Fall risk assessment method and system based on individualized information
By acquiring multi-channel plantar pressure data and demographic information, and combining threshold prediction and deep feature extraction for individualized fusion, the problems of poor individual adaptability and low feature extraction resolution in existing technologies are solved, and accurate fall risk assessment is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2026-04-27
- Publication Date
- 2026-06-26
AI Technical Summary
Existing fall risk assessment methods are not robust enough to adapt to individuals of different weights and ages, have low spatial resolution of feature extraction, and cannot achieve continuous daily monitoring and personalized assessment.
By acquiring multi-channel plantar pressure time-series data and demographic information of subjects, individualized feature fusion is performed using a threshold prediction sub-network and a deep feature extraction network. Combined with a time-series coding model, support phase detection, gait cycle segmentation, and multi-dimensional feature extraction are achieved. An adaptive training strategy is used for fall risk assessment.
It achieves accurate fall risk assessment for different individuals, solves the problems of poor adaptability of fixed thresholds and low feature extraction resolution, and can capture gait deterioration trends and provide personalized assessment results.
Smart Images
Figure CN122074969B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical assessment and artificial intelligence technology, specifically a fall risk assessment method and system based on individualized information. Background Technology
[0002] With the acceleration of global aging, falls have become the leading health risk factor threatening the health of the elderly population.
[0003] Traditional fall risk assessment methods are mainly divided into three categories. The first category is clinical scale assessment methods, such as the commonly used Berg Balance Scale and the Timed Stand-Up-Walk Test. These methods rely on the subjective rating of clinicians, and the reliability varies greatly between different assessors. Moreover, they can only be assessed once and cannot achieve continuous daily monitoring or capture gait changes in daily activities.
[0004] The second category is objective evaluation methods based on laboratory equipment, such as 3D motion capture systems and force tables. These methods can achieve high-precision gait analysis, but the equipment costs are extremely high and the operation is complex. They can only be used in professional laboratory environments and cannot be applied to everyday home monitoring scenarios.
[0005] The third category is assessment methods based on portable devices. These methods attempt to collect daily gait data using portable devices to assess the risk of falls. However, these methods have many technical shortcomings that are difficult to solve in practical applications.
[0006] First, the fixed threshold of existing methods has poor robustness. The globally fixed support phase threshold has extremely poor adaptability for subjects of different weights and ages, resulting in large errors in support phase detection. Second, the feature extraction spatial resolution of existing methods is extremely low. Most methods directly compress multi-channel plantar pressure data into pressure center trajectory, masking local pressure abnormalities in the plantar surface. Summary of the Invention
[0007] The purpose of this invention is to propose a fall risk assessment method based on individualized information, comprising the following steps:
[0008] The data acquisition step involves acquiring multi-channel plantar pressure time-series data of the subjects, as well as the demographic characteristics of the subjects, including age, gender, weight, and body mass index.
[0009] The support phase detection step involves determining a support phase detection threshold for the subject based on the demographic characteristics information using a threshold prediction subnetwork, and then detecting the support phase boundary in the plantar pressure time series data based on the threshold. The support phase boundary is the time interval between foot contact with the ground.
[0010] The individualized feature fusion step involves cascading and fusing the demographic feature information with the feature map of the current layer during each convolutional processing step of the deep feature extraction, thereby achieving feature extraction for the subject.
[0011] The gait cycle segmentation step involves segmenting the plantar pressure time series data into gait cycles to obtain standardized gait cycle data.
[0012] The feature extraction step involves extracting manual feature vectors and deep feature vectors from the standardized gait cycle data, respectively.
[0013] The temporal coding step involves assembling the features of all gait cycles of the same subject into a temporal sequence, which is then input into a temporal model for temporal coding.
[0014] The risk assessment step, based on the coded features, outputs the fall risk assessment results for the subject.
[0015] This invention also proposes a fall risk assessment system based on individualized information, comprising:
[0016] The data acquisition unit is used to collect multi-channel plantar pressure time-series data from the subjects;
[0017] Individual information storage unit, used to store the subject's demographic characteristics;
[0018] A threshold prediction unit is used to determine the support period detection threshold based on the demographic characteristics information.
[0019] The period segmentation unit is used to segment the plantar pressure time series data into gait periods to obtain standardized gait period data.
