A method and system for early warning of fall risk of the elderly based on motion state change and physiological response analysis

Abstract

The application discloses a kind of old person's risk warning method and system before falling based on motion state change and physiological response analysis, comprising: S1: the motion data and physiological data of target old person are acquired in real time;S2: after time window segmentation to motion data, motion state classification identification is carried out using pre-constructed motion state recognition model;And with motion state division as multiple motion state segments;S3: pre-constructed activity baseline is obtained, and activity baseline is used to characterize the historical normal performance of target old person;Current motion feature and current physiological response feature are extracted, and compared with corresponding activity baseline, obtain motion abnormality degree and physiological response abnormality degree, calculate the risk value before falling;S4: determine the abnormal state before falling according to the risk value before falling and carry out early warning.The application realizes the accurate determination of the abnormal state before falling of old people, can identify the abnormal performance before falling such as getting up slowly, improve pertinence, reliability and timeliness.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare, specifically to a method and system for early warning of fall risks in the elderly based on changes in movement state and physiological response analysis. Background Technology

[0002] With the increasing number of elderly people, falls have become a significant safety risk in home-based, community-based, and institutional care settings. Falls are often sudden and can easily lead to fractures, head injuries, decreased mobility, and increased subsequent care burdens. Therefore, identifying abnormal movement patterns in the elderly before a fall occurs and issuing timely alerts is a crucial issue that needs to be addressed in the field of elderly health monitoring.

[0003] Existing fall monitoring technologies for the elderly mostly rely on wearable devices to collect data such as triaxial acceleration, triaxial angular velocity, and heart rate, and determine whether a fall has occurred by using methods such as combined acceleration thresholds, angular velocity thresholds, changes in posture angles, and sudden changes in exercise intensity. For example, Chinese invention patent CN 119360547 A discloses a multi-sensor-based fall alarm system and method for the elderly. This system collects the elderly person's body data through a smart bracelet and sensor array, calculates the elderly person's acceleration modulus, angular velocity modulus, gait period, and posture angle, calculates a fall risk assessment coefficient based on this data, compares it with a preset fall risk threshold to assess the elderly person's fall risk, and triggers an alarm to notify the caregiver when the risk exceeds the threshold. However, while this type of method has a certain ability to identify the impact, posture changes, and stationary state after a fall, its judgment usually focuses on the fall event itself and is difficult to reflect early risk signs such as unsteadiness when getting up, abnormal starting, increased turning fluctuations, and increased pauses during walking.

[0004] Some existing technologies attempt to incorporate physiological data such as heart rate, blood oxygen, and pulse into fall risk assessment. However, these typically treat physiological data as additional thresholds or auxiliary features, simply concatenating them with motion data before inputting them into a classification model. Such methods fail to adequately distinguish the normal differences in physiological responses under different activity states. For example, changes in posture (sitting / standing), starting to walk, and turning themselves cause varying degrees of heart rate changes and pulse interval fluctuations. If a uniform threshold is used for assessment, normal activity responses are easily misclassified as risk states, and individual differences among the same elderly person under different activity states can be easily overlooked.

[0005] Furthermore, existing technologies suffer from insufficient adaptation to individual differences among older adults. Gait habits, standing speed, turning range, heart rate response, and pulse interval fluctuations vary significantly among different older adults. Even within the same older adult, their recent physical condition, fatigue level, and activity level can change. Using only fixed thresholds, uniform models, or static risk grading makes it difficult to accurately characterize whether a target older adult deviates from their usual performance in their current movement state, thus affecting the accuracy and timeliness of fall risk warnings.

[0006] Therefore, there is a need to provide an early warning method for fall risk in the elderly, which can promptly reflect early risk signs such as unsteady standing, abnormal starting, increased turning fluctuations, and more pauses during walking, and fully distinguish the normal differences in physiological responses under different exercise states; improve the adaptability of individual differences among the elderly, and enhance the pertinence and reliability of fall risk identification. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and solve the problem that existing fall monitoring technologies mainly focus on detection and alarm after a fall occurs and have difficulty in timely identifying abnormal states before a fall. This invention provides a method and system for early warning of fall risks in the elderly based on changes in movement state and physiological response analysis.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A method for predicting fall risk in older adults based on changes in motor state and physiological response analysis includes the following steps:

[0010] S1: Real-time acquisition of the target elderly person's movement and physiological data;

[0011] S2: After segmenting the motion data into time windows, a pre-built motion state recognition model is used to classify and identify the motion state corresponding to each time window; adjacent time windows with the same motion state category in all time windows are merged to obtain multiple motion state segments divided by motion state, wherein each motion state segment contains data corresponding to one or more time windows.

[0012] S3: For each motion state segment, extract the current motion features and the current physiological response features respectively;

[0013] A pre-constructed activity baseline is obtained, and the current motor characteristics and current physiological response characteristics are compared with the corresponding activity baseline to obtain the degree of motor abnormality and the degree of physiological response abnormality for each motor state segment. The pre-fall risk value for each motor state segment is calculated based on the degree of motor abnormality and the degree of physiological response abnormality. The activity baseline is established based on the target elderly person's historical valid data and according to different motor states and different motor characteristics and physiological response characteristics. The activity baseline is used to characterize the target elderly person's historical normal performance.

[0014] S4: Determine the pre-fall risk value based on the pre-fall risk value of each motion state segment, and / or the pre-fall risk values ​​of multiple consecutive motion state segments, and issue an early warning.

[0015] Preferably, in step S1, the motion data of the target elderly person acquired in real time includes derived motion data; the derived motion data is calculated based on the real-time collected triaxial acceleration data and triaxial angular velocity data.

[0016] The derived motion data includes resultant acceleration, resultant angular velocity, posture changes, number of steps, stride frequency, motion intensity, and pause time.

[0017] Preferably, in step S1, the physiological data of the target elderly person acquired in real time includes heart rate data and pulse interval data.

