors, and the radar features are mapped to key vectors and value vectors; By forcing the cranial abnormality risk assessment model to learn the radar micro-motion feature pattern corresponding to the key vector when the spatial pose corresponding to the query vector is abnormal, an attention mechanism is used to establish a mapping relationship from radar feature abnormality to video feature abnormality, thereby encoding medical visual semantics into the radar feature processing logic.

5. The method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation according to claim 1, characterized in that, The parameter updates of the craniocerebral abnormality risk assessment model employ a federated learning aggregation mechanism, which includes: For each household edge device, the contribution score is calculated from three dimensions: data reliability, task performance, and coverage generalization. Hard gating mechanisms are applied to filter data sources that do not meet preset quality thresholds, while soft gating mechanisms are applied to assign protective weights to data from home edge devices identified as high-risk groups. Based on the contribution scores and gating results, the aggregate weights of each household edge device are dynamically calculated and weighted averaged to generate a new generation of global model.

6. The method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation according to claim 5, characterized in that, The hard gating mechanism includes: when the false alarm rate of the local cranial abnormality risk assessment model of the home terminal device is higher than the preset false alarm rate tolerance threshold, or the data signal-to-noise ratio is lower than the preset signal-to-noise ratio qualified threshold, the aggregation weight of the home terminal device is reset to zero. The soft gating mechanism includes: identifying whether a home device belongs to a high-risk group; if it belongs to a high-risk group and the model recall rate is lower than a preset value, then a protective coefficient is introduced to increase the aggregation weight of the home device.

7. The method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation according to claim 5, characterized in that, The contribution score is calculated from three dimensions: data reliability, task performance, and coverage generalization, including: Through calculation formula Get home devices Data reliability score The data reliability score is used to assess the cleanliness and reliability of home-based devices. Indicates the weighting coefficient. This represents the Sigmoid activation function. Indicates home terminal device The amount of data, Indicates home terminal device signal-to-noise ratio, Indicates home terminal device Data missing rate, Indicates the device number for the home device; Through calculation formula Get home devices Task performance score ;in, Indicates the weighting coefficient. Indicates home terminal device The AUC value on the validation set, Indicates home terminal device Recall rate on the validation set, Indicates home terminal device False positive rate on the validation set; Through calculation formula Get home devices Coverage generalization score ;in, Indicates the weighting coefficient. Indicates home terminal device The diversity of coverage scenarios, Indicates home terminal device The population coverage scarcity indicator corresponding to the monitored subjects. This is used to penalize home devices that deviate too much from the global data distribution, preventing the negative impact of extreme outliers; Home devices Contribution score Data reliability score Task performance score and coverage generalization score The weighted average is obtained.

8. The method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation according to claim 2, characterized in that, The selection strategy for the key plotting points is as follows: Through calculation formula Obtain key points ;in, Indicates a predefined search area. Represents a node curvature, Represents a node Point density, Indicates short-term stability index, This represents the weighting coefficient of each indicator.

9. The method for assessing the risk of cranial abnormalities based on transfer learning and posture estimation according to claim 1, characterized in that, The initialization and training process of the craniocerebral abnormality risk assessment model includes: A medical attention weight matrix was constructed, and based on prior knowledge of clinical medicine, higher initial weights were assigned to feature channels that were strongly correlated with cranial abnormalities, and lower initial weights were assigned to feature channels that were weakly correlated. A dynamic risk-sensitive penalty term is introduced into the objective function of model training to enhance the attention to high-risk samples during backpropagation; the coefficient of the penalty term is positively correlated with the model's misclassification rate for samples of different risk levels on the validation set.

10. A device for assessing the risk of intracranial abnormalities, characterized in that, include: Millimeter-wave radar modules are used to collect radar point cloud data of the monitored objects; Memory, used to store computer programs and brain abnormality risk assessment models; A processor, electrically connected to the memory and the millimeter-wave radar module, is used to execute the computer program to implement the steps of the method as described in any one of claims 1 to 9; The early warning module is used to output early warning information when the risk assessment results of craniocerebral abnormalities meet preset triggering conditions.

Images