[0020] The feature extraction unit is used to extract manual feature vectors and deep feature vectors from standardized gait cycle data, respectively. In each convolutional processing layer of deep feature extraction, the demographic feature information is cascaded and fused with the feature map of the current layer.
[0021] The temporal coding unit is used to combine the features of all gait cycles of the same subject into a temporal sequence, which is then input into the temporal model for temporal coding.
[0022] An assessment output unit is used to output the fall risk assessment result of the subject based on the encoded features.
[0023] The technical solution of the present invention brings at least the following beneficial effects:
[0024] First, multi-channel plantar pressure time-series data and individual information of the subjects are obtained. Then, through a five-layer progressive processing, the problems of adaptive detection of the support phase, accurate segmentation of gait cycle, multi-dimensional feature extraction, temporal dynamic modeling, and adaptive model training are solved respectively, and finally, accurate individualized fall risk assessment is achieved.
[0025] By employing adaptive threshold detection through joint training, the problem of poor robustness of fixed thresholds is completely solved. Thresholds are automatically adapted for subjects with different physiological characteristics. Through periodic dual-stream temporal modeling, the trend of gait deterioration caused by fatigue is captured. Through cascaded individualized information fusion layer by layer, adaptive personalized feature learning is achieved. Through multi-scale convolution and adaptive channel attention mechanism, key stress areas are automatically focused on, avoiding local anomalies being masked by global averaging. By adopting a training strategy that divides the training according to the subjects independently, data leakage is avoided from the root, ensuring the rigor of model performance evaluation. Attached Figure Description
[0026] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0027] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0028] Please see Figure 1 This application provides a fall risk assessment method based on individualized information, comprising the following steps:
[0029] The data acquisition step involves acquiring multi-channel plantar pressure time-series data of the subjects, as well as the demographic characteristics of the subjects, including age, gender, weight, and body mass index.
[0030] The support phase detection step involves determining a support phase detection threshold for the subject based on the demographic characteristics information using a threshold prediction subnetwork, and then detecting the support phase boundary in the plantar pressure time series data based on the threshold. The support phase boundary is the time interval between foot contact with the ground.
[0031] The individualized feature fusion step involves cascading and fusing the demographic feature information with the feature map of the current layer during each convolutional processing step of the deep feature extraction, thereby achieving feature extraction for the subject.
[0032] The gait cycle segmentation step involves segmenting the plantar pressure time series data into gait cycles to obtain standardized gait cycle data.
[0033] The feature extraction step involves extracting manual feature vectors and deep feature vectors from the standardized gait cycle data, respectively.
[0034] The temporal coding step involves assembling the features of all gait cycles of the same subject into a temporal sequence, which is then input into a temporal model for temporal coding.
[0035] The risk assessment step, based on the coded features, outputs the fall risk assessment results for the subject.
[0036] As an optional embodiment, the threshold prediction subnetwork is an independent fully connected subnetwork. Its input is normalized subject demographic information, and its output is normalized support period detection threshold. The threshold prediction subnetwork is jointly trained end-to-end with the deep feature extraction subnetwork and the subsequent classification subnetwork. The output of the threshold prediction subnetwork is directly used as the input for support period detection, and the result of support period detection is directly input into the subsequent feature extraction subnetwork.
[0037] It should be noted that the entire network is optimized end-to-end using the loss function of the classification task, without the need for separate threshold labeling.
[0038] It should also be noted that the normalized support phase detection threshold refers to the threshold being normalized to a range of 0 to 1, corresponding to the normalized range of total pressure, ensuring that the threshold can be adapted to the pressure range of subjects with different weights.
[0039] End-to-end joint training: This refers to all sub-networks of the entire network, including the threshold prediction sub-network, feature extraction sub-network, and classification sub-network, being optimized together using the same loss function without the need for separate labeling.
[0040] The specific calculation steps of the threshold prediction subnetwork are as follows: The input is the normalized 4-dimensional demographic features, namely age, gender, weight, and body mass index. First, it passes through the first fully connected layer, which maps the 4-dimensional features to 16-dimensional features. The activation function is ReLU. Then, it passes through the second fully connected layer, which maps the 16-dimensional features to 1-dimensional features. The activation function is Sigmoid. Finally, the normalized support period detection threshold is output.