[0018] Preferably, in step S2, a temporal convolutional network is used to construct the motion state recognition model;

[0019] The temporal convolutional network includes an input layer, multiple residual convolutional modules, a global average pooling layer, a fully connected layer, and a softmax output layer connected in sequence. The input layer is used to input the motion data segmented based on a time window. The multiple residual convolutional modules are used to capture motion dependencies. Each residual convolutional module includes a one-dimensional dilated convolutional layer, a batch normalization layer, a ReLU activation layer, and a residual connection layer. The residual connection layer is used to connect the input of the current residual convolutional module and the output of the current ReLU activation layer.

[0020] Preferably, in step S3, the method for constructing the active baseline includes:

[0021] Historical monitoring data of the target elderly person is obtained, including historical exercise data and historical physiological data; the wearing status and data quality status are determined based on the historical monitoring data and its collection time, and basic validity screening is performed to obtain the historical valid data;

[0022] After segmenting the historical valid data into time windows, the motion state recognition model is used to classify and identify motion states. Adjacent time windows with the same motion state category in all time windows are merged to obtain multiple historical segments divided by motion state. Abnormal action segments in the historical segments are removed to obtain valid historical segments.

[0023] Extract motion features and physiological response features from the effective historical segments;

[0024] For the target elderly in the first The first motion state The first valid historical segment, the first of the valid historical segments The eigenvalues ​​are denoted as ;

[0025] Segment weights are calculated based on the data quality and time distance of each valid historical segment. Its expression is:

[0026] ;

[0027] in, ; Indicates the first The first motion state Data quality score for each valid historical segment. Indicates the first The first motion state The total duration of each valid historical segment This indicates the invalid collection duration within the valid historical segment. This indicates the time interval between the start or end time of the valid historical segment and the baseline calculation time. Indicates the time decay coefficient. Indicates the target elderly in the first The number of valid historical segments used to establish a baseline under motion-like conditions;

[0028] The first The first effective historical segment under the motion state The features are weighted and statistically analyzed according to the segment weights to obtain the target elderly person in the [number]th segment. The first motion state The active baseline mean of each feature Its expression is:

[0029] ;

[0030] The baseline mean activity level is used to reflect the target older adult's activity level during the period. The first motion state Historical benchmark levels for each feature;

[0031] The corresponding active baseline fluctuation range is calculated based on the segment weight and the mean of the active baseline. Its expression is:

[0032] ;

[0033] in, To prevent constants with a fluctuation range of zero;

[0034] This yields the activity baseline. .

[0035] Preferably, in step S3, the method for obtaining the degree of motion anomaly of each motion state segment includes the following steps:

[0036] For each motion state segment, extract the current motion features;

[0037] Let the motion state category to which the current motion state segment belongs be . , its first Each motion feature value is denoted as ;

[0038] The first segment within the current motion state Each motion feature is compared with its corresponding activity baseline to calculate the degree of deviation of the current motion feature. Its expression is:

[0039] ;

[0040] Set motion feature deviation threshold ,when When an abnormality occurs, it is considered a motion anomaly, and the individual motion anomaly value for the current motion feature is calculated. The expression for this value is:

[0041] ;

[0042] in, Indicates the first segment of the current motion state. Individual motion anomalies of each motion feature;

[0043] Based on the individual motion anomaly values ​​of each motion feature, calculate the degree of motion anomaly in the current motion state segment. Its expression is:

[0044] ;

[0045] in, This indicates the number of motion features in the current motion state segment.

[0046] Preferably, in step S3, the method for obtaining the degree of physiological response abnormality for each motion state segment includes the following steps:

[0047] For each motion state segment, extract the current physiological response features, which include heart rate features and pulse interval fluctuation value features;

[0048] Let the motion state category to which the current motion state segment belongs be . , No. The Hth feature in the exercise-like state is the heart rate feature, and the corresponding mean and range of activity baseline are respectively... and ;

[0049] Compare the heart rate characteristics within the current motion segment with the corresponding [number]. By comparing heart rate characteristics under exercise-like conditions with an active baseline, the heart rate elevation load within the current exercise segment can be calculated. Its expression is:

[0050] ;

[0051] in, This indicates the number of valid heart rate values ​​within the current motion segment; Indicates the first segment of the current motion state. One effective heart rate value; To stabilize heart rate, based on the current exercise state before the segment begins. Calculation of the average value of effective heart rate data within seconds; Allowable deviation for increased heart rate;

[0052] The characteristics of pulse interval fluctuation values ​​are obtained based on pulse interval data, namely, adjacent pulse interval fluctuation values. The expression is:

[0053] ;

[0054] in, This indicates the number of effective pulse intervals within the current motion segment; This represents the r-th effective pulse interval value within the current motion state segment; To prevent constants with a fluctuation range of zero;

[0055] The pulse interval fluctuation values ​​within the current motion segment With the corresponding first Adjacent fluctuations in pulse interval during exercise-like conditions Compare the activity baseline to calculate abnormal values ​​of pulse interval fluctuations. Its expression is:

[0056] ;

[0057] in, Indicates the first The mean of the activity baseline corresponding to the characteristics of pulse interval fluctuations under exercise-like conditions. Indicates the first The range of activity baseline fluctuations corresponding to the characteristics of pulse interval fluctuations under motion-like conditions;

[0058] Based on the aforementioned heart rate increase load Abnormal values ​​of pulse interval fluctuations Calculate the degree of abnormality in the physiological response of the current motion segment. Its expression is:

[0059] ;

[0060] in, It represents the combined enhancement coefficient when elevated heart rate load and abnormal pulse interval fluctuations occur simultaneously.

[0061] Preferably, in step S3, the method for calculating the pre-fall risk value for each motion state segment is as follows:

[0062] For each motion state segment, the degree of motion anomaly is normalized to obtain the normalized degree of motion anomaly. Its expression is:

[0063] ;

[0064] in, The degree of motor abnormality;

[0065] Based on the normalized degree of motion abnormality and the degree of abnormal physiological response The pre-fall risk value for the current motion segment is calculated using the following expression:

[0066] ;

[0067] in, This indicates the risk value before a fall in the current motion segment. This represents the enhancement coefficient when abnormal movement and abnormal physiological response occur simultaneously.