[0041] The specific process of joint training is as follows: During the model training phase, the output of the threshold prediction sub-network is used as the input of the support period detection. Then, the result of the support period detection is input into the subsequent feature extraction sub-network. The loss function of the entire network adopts the cross-entropy loss function. The parameters of all sub-networks are updated through backpropagation using this loss function. This end-to-end joint training does not require separate threshold labeling.
[0042] As an optional embodiment, the gait period segmentation includes:
[0043] Calculate the total pressure time-series curves for the left and right feet respectively, and apply a moving average filter to the total pressure curves. The size of the filter window is determined based on the sampling frequency.
[0044] The local minimum point of the total pressure curve is detected as the critical point for gait segmentation. The minimum segmentation interval is determined based on the average gait cycle of the subjects, and the minimum spurt of the segmentation is determined based on the pressure fluctuation characteristics of the subjects.
[0045] The cycles of the left and right feet are paired according to the time sequence, and all cycles are interpolated to a length that matches the average gait cycle.
[0046] It should be noted that the window size of the moving average filter is specifically 1 / 10 of the sampling frequency, which is a 100ms window. This is automatically calculated based on the sampling frequency. The filtering calculation method is as follows: for each time point t, take all pressure values from t50ms to t+50ms, calculate the average value, and use it as the filtered pressure value.
[0047] The minimum interval is specifically 0.6 times the average gait period, which ensures that the interval between two critical points will not be less than 0.6 times the average period, thus avoiding false detections.
[0048] The minimum spuriousness is specifically 1.5 times the standard deviation of the pressure fluctuation. In other words, only when the spuriousness of the minimum value is greater than 1.5 times the pressure fluctuation will it be judged as the critical point.
[0049] The pairing of left and right foot cycles is specifically in chronological order. The first right cycle after the left cycle is the corresponding pairing cycle, which is the natural order of gait.
[0050] The interpolation length is specifically the length of the average gait cycle of all cycles. All cycles are interpolated to this length to ensure consistent input size. The interpolation is calculated using linear interpolation, which interpolates the original sequence of length L to a sequence of length L_avg.
[0051] As an optional embodiment, the handcrafted feature vector includes:
[0052] For the pressure sequence of each channel, extract the peak pressure, average pressure, and pressure-time integral to form the basic static features;
[0053] Extract the maximum pressure gradient, minimum pressure gradient, and full width at half maximum (FWHM) to form a pressure change rate feature;
[0054] Extract pressure transfer time and transfer rate to form time-series event features;
[0055] Extract the load ratio of forefoot and hindfoot, the load ratio of medial and lateral sides, and the symmetry index of left and right feet to form the load distribution and asymmetry characteristics;
[0056] All features are concatenated into a handcrafted feature vector for that period.
[0057] It should be noted that peak pressure is the maximum value of the pressure sequence for each channel.
[0058] Average pressure: The average pressure sequence for each channel.
[0059] Pressure-time integral: The integral of the pressure sequence for each channel, which is the sum of the pressure values at all time points.
[0060] Maximum pressure gradient: The maximum value of the first difference of the pressure sequence.
[0061] Minimum pressure gradient: The minimum value of the first difference of the pressure sequence.
[0062] Full width at half maximum (FWHM): The length of the time interval in a pressure sequence where the pressure value is greater than half of the peak value.
[0063] Pressure transfer time: The time it takes for pressure to transfer from the hind foot to the forefoot.
[0064] Transfer rate: The speed at which pressure is transferred, which is the reciprocal of the pressure transfer time.
[0065] Forefoot-to-hindfoot load ratio: Total pressure in the forefoot area divided by the total pressure in the hindfoot area.
[0066] Inner-outer load ratio: Total pressure in the inner region divided by the total pressure in the outer region.
[0067] Left-right foot symmetry index: the absolute value of the difference between all features of the left and right feet, divided by the sum of the two features;
[0068] As an optional embodiment, the timing event features further include:
[0069] Based on the support period boundary, the support period is divided into the first half and the second half. The pressure change rate of the first half is calculated to obtain the pressure rise rate of the support period, and the pressure change rate of the second half is calculated to obtain the pressure fall rate of the support period.
[0070] The plantar pressure channel is divided into forefoot and hindfoot regions. The temporal change of the pressure ratio in the forefoot region is calculated to obtain the forefoot and hindfoot pressure transfer characteristics.
[0071] Based on the pressure values of all channels, the temporal coordinates of the pressure center are calculated to obtain the center of gravity shift characteristics.
[0072] It should be noted that the forefoot area and the hindfoot area refer to dividing the pressure channels of the sole of the foot into the forefoot area (the front half) and the hindfoot area (the back half). The forefoot area corresponds to the toes and forefoot, while the hindfoot area corresponds to the heel.
[0073] The time-series coordinates of the pressure center refer to the coordinates of the center of gravity calculated based on the pressure values of all channels. These coordinates change over time, hence they are time-series coordinates. The specific calculation method is as follows: the x-coordinate is the sum of the x-coordinates of all channels multiplied by the pressure value of that channel, divided by the total pressure. The y-coordinate is calculated similarly.
[0074] All custom terms have been defined: the rate of increase of pressure during the support period is specifically the rate of change of pressure in the first half of the support period, which is the average of the first difference of the pressure in the first half; the rate of decrease of pressure during the support period is specifically the rate of change of pressure in the second half of the support period, which is the average of the first difference of the pressure in the second half.
[0075] As an optional embodiment, the deep feature vector is extracted through parallel one-dimensional convolutional layers. The size of the convolutional layers is determined according to the gait cycle length of the subject. When the gait cycle length is greater than the average length of all cycles, a larger convolutional layer is automatically added. Convolutional layers of different sizes process the input stress sequence respectively. The features output by different convolutional layers are concatenated in the channel dimension to obtain multi-scale fused features.
[0076] It should be noted that small-sized convolutional layers refer to convolutional layers with a size of 3, which are used to capture local, short-term details of pressure changes.
[0077] Large-size convolutional layers: These refer to convolutional layers with sizes of 5, 7, and 9, used to capture global, long-term stress change patterns.
[0078] The size of the convolutional layers is automatically set to three scales of 3, 5, and 7 based on the length of the gait cycle. When the cycle length is greater than the average length of all cycles, the scale is automatically increased to 9, which is completely adaptive. The stride of each convolutional layer is 1, and the padding is the same to ensure that the temporal dimension of the output is consistent with the input. Then, the outputs of different convolutional layers are concatenated in the channel dimension to obtain multi-scale fused features.
[0079] As an optional embodiment, the deep feature extraction further includes:
[0080] The feature map obtained by convolution is subjected to global average pooling along the time dimension, and the average value of the feature values at all time points of each channel is calculated to obtain the channel description vector.
[0081] The channel description vector is concatenated with demographic feature information and then input into a fully connected layer. The number of fully connected layers is adaptively adjusted according to the number of channel dimensions to learn the attention weight for each channel.
[0082] The weights are copied along the time dimension to obtain a weight matrix of the same size as the original feature map. The weight matrix is then multiplied element-wise with the original feature map to complete channel weighting.
[0083] It should be noted that the layer adaptively adjusts based on the number of channel dimensions. When the channel dimension is less than half the number of input channels in the current layer, one fully connected layer is used; when the channel dimension is greater than or equal to half the number of input channels in the current layer, two fully connected layers are used. This is entirely adaptive. The specific calculation steps are as follows:
[0084] The input feature map has a size of T×C. First, global average pooling is performed along the time dimension to obtain a 1×C channel description vector. Then, this vector is concatenated with 4-dimensional demographic features to obtain a 1×(C+4) vector. Then, it passes through the first fully connected layer to compress the dimension to 1 / 16 of C. This ratio is adaptively adjusted according to the channel dimension. The activation function is ReLU. Then, it passes through the second fully connected layer to restore the dimension to C. The activation function is Sigmoid to obtain 1×C attention weights. Then, the weights are copied T times along the time dimension to obtain a T×C weight matrix. Then, the weight matrix is multiplied element-wise with the original feature map to complete the channel weighting.
[0085] As an optional embodiment, the timing coding includes:
[0086] The handcrafted feature vectors are reduced to the same dimension as the deep features through a fully connected layer;
[0087] The deep feature sequence and the dimensionality-reduced handmade feature sequence are respectively input into the bidirectional time series model, and the number of bidirectional time series models is adaptively adjusted according to the dimension of the features;
[0088] Take the output of the last time step as the encoding result of the two streams.