[0068] Preferably, in step S4, the method for determining the abnormal state before the fall and issuing an early warning includes:

[0069] The abnormal state before the fall includes low-risk, medium-risk and high-risk states, with the corresponding risk levels increasing sequentially.

[0070] when When the risk level is low, only the risk value is recorded. The low-risk threshold; when When the risk level is determined to be medium, an alert is sent. The high-risk threshold; when When the situation is deemed high-risk, an alarm command is sent.

[0071] If the target elderly person is identified as having an abnormal state before a fall in M ​​consecutive motion state segments, the risk level of the latest motion state segment will be increased by one level. If the latest motion state segment was already in a high-risk state before the increase, the high-risk state will be maintained and an alarm command will be sent.

[0072] A fall risk warning system for the elderly based on changes in motor state and physiological response analysis, used to implement the method, includes:

[0073] The data acquisition module is used to acquire the target elderly person's exercise and physiological data in real time;

[0074] The motion state recognition module is used to segment the motion data into time windows and then use a pre-built motion state recognition model to classify and recognize the motion state, so as to identify the motion state corresponding to each time window; and to merge adjacent time windows with the same motion state category in all time windows to obtain multiple motion state segments divided by motion state, wherein each motion state segment contains data corresponding to one or more time windows.

[0075] The fall risk estimation module is used to extract the current movement features and current physiological response features for each movement state segment; obtain a pre-constructed activity baseline; compare the current movement features and current physiological response features with the corresponding activity baseline to obtain the degree of movement abnormality and the degree of physiological response abnormality for each movement state segment; and calculate the pre-fall risk value for each movement state segment based on the degree of movement abnormality and the degree of physiological response abnormality. The activity baseline is established based on the target elderly person's historical valid data and according to different movement states and different movement features and physiological response features. The activity baseline is used to characterize the target elderly person's historical normal performance.

[0076] The early warning module is used to determine the abnormal state before falling and issue an early warning based on the pre-fall risk value of each motion state segment and / or the pre-fall risk values ​​of multiple consecutive motion state segments.

[0077] This invention establishes activity baselines for different movement states, collects movement and physiological data of the target elderly person through a wearable monitoring terminal, identifies the movement state of the target elderly person, and combines the degree of movement abnormality and physiological response abnormality of the current movement state segment to determine the abnormal state before a fall, thereby outputting graded reminders and notifications before a fall occurs, improving the pertinence, reliability and timeliness of fall risk identification; it solves the problem that existing fall monitoring technologies mainly focus on detection and alarm after a fall occurs and have difficulty in timely identifying abnormal states before a fall.

[0078] This invention identifies the static, sitting-standing, starting, walking, turning, and paused movement states of target elderly individuals, and calculates the degree of movement abnormality and physiological response abnormality in the current movement state segment. It can identify abnormal pre-fall signs such as slower standing, unsteady starting, increased turning fluctuations, and more frequent pauses in walking, improving the early warning capability. This invention constructs activity baselines for target elderly individuals under different movement states, comparing the current segment with their historical performance under similar movement states, reducing misjudgments due to individual differences and activity habits. By combining the degree of movement abnormality and the degree of physiological response abnormality in the calculation of pre-fall risk values, and outputting low-risk records, medium-risk local alerts, and high-risk notifications based on the risk values, it avoids relying solely on a single movement mutation or a single physiological indicator to judge risk, making the warning results more suitable for tiered management in home-based, community-based, and institutional care scenarios. Attached Figure Description

[0079] The present invention will now be described in further detail with reference to the accompanying drawings:

[0080] Figure 1 This is a method block diagram of the present invention;

[0081] Figure 2 This is a schematic diagram of the motion state recognition model of the present invention;

[0082] Figure 3 This is a system principle block diagram of the present invention. Detailed Implementation

[0083] like Figure 1As shown, this invention provides a method for early warning of fall risks in the elderly based on changes in movement state and physiological response analysis. It establishes activity baselines for different movement states, collects movement and physiological data of the target elderly person through a wearable monitoring terminal, identifies the target elderly person's movement state, and determines the abnormal state before a fall by combining the degree of movement abnormality and physiological response abnormality of the current movement state segment. This allows for the output of tiered reminders and notifications before a fall occurs, improving the targeting, reliability, and timeliness of fall risk identification. It addresses the problem that existing fall monitoring technologies mainly focus on detection and alarm after a fall occurs, making it difficult to identify abnormal states before a fall. This invention specifically includes the following steps.

[0084] S1: Real-time acquisition of motion and physiological data of the target elderly individual. The real-time acquired motion data specifically refers to derived motion data, which is calculated based on real-time collected triaxial acceleration and triaxial angular velocity data. This triaxial acceleration, triaxial angular velocity, and physiological data are collected through a wearable monitoring terminal worn by the target elderly individual.

[0085] The derived motion data includes resultant acceleration, resultant angular velocity, posture changes, number of steps, stride frequency, motion intensity, and pause time.

[0086] Specifically, derived motion data is obtained by processing triaxial acceleration and triaxial angular velocity data in units of time windows; each time window includes The nth sampling point. Let the nth sampling point be... The triaxial acceleration of each sampling point is The angular velocity of the three axes is .

[0087] Resultant acceleration We obtain it by taking the square root of the sum of the squares of the triaxial accelerations:

[0088] ;

[0089] Resultant angular velocity We obtain it by taking the square root of the sum of the squares of the three-axis angular velocities:

[0090] ;

[0091] Attitude changes are obtained from the amplitude of attitude angle changes at the wearing site within a time window: the pitch, roll, and yaw angle changes within the time window are estimated based on the triaxial acceleration direction and the triaxial angular velocity integral results, and the amplitude of attitude angle changes within the time window is used as attitude change data. For example, when the target elderly person gets up from a sitting position, the wrist direction of the wristband will change significantly with body support and arm swinging movements; when the target elderly person turns around, the wrist will produce angular velocity and yaw angle changes with torso rotation, and these changes are used as the corresponding attitude change data.