[0089] It should be noted that the adjustment is adaptive based on the feature dimension. When the feature dimension is greater than half of the input feature dimension, two independent bidirectional time series models are used; when the feature dimension is less than or equal to half of the input feature dimension, a shared bidirectional time series model is used. This is entirely adaptive. The specific calculation steps are as follows:
[0090] The bidirectional temporal model uses a bidirectional LSTM with a hidden layer dimension of 64. The input sequence size is N×D, where N is the number of gait cycles and D is the dimension of the features. The output size of the bidirectional LSTM is N×128. Then, the output of the last time step is taken as the result of temporal encoding to obtain a 128-dimensional encoding vector.
[0091] As an optional embodiment, it also includes:
[0092] Cross-validation was performed using an independent partitioning strategy based on the subjects. All subjects were divided into multiple groups, with the number of groups adaptively adjusted according to the total number of subjects. In each round, one group was used as the test set, and the rest were used as the training and validation set.
[0093] Automatically perform adaptive tuning of hyperparameters within the training and validation sets;
[0094] The encoding results are concatenated with demographic information and input into an adaptive classification layer that is jointly trained with the entire network. This classification layer is jointly trained with all subnetworks of the entire network and automatically outputs the fall risk assessment results of the subjects.
[0095] It should be noted that the training phase is divided according to the number of subjects. That is, 10% of all subjects are divided into the test set, and the remaining 90% are used as the training and validation set. The number of divisions is adaptively adjusted according to the total number of subjects. When the total number of subjects is greater than 10% of the total sample size, 10-fold cross-validation is used, that is, the subjects are randomly divided into 10 groups, one group is used as the test set in each round, and the rest are used as the training set, and this is repeated 10 times. When the total number of subjects is less than or equal to 10% of the total sample size, leave-one-out cross-validation is used, that is, one subject is used as the test set in each round, and the rest are used as the training set, and this is repeated N times. It is completely adaptive.
[0096] Hyperparameter tuning, specifically using Bayesian optimization, involves tuning hyperparameters such as learning rate and batch size within the training and validation sets. The search range is: learning rate 195 to 193, batch size 8 to 64, with the accuracy of 5-fold cross-validation as the optimization target.
[0097] This invention also proposes a fall risk assessment system based on individualized information, comprising:
[0098] The data acquisition unit is used to collect multi-channel plantar pressure time-series data from the subjects;
[0099] Individual information storage unit, used to store the subject's demographic characteristics;
[0100] A threshold prediction unit is used to determine the support period detection threshold based on the demographic characteristics information.
[0101] The period segmentation unit is used to segment the plantar pressure time series data into gait periods to obtain standardized gait period data.
[0102] The feature extraction unit is used to extract manual feature vectors and deep feature vectors from standardized gait cycle data, respectively. In each convolutional processing layer of deep feature extraction, the demographic feature information is cascaded and fused with the feature map of the current layer.
[0103] The temporal coding unit is used to combine the features of all gait cycles of the same subject into a temporal sequence, which is then input into the temporal model for temporal coding.
[0104] An assessment output unit is used to output the fall risk assessment result of the subject based on the encoded features.
[0105] It should be noted that the specific operation of layer-by-layer fusion of individualized information is as follows: the demographic characteristics of the subjects are normalized to the range of 0 and 1, and then mapped to an embedding vector adapted to the current layer through a fully connected layer. The dimension E of the embedding vector is 1 / 4 of the number of channels C of the current layer, and this ratio is adaptively adjusted according to the channel dimension. Then, for the feature map output by each convolutional layer, assuming the size of the feature map is T×C, the embedding vector is copied T times in the time dimension to obtain a T×E vector. Then, this vector is concatenated with the original T×C feature map in the channel dimension to obtain a new feature map of T×(C+E), which is used as the input of the next convolutional layer.
[0106] Working principle
[0107] Input layer: First, multi-channel plantar pressure time-series data of the subjects are acquired, as well as the subjects' demographic characteristics, including age, gender, weight, and body mass index. This information serves as the input to the entire network, providing a foundation for subsequent adaptive processing.