[0092] The number of steps is obtained by detecting periodic peaks in the resultant acceleration sequence. First, the resultant acceleration sequence is filtered to identify peak points that satisfy the gait periodicity condition, and the number of peak points is taken as the number of steps within that time window.

[0093] Step frequency Calculated from the number of steps and the time window length:

[0094] ;

[0095] in, Indicates the first The number of steps identified within a time window This indicates the length of the time window. In this embodiment, s.

[0096] The intensity of motion is represented by the average amplitude of the resultant acceleration within the time window, and its expression is:

[0097] ;

[0098] in, Indicates the first Motion intensity within a time window This represents gravitational acceleration. The greater the intensity of the motion, the more pronounced the motion within that time window. For example, in a stationary state... Lower, when starting, turning or walking quickly Increase.

[0099] The pause time is calculated from the duration of the interval or stillness between actions. Specifically, when the resultant acceleration corresponding to a certain sampling point is lower than the acceleration threshold and the resultant angular velocity is lower than the angular velocity threshold, the sampling point is determined to be a low-motion sampling point; the cumulative duration corresponding to consecutive low-motion sampling points is used as the pause time. In this embodiment, the acceleration threshold is set to 0.15. Angular velocity threshold is taken .

[0100] Physiological data includes heart rate data and pulse interval data. This data is used to assess the physical response of older adults under different exercise conditions, such as whether their heart rate abnormally increases after standing up, whether heart rate fluctuations significantly increase during walking, whether their heart rate remains persistently high after pausing, and whether their heart rate recovery is slow after activity.

[0101] During data acquisition, the acquisition time is recorded synchronously. Furthermore, the wearing status and data quality are determined based on motion data, physiological data, and acquisition time, and the corresponding time window data is marked.

[0102] The wearing status is determined based on a combination of heart rate and pulse interval data collection, motion signal continuity, and timestamp continuity. If heart rate and pulse interval can be continuously collected, and the motion data timestamps are continuous, it is considered normal wearing. If heart rate and pulse interval cannot be read for a continuous period, or if the motion data remains unchanged for a long time, it is considered not wearing or abnormal wearing. In this embodiment, continuous wearing is indicated when the duration exceeds a preset continuous time threshold. The preset continuous time threshold is 5 seconds. This time length is slightly longer than a single time window (4 seconds), which can filter out brief signal loss caused by momentary poor contact, while not significantly affecting the real-time performance of subsequent segment quality assessment.

[0103] Data quality status is determined by a remote server based on whether data is continuously missing, whether timestamps are continuous, whether values ​​exceed reasonable ranges, and whether signals are abnormally interrupted. Data quality status can be marked as valid, missing, or abnormal, and is used for subsequent removal of low-quality segments (data quality status below set requirements) or calculation of data quality scores. The methods described above for determining wearing status and data quality status are mature existing technologies in the field of wearable monitoring, and are not elaborated upon here.

[0104] When the wearable monitoring terminal is detected as not being worn, being worn abnormally, or having data quality below the set requirements, the data for the corresponding time period is marked for segment rejection before subsequent anomaly calculation; the data for this time period is not directly used for determining the abnormal state before the fall. This method reduces false judgments caused by improper device wearing or signal abnormalities.

[0105] Through the above data acquisition steps, continuous motion status information and physiological response information of the elderly during daily activities are obtained, providing a data foundation for subsequent motion status identification, abnormality calculation, and judgment and early warning of abnormal states before falls.

[0106] S2: After segmenting the motion data into time windows, a pre-built motion state recognition model is used to classify and identify the motion state corresponding to each time window. Adjacent time windows with the same motion state category are merged to obtain multiple motion state segments, where each motion state segment contains data corresponding to one or more time windows. By performing state recognition on the current motion data, the motion state category is obtained, denoted as... In this embodiment, there are six motion state categories: stationary, sitting / standing transition, starting, walking, turning, and paused. For example, if the current time window data is identified as turning, then... Indicates a turning state; if the current segment is identified as a change from sitting to standing, then... Indicates the change in sitting or standing status.

[0107] Specifically, the motion data of the target elderly person obtained in step S1 is segmented into time windows. Each time window is set to a length of 4 seconds, with a sampling frequency of 50Hz, and a 1-second interval between adjacent time windows. For the first... Extract motion data matrix within a time window. Where 200 represents the number of sampling points within the time window, and 7 represents the number of motion data channels. In this embodiment, the 7 motion data channels are: resultant acceleration, resultant angular velocity, attitude change, number of steps, step frequency, motion intensity, and pause time.

[0108] The motion data matrix is ​​input into a motion state recognition model for motion state classification and identification. Specifically, a temporal convolutional network is used to construct the motion state recognition model. Figure 2 As shown, the temporal convolutional network includes an input layer, multiple residual convolutional modules, a global average pooling layer, a fully connected layer, and a softmax output layer connected in sequence; the input layer is used to input motion data segmented based on a time window; the multiple residual convolutional modules are used to capture motion dependencies.

[0109] Each residual convolutional module includes a one-dimensional dilated convolutional layer, a batch normalization layer, a ReLU activation layer, and a residual connection layer, used to extract the variation features of motion amplitude, motion direction, and motion rhythm within a time window. The residual connection layer connects the input of the current residual convolutional module and the output of the current ReLU activation layer. In this embodiment, there are three residual convolutional modules: a first residual convolutional module, a second residual convolutional module, and a third residual convolutional module. The kernel size of the three corresponding one-dimensional dilated convolutional layers is 3, the dilation coefficients are 1, 2, and 4 respectively, and the number of output channels is 32, 64, and 64 respectively.

[0110] The motion state recognition model is constructed using cloud-based pre-training and continuous fine-tuning with user data. First, a basic motion state recognition network is trained using existing general wrist-worn motion data to obtain the basic model. The training data includes three-axis acceleration data and three-axis angular velocity data collected by the wrist-worn device, as well as the corresponding six motion state labels. The training objective is to classify six motion states: stationary, sitting / standing transition, starting, walking, turning, and pausing.