[0108] Adaptive Support Phase Detection Layer: To address the issue of poor robustness of fixed thresholds in existing technologies, an end-to-end jointly trained threshold prediction sub-network is used. This sub-network takes the demographic characteristics of the subject as input and automatically outputs a support phase detection threshold for that subject. Then, based on this threshold, the support phase boundary in the plantar pressure time series data is detected, achieving adaptive support phase detection for different subjects and solving the problem of poor adaptability of fixed thresholds.
[0109] Gait cycle segmentation layer: For the detected support phase data, this invention automatically calculates the average gait cycle of the subject, and then adaptively adjusts the minimum interval and minimum spurs of the segmentation according to the average cycle to automatically complete the gait cycle segmentation. Then, all cycles are interpolated to a uniform length to obtain standardized gait cycle data, providing a uniform input for subsequent feature extraction.
[0110] Multi-dimensional feature extraction layer: To address the low resolution of existing feature extraction technologies, a parallel architecture of handcrafted and deep feature extraction is adopted. The handcrafted feature extraction includes multi-dimensional handcrafted features such as basic static features, pressure change rate features, time-series event features, and load distribution features. The deep feature extraction adopts multi-scale parallel convolution to automatically capture pressure change patterns at different scales. At the same time, an adaptive channel attention mechanism is added to automatically focus on key pressure regions and avoid local anomalies being masked by global averaging. Furthermore, during the processing of each convolution layer, the demographic characteristics of the subjects are cascaded and fused with the feature map of the current layer to achieve deep personalized feature learning and solve the problem of poor generalization ability of general models.
[0111] Temporal dynamic modeling layer: In response to the problem that existing technologies cannot capture the trend of gait deterioration, this invention combines the features of all gait cycles of the same subject into a temporal sequence, inputs it into a bidirectional temporal model, automatically models the temporal changes between cycles, captures the trend of gait deterioration caused by fatigue, and realizes dynamic temporal encoding.
[0112] Adaptive Model Training Layer: To address the data leakage problem in existing technologies, this invention employs a cross-validation strategy based on independent partitioning of subjects, ensuring that the subjects in the training set and test set are completely non-overlapping. Simultaneously, adaptive tuning of hyperparameters is automatically performed within the training and validation sets. Finally, the encoding results are concatenated with demographic features and input into the adaptive classification layer, automatically outputting the subject's fall risk assessment results.
[0113] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A fall risk assessment method based on individualized information, characterized in that, Includes the following steps: The data acquisition step involves acquiring multi-channel plantar pressure time-series data of the subjects, as well as the demographic characteristics of the subjects, including age, gender, weight, and body mass index. The support phase detection step involves determining a support phase detection threshold for the subject based on the demographic characteristics information using a threshold prediction subnetwork, and then detecting the support phase boundary in the plantar pressure time series data based on the threshold. The support phase boundary is the time interval between foot contact with the ground. The individualized feature fusion step involves cascading and fusing the demographic feature information with the feature map of the current layer during each convolutional processing step of the deep feature extraction, thereby achieving feature extraction for the subject. The gait cycle segmentation step involves segmenting the plantar pressure time series data into gait cycles to obtain standardized gait cycle data. The feature extraction step involves extracting manual feature vectors and depth feature vectors from the standardized gait cycle data, respectively. The temporal coding step involves assembling the features of all gait cycles of the same subject into a temporal sequence, which is then input into a temporal model for temporal coding. The risk assessment step, based on the coded features, outputs the fall risk assessment results for the subject; The threshold prediction subnetwork is an independent fully connected subnetwork. Its input is normalized subject demographic information, and its output is normalized support period detection threshold. The threshold prediction subnetwork is jointly trained end-to-end with the deep feature extraction subnetwork and the subsequent classification subnetwork. The output of the threshold prediction subnetwork is directly used as the input for support period detection, and the result of support period detection is directly input into the subsequent feature extraction subnetwork. Characteristics of time-series events include: Based on the support period boundary, the support period is divided into the first half and the second half. The pressure change rate of the first half is calculated to obtain the pressure rise rate of the support period, and the pressure change rate of the second half is calculated to obtain the pressure fall rate of the support period. The plantar pressure channel is divided into forefoot and hindfoot regions. The temporal change of the pressure ratio in the forefoot region is calculated to obtain the forefoot and hindfoot pressure transfer characteristics. Based on the pressure values of all channels, the temporal coordinates of the pressure center are calculated to obtain the center of gravity shift characteristics. The deep feature vector is extracted through parallel one-dimensional convolutional layers. The size of the convolutional layers is determined according to the length of the subject's gait cycle. When the length of the gait cycle is greater than the average length of all cycles, a larger convolutional layer is automatically added. Convolutional layers of different sizes process the input stress sequence respectively. The features output by different convolutional layers are concatenated in the channel dimension to obtain multi-scale fusion features. The deep feature extraction also includes: The feature map obtained by convolution is subjected to global average pooling along the time dimension, and the average value of the feature values at all time points of each channel is calculated to obtain the channel description vector. The channel description vector is concatenated with demographic feature information and then input into a fully connected layer. The number of fully connected layers is adaptively adjusted according to the number of channel dimensions to learn the attention weight for each channel. The weights are copied along the time dimension to obtain a weight matrix of the same size as the original feature map. The weight matrix is then multiplied element-wise with the original feature map to complete channel weighting.