[0111] After the target elderly person uses the device, the remote server continuously collects their daily exercise data and uses a basic model to identify the exercise state based on the data. Based on the identification results, a personal exercise state sample is generated for that elderly person. The remote server then uses the accumulated personal exercise state samples to continuously fine-tune the basic model, gradually adapting it to the elderly person's range of motion, wearing position, and exercise habits. The fine-tuning target remains the six-category exercise state classification.

[0112] The motion state recognition model outputs the probability of the current time window belonging to different motion states. The motion state category with the highest probability is selected as the motion state of the current time window. Then, the motion states of all time windows are concatenated in chronological order to form a motion state sequence, thus forming the motion state sequence of the target elderly person.

[0113] The motion state sequence of the target elderly was divided into multiple motion state segments, each corresponding to a specific motion state. Motion state segments were obtained by merging adjacent time windows with the same motion state category within the sequence. Specifically, the time windows were traversed in chronological order of data collection; when adjacent time windows had the same motion state category, they were merged into the same motion state segment. When the motion state category changed in a later time window, the previous motion state segment ended, and a new motion state segment was established starting from that time window. Each motion state segment included the motion state category, segment start time, segment end time, and the corresponding motion and physiological data for the duration of that segment.

[0114] Through the above-mentioned motion state recognition steps, the remote server obtains fragments of the target elderly person's motion state during daily activities, providing a basis for subsequently establishing personal baselines according to different motion states, calculating the degree of motion abnormality, and judging the risk of falling.

[0115] To further improve the accuracy of subsequent abnormal state determination before falls, the movement state segments of the target elderly are screened based on the above-mentioned wearing status and data quality status marker data. Movement state segments containing time periods of no wearing, abnormal wearing, or data quality status below the set requirements are screened out to avoid misjudgment caused by improper device wearing or abnormal signal.

[0116] S3: For each motor state segment, extract the current motor features and current physiological response features; obtain a pre-constructed activity baseline; compare the current motor features and current physiological response features with the corresponding activity baseline to obtain the degree of motor abnormality and physiological response abnormality for each motor state segment; calculate the pre-fall risk value for each motor state segment based on the degree of motor abnormality and physiological response abnormality. The activity baseline is established based on the target elderly person's historical valid data and according to different motor states and different motor features and physiological response features. The activity baseline is used to characterize the target elderly person's historical normal performance.

[0117] The method for constructing an active baseline specifically includes the following steps:

[0118] Historical monitoring data of the target elderly is acquired, including historical exercise data and historical physiological data. Based on the historical monitoring data and its collection time, the wearing status and data quality are assessed, and a basic validity screening is performed to obtain valid historical data. Through this basic validity screening, invalid time periods in the historical monitoring data are removed, such as periods of no wearing, prolonged signal interruption, or data missing exceeding a preset removal threshold.

[0119] After segmenting the historical valid data into time windows, a motion state recognition model is used to classify and identify motion states. Adjacent time windows with the same motion state category are merged to obtain multiple historical segments divided by motion state. Abnormal motion segments are then removed from the historical segments to obtain valid historical segments. Abnormal motion segments refer to segments whose motion state probability values ​​obtained through motion state classification and identification are less than preset values ​​(clearly not belonging to the six types of motion states) or which have been manually marked as high-risk abnormalities, in order to avoid abnormal data affecting the normal baseline.

[0120] The study extracts motion and physiological response features from valid historical segments, as well as segment duration. Specifically, the extracted features include: segment duration, changes in resultant acceleration, changes in resultant angular velocity, amplitude of posture changes, cadence, number of steps, exercise intensity, pause time, heart rate, and pulse interval fluctuations.

[0121] For the target elderly in the first The first motion state The first valid historical segment, the first of the valid historical segments The eigenvalues ​​are denoted as For example, the first feature corresponds to the duration of the segment. The value of the first feature of this valid historical segment, "segment duration", is 4.

[0122] To ensure that the personalized activity baseline reflects the recent status of the target older adults while reducing the impact of low-quality data on the baseline, segment weights are calculated based on the data quality and temporal distance of each valid historical segment. Its expression is:

[0123] ;

[0124] in, This indicates the time interval between the start or end time of the valid historical segment and the baseline calculation time. Indicates the time decay coefficient. Indicates the target elderly in the first The number of valid historical segments used to establish a baseline under motion-like conditions; Indicates the first The first motion state Data quality score for each valid historical segment.

[0125] The expression for data quality scoring is: ;

[0126] in, Indicates the first The first motion state The total duration of each valid historical segment This indicates the invalid acquisition duration within the valid historical segment. The invalid acquisition duration is calculated for the duration of locally failed data retained within the valid historical segment, such as short-term signal loss or abnormal individual sampling points. Since the degree of failure does not reach the removal threshold of the basic validity screening mentioned above, the valid historical segment is still retained, and its weight in the construction of the active baseline is reduced by the invalid acquisition duration.

[0127] By introducing in fragment weights and time decay coefficient To ensure that the activity baseline more accurately reflects the recent condition of the target elderly individual, factors such as cadence, standing speed, turning range, and heart rate response can vary depending on recent physical condition, fatigue level, medication use, or rehabilitation progress. If all historical data segments are weighted equally in baseline construction, earlier historical data may skew the current activity baseline, causing recent normal activities to be misjudged as abnormal. Therefore, during activity baseline initialization or periodic updates, a time decay weight is set based on the time interval between the start or end time of a valid historical segment and the baseline calculation time. This gives more weight to recent valid historical segments in activity baseline construction, making them have a greater impact on the baseline, while the influence of earlier valid historical segments gradually decreases.

[0128] The first The first effective historical segment under the motion state Each feature is weighted statistically according to its segment weight to obtain the target elderly person's [number]th [feature]. The first motion state The active baseline mean of each feature Its expression is:

[0129] .

[0130] The baseline mean of activity is used to reflect the target older age at the time of activity. The first motion state Historical benchmark levels for each feature;

[0131] The corresponding active baseline fluctuation range is calculated based on the segment weights and the mean of the active baseline. Its expression is:

[0132] ;

[0133] in, To prevent a constant with a fluctuation range of zero, this embodiment sets... .

[0134] Through the above processing, activity baselines are established for the target elderly individuals under different exercise states, mathematically represented as the mean and fluctuation range. For example: what is the approximate cadence, exercise intensity, and heart rate of the target elderly person when they normally "walk"? What is the approximate change in angular velocity and posture when they normally "turn around"?

[0135] By utilizing effective historical fragments of the target elderly, the historical baseline mean and fluctuation range of each movement characteristic and physiological response characteristic are calculated according to the movement states of stillness, sitting and standing changes, starting, walking, turning and stopping. The personalized activity baseline of the target elderly is obtained, which provides a normal reference benchmark for subsequent judgment on whether a certain characteristic of the target elderly deviates from their own historical normal performance in a certain movement state.

[0136] S3 specifically includes the following steps:

[0137] S3.1: Obtain the degree of motion anomaly for each motion state segment, specifically:

[0138] For each motion state segment, extract the current motion features;

[0139] Let the motion state category to which the current motion state segment belongs be . , its first Each motion feature value is denoted as To indicate the target elderly in the current number The first motion state The current value of each motion feature.

[0140] Call the above-obtained first The first motion state Baseline mean of each feature and baseline fluctuation range The first segment of the current motion state Each motion feature is compared with its corresponding activity baseline to calculate the degree of deviation of the current motion feature. Its expression is:

[0141] .

[0142] Set motion feature deviation threshold In this embodiment, the following settings are provided: .

[0143] when When this motion characteristic is not counted as a motion abnormality; when When an abnormality occurs, it is considered a motion anomaly, and the individual motion anomaly value for the current motion feature is calculated. The expression for this value is:

[0144] ;

[0145] in, Indicates the first segment of the current motion state. Individual motion anomalies of each motion feature.

[0146] Based on the individual motion anomaly values ​​of each motion feature, calculate the degree of motion anomaly in the current motion state segment. Its expression is:

[0147] ;

[0148] in, This indicates the number of motion features in the current motion state segment.

[0149] S3.2: Obtain the degree of physiological response abnormality for each motion state segment, specifically:

[0150] For each motion state segment, the current physiological response features are extracted, including heart rate features and pulse interval fluctuation value features.

[0151] Let the motion state category to which the current motion state segment belongs be . , No. The Hth feature in the exercise-like state is the heart rate feature, and the corresponding mean and range of activity baseline are respectively... and ;

[0152] Compare the heart rate characteristics within the current motion segment with the corresponding [number]. By comparing heart rate characteristics under exercise-like conditions with an active baseline, the heart rate elevation load within the current exercise segment can be calculated. Its expression is:

[0153] ;

[0154] in, This indicates the number of valid heart rate values ​​within the current motion segment; Indicates the first segment of the current motion state. One effective heart rate value; To stabilize heart rate, based on the current exercise state before the segment begins. Calculation of the average value of effective heart rate data within seconds; To account for the allowable deviation in heart rate elevation, in this embodiment, the average of the effective heart rate data within 30 seconds prior to the start of the current motion segment is taken as the stable heart rate. A valid heart rate value refers to a heart rate value successfully collected by the wearable monitoring terminal within a specific time segment of the exercise state, with continuous timestamps, and not marked as missing or abnormal. If the heart rate data is empty, the timestamps are interrupted, or the device marks it as invalid, it will not be included in the valid heart rate value. In this embodiment, the allowable deviation of heart rate increase is taken. =5 bpm.

[0155] Based on pulse interval data, the characteristics of pulse interval fluctuation values ​​are obtained, including adjacent fluctuation values ​​of pulse intervals. The expression is:

[0156] ;

[0157] in, This indicates the number of effective pulse intervals within the current motion segment; This represents the r-th effective pulse interval value within the current motion state segment; To prevent constants with a fluctuation range of zero;

[0158] The pulse interval fluctuation values ​​within the current motion segment With the corresponding first Adjacent fluctuations in pulse interval during exercise-like conditions Compare the activity baseline to calculate abnormal values ​​of pulse interval fluctuations. Its expression is:

[0159] ;

[0160] in, Indicates the first The mean of the activity baseline corresponding to the characteristics of pulse interval fluctuations under exercise-like conditions. Indicates the first The range of activity baseline fluctuations corresponding to the pulse interval fluctuation characteristics under motion-like conditions.

[0161] Based on heart rate increase load Abnormal values ​​of pulse interval fluctuations Calculate the degree of abnormality in the physiological response of the current motion segment. Its expression is:

[0162] ;

[0163] in, This represents the combined enhancement coefficient when elevated heart rate load and abnormal pulse interval variability occur simultaneously. In this embodiment, Take 0.5.

[0164] S3.3: Calculate the pre-fall risk value for each motor state segment based on the degree of motor abnormality and the degree of physiological response abnormality, specifically as follows:

[0165] For each motion state segment, the degree of motion anomaly is normalized to obtain the normalized degree of motion anomaly. Its expression is:

[0166] ;

[0167] in, The degree of motor abnormality;

[0168] Based on the normalized degree of motion abnormality and the degree of abnormal physiological response The pre-fall risk value for the current motion segment is calculated using the following expression:

[0169] ;

[0170] in, This indicates the risk value before a fall in the current motion segment. The enhancement coefficient represents the simultaneous occurrence of abnormal movement and abnormal physiological response. In this embodiment, Take 0.3.

[0171] S4: Determine the pre-fall risk value based on the pre-fall risk value of each motion state segment, and / or the pre-fall risk values ​​of multiple consecutive motion state segments, and issue an early warning.

[0172] Specifically, the abnormal state before a fall includes low-risk, medium-risk, and high-risk states, with the corresponding risk levels increasing sequentially.

[0173] when When the risk level is low, only the risk value is recorded. The low-risk threshold is set to 0.4; when When the risk level is determined to be medium, an alert is sent. The high-risk threshold is set to 0.7; when When the situation is deemed high-risk, an alarm command is sent.