2. The fall risk assessment method based on individualized information according to claim 1, characterized in that, The gait cycle segmentation includes: Calculate the total pressure time-series curves for the left and right feet respectively, and apply a moving average filter to the total pressure curves. The size of the filter window is determined based on the sampling frequency. The local minimum point of the total pressure curve is detected as the critical point for gait segmentation. The minimum segmentation interval is determined based on the average gait cycle of the subjects, and the minimum spurt of the segmentation is determined based on the pressure fluctuation characteristics of the subjects. The cycles of the left and right feet are paired according to the time sequence, and all cycles are interpolated to a length that matches the average gait cycle.
3. The fall risk assessment method based on individualized information according to claim 2, characterized in that, The handcrafted feature vector includes: For the pressure sequence of each channel, extract the peak pressure, average pressure, and pressure-time integral to form the basic static features; Extract the maximum pressure gradient, minimum pressure gradient, and full width at half maximum (FWHM) to form a pressure change rate feature; Extract pressure transfer time and transfer rate to form time-series event features; Extract the load ratio of forefoot and hindfoot, the load ratio of medial and lateral sides, and the symmetry index of left and right feet to form the load distribution and asymmetry characteristics; All features are concatenated into a handcrafted feature vector for that period.
4. The fall risk assessment method based on individualized information according to claim 1, characterized in that, The timing coding includes: The handcrafted feature vectors are reduced to the same dimension as the deep features through a fully connected layer; The deep feature sequence and the dimensionality-reduced handmade feature sequence are respectively input into the bidirectional time series model, and the number of bidirectional time series models is adaptively adjusted according to the dimension of the features; Take the output of the last time step as the encoding result of the two streams.
5. The fall risk assessment method and system based on individualized information according to claim 4, characterized in that, Also includes: Cross-validation was performed using an independent partitioning strategy based on the subjects. All subjects were divided into multiple groups, with the number of groups adaptively adjusted according to the total number of subjects. In each round, one group was used as the test set, and the rest were used as the training and validation set. Automatically perform adaptive tuning of hyperparameters within the training and validation sets; The encoding results are concatenated with demographic information and input into an adaptive classification layer that is jointly trained with the entire network. This classification layer is jointly trained with all subnetworks of the entire network and automatically outputs the fall risk assessment results of the subjects.
6. A fall risk assessment system based on individualized information, applicable to the fall risk assessment method based on individualized information as described in any one of claims 1 to 5, characterized in that, include: The data acquisition unit is used to collect multi-channel plantar pressure time-series data from the subjects; Individual information storage unit, used to store the subject's demographic characteristics; A threshold prediction unit is used to determine the support period detection threshold based on the demographic characteristics information. The period segmentation unit is used to segment the plantar pressure time series data into gait periods to obtain standardized gait period data. The feature extraction unit is used to extract manual feature vectors and deep feature vectors from standardized gait cycle data, respectively. In each convolutional processing layer of deep feature extraction, the demographic feature information is cascaded and fused with the feature map of the current layer. The temporal coding unit is used to combine the features of all gait cycles of the same subject into a temporal sequence, which is then input into the temporal model for temporal coding. An assessment output unit is used to output the fall risk assessment result of the subject based on the encoded features.