[0174] If an elderly person is identified as having an abnormal pre-fall condition in M ​​consecutive motion state segments, the risk level of the latest motion state segment is raised by one level. If the latest motion state segment was already in a high-risk state before the upgrade, it remains in a high-risk state and an alarm command is sent. For example, if three consecutive motion state segments are identified as low-risk, the latest motion state segment (most recent in the time window) is upgraded to a medium-risk state and an alert command is sent. This enables timely and effective identification and early warning of abnormal pre-fall conditions, allowing for the prompt detection of abnormal behaviors in the elderly.

[0175] This invention identifies the static, sitting-standing, starting, walking, turning, and paused movement states of target elderly individuals, and calculates the degree of movement abnormality and physiological response abnormality in the current movement state segment. It can identify abnormal pre-fall signs such as slower standing, unsteady starting, increased turning fluctuations, and more frequent pauses in walking, improving the early warning capability. This invention constructs activity baselines for target elderly individuals under different movement states, comparing the current segment with their historical performance under similar movement states, reducing misjudgments due to individual differences and activity habits. By combining the degree of movement abnormality and the degree of physiological response abnormality in the calculation of pre-fall risk values, and outputting low-risk records, medium-risk local alerts, and high-risk notifications based on the risk values, it avoids relying solely on a single movement mutation or a single physiological indicator to judge risk, making the warning results more suitable for tiered management in home-based, community-based, and institutional care scenarios.

[0176] like Figure 3 As shown, the present invention also provides a fall risk warning system for the elderly based on changes in movement state and physiological response analysis, comprising:

[0177] The data acquisition module is used to acquire the target elderly person's exercise and physiological data in real time;

[0178] The motion state recognition module is used to classify and recognize motion states by segmenting motion data into time windows and then using a pre-built motion state recognition model to identify the motion state corresponding to each time window. It merges adjacent time windows with the same motion state category in all time windows to obtain multiple motion state segments divided by motion state. Each motion state segment contains data corresponding to one or more time windows.

[0179] The fall risk estimation module extracts current motor features and current physiological response features for each motor state segment; it obtains a pre-constructed activity baseline, compares the current motor features and current physiological response features with the corresponding activity baseline to obtain the degree of motor abnormality and physiological response abnormality for each motor state segment, and calculates the pre-fall risk value for each motor state segment based on the degree of motor abnormality and physiological response abnormality; the activity baseline is established based on the target elderly person's historical valid data and according to different motor states and different motor features and physiological response features, and the activity baseline is used to characterize the target elderly person's historical normal performance;

[0180] The early warning module is used to determine the abnormal state before a fall and issue an early warning based on the pre-fall risk value of each motion state segment and / or the pre-fall risk values ​​of multiple consecutive motion state segments.

[0181] The system of the present invention is used to implement the above method, and will not be described in detail here.

[0182] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention 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 the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for early warning of fall risk in the elderly based on changes in motor state and physiological response analysis, characterized in that, Includes the following steps: S1: Real-time acquisition of the target elderly person's movement and physiological data; S2: After segmenting the motion data into time windows, a pre-built motion state recognition model is used to classify and identify the motion state corresponding to each time window; adjacent time windows with the same motion state category in all time windows are merged to obtain multiple motion state segments divided by motion state, wherein each motion state segment contains data corresponding to one or more time windows. S3: For each motion state segment, extract the current motion features and the current physiological response features respectively; A pre-constructed activity baseline is obtained, and the current motor characteristics and current physiological response characteristics are compared with the corresponding activity baseline to obtain the degree of motor abnormality and the degree of physiological response abnormality for each motor state segment. The pre-fall risk value for each motor state segment is calculated based on the degree of motor abnormality and the degree of physiological response abnormality. The activity baseline is established based on the target elderly person's historical valid data and according to different motor states and different motor characteristics and physiological response characteristics. The activity baseline is used to characterize the target elderly person's historical normal performance. S4: Determine the pre-fall risk value based on the pre-fall risk value of each motion state segment, and / or the pre-fall risk values ​​of multiple consecutive motion state segments, and issue an early warning.

2. The method according to claim 1, characterized in that, In step S1, the real-time acquired motion data of the target elderly person includes derived motion data; the derived motion data is calculated based on the real-time collected triaxial acceleration data and triaxial angular velocity data. The derived motion data includes resultant acceleration, resultant angular velocity, posture changes, number of steps, stride frequency, motion intensity, and pause time.

3. The method according to claim 1, characterized in that, In step S1, the physiological data of the target elderly person acquired in real time includes heart rate data and pulse interval data.

4. The method according to claim 1, characterized in that, In step S2, a temporal convolutional network is used to construct the motion state recognition model; The temporal convolutional network includes an input layer, multiple residual convolutional modules, a global average pooling layer, a fully connected layer, and a softmax output layer connected in sequence. The input layer is used to input the motion data segmented based on a time window. The multiple residual convolutional modules are used to capture motion dependencies. Each residual convolutional module includes a one-dimensional dilated convolutional layer, a batch normalization layer, a ReLU activation layer, and a residual connection layer. The residual connection layer is used to connect the input of the current residual convolutional module and the output of the current ReLU activation layer.

5. The method according to claim 1, characterized in that, In step S3, the method for constructing the active baseline includes: Historical monitoring data of the target elderly person is obtained, including historical exercise data and historical physiological data; the wearing status and data quality status are determined based on the historical monitoring data and its collection time, and basic validity screening is performed to obtain the historical valid data; After segmenting the historical valid data into time windows, the motion state recognition model is used to classify and identify motion states. Adjacent time windows with the same motion state category in all time windows are merged to obtain multiple historical segments divided by motion state. Abnormal action segments in the historical segments are removed to obtain valid historical segments. Extract motion features and physiological response features from the effective historical segments; For the target elderly in the first The first motion state The first valid historical segment, the first of the valid historical segments The eigenvalues ​​are denoted as ; Segment weights are calculated based on the data quality and time distance of each valid historical segment. Its expression is: ； in, ; Indicates the first The first motion state Data quality score for each valid historical segment. Indicates the first The first motion state The total duration of each valid historical segment This indicates the invalid collection duration within the valid historical segment. This indicates the time interval between the start or end time of the valid historical segment and the baseline calculation time. Indicates the time decay coefficient. Indicates the target elderly in the first The number of valid historical segments used to establish a baseline under motion-like conditions; The first The first effective historical segment under the motion state The features are weighted and statistically analyzed according to the segment weights to obtain the target elderly person in the [number]th segment. The first motion state The active baseline mean of each feature Its expression is: ； The baseline mean activity level is used to reflect the target older adult's activity level during the period. The first motion state Historical benchmark levels for each feature; The corresponding active baseline fluctuation range is calculated based on the segment weight and the mean of the active baseline. Its expression is: ； in, To prevent constants with a fluctuation range of zero; This yields the activity baseline. .

6. The method according to claim 5, characterized in that, In step S3, the method for obtaining the degree of motion anomaly of each motion state segment includes the following steps: For each motion state segment, extract the current motion features; Let the motion state category to which the current motion state segment belongs be . , its first Each motion feature value is denoted as ; The first segment within the current motion state Each motion feature is compared with its corresponding activity baseline to calculate the degree of deviation of the current motion feature. Its expression is: ； Set motion feature deviation threshold ,when When an abnormality occurs, it is considered a motion anomaly, and the individual motion anomaly value for the current motion feature is calculated. The expression for this value is: ； in, Indicates the first segment of the current motion state. Individual motion anomalies of each motion feature; Based on the individual motion anomaly values ​​of each motion feature, calculate the degree of motion anomaly in the current motion state segment. Its expression is: ； in, This indicates the number of motion features in the current motion state segment.

7. The method according to claim 6, characterized in that, In step S3, the method for obtaining the degree of physiological response abnormality for each motion state segment includes the following steps: For each motion state segment, extract the current physiological response features, which include heart rate features and pulse interval fluctuation value features; Let the motion state category to which the current motion state segment belongs be . , No. The Hth feature in the exercise-like state is the heart rate feature, and the corresponding mean and range of activity baseline are respectively... and ; Compare the heart rate characteristics within the current motion segment with the corresponding [number]. By comparing heart rate characteristics under exercise-like conditions with an active baseline, the heart rate elevation load within the current exercise segment can be calculated. Its expression is: ； in, This indicates the number of valid heart rate values ​​within the current motion segment; Indicates the first segment of the current motion state. One effective heart rate value; To stabilize heart rate, based on the current exercise state before the segment begins. Calculation of the average value of effective heart rate data within seconds; Allowable deviation for increased heart rate; The characteristics of pulse interval fluctuation values ​​are obtained based on pulse interval data, namely, adjacent pulse interval fluctuation values. The expression is: ； in, This indicates the number of effective pulse intervals within the current motion segment; This represents the r-th effective pulse interval value within the current motion state segment; To prevent constants with a fluctuation range of zero; The pulse interval fluctuation values ​​within the current motion segment With the corresponding first Adjacent fluctuations in pulse interval during exercise-like conditions Compare the activity baseline to calculate abnormal values ​​of pulse interval fluctuations. Its expression is: ； in, Indicates the first The mean of the activity baseline corresponding to the characteristics of pulse interval fluctuations under exercise-like conditions. Indicates the first The range of activity baseline fluctuations corresponding to the characteristics of pulse interval fluctuations under motion-like conditions; Based on the aforementioned heart rate increase load Abnormal values ​​of pulse interval fluctuations Calculate the degree of abnormality in the physiological response of the current motion segment. Its expression is: ； in, It represents the combined enhancement coefficient when elevated heart rate load and abnormal pulse interval fluctuations occur simultaneously.

8. The method according to claim 7, characterized in that, In step S3, the method for calculating the pre-fall risk value for each motion state segment is as follows: For each motion state segment, the degree of motion anomaly is normalized to obtain the normalized degree of motion anomaly. Its expression is: ； in, The degree of motor abnormality; Based on the normalized degree of motion abnormality and the degree of abnormal physiological response The pre-fall risk value for the current motion segment is calculated using the following expression: ； in, This indicates the risk value before a fall in the current motion segment. This represents the enhancement coefficient when abnormal movement and abnormal physiological response occur simultaneously.

9. The method according to claim 8, characterized in that, In step S4, the method for determining the abnormal state before the fall and issuing an early warning includes: The abnormal state before the fall includes low-risk, medium-risk and high-risk states, with the corresponding risk levels increasing sequentially. when When the risk level is low, only the risk value is recorded. The low-risk threshold; when When the risk level is determined to be medium, an alert is sent. The high-risk threshold; when When the situation is deemed high-risk, an alarm command is sent. If the target elderly person is identified as having an abnormal state before a fall in M ​​consecutive motion state segments, the risk level of the latest motion state segment will be increased by one level. If the latest motion state segment was already in a high-risk state before the increase, the high-risk state will be maintained and an alarm command will be sent.

10. A fall risk warning system for the elderly based on changes in motion state and physiological response analysis, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire the target elderly person's exercise and physiological data in real time; The motion state recognition module is used to segment the motion data into time windows and then use a pre-built motion state recognition model to classify and recognize the motion state, so as to identify the motion state corresponding to each time window; and to merge adjacent time windows with the same motion state category in all time windows to obtain multiple motion state segments divided by motion state, wherein each motion state segment contains data corresponding to one or more time windows. The fall risk estimation module is used to extract current movement features and current physiological response features for each movement state segment, obtain a pre-constructed activity baseline, compare the current movement features and current physiological response features with the corresponding activity baseline to obtain the degree of movement abnormality and the degree of physiological response abnormality for each movement state segment, and calculate the pre-fall risk value for each movement state segment based on the degree of movement abnormality and the degree of physiological response abnormality. The activity baseline is established based on the target elderly person's historical valid data and is established separately according to different movement states and different movement features and physiological response features. The activity baseline is used to characterize the target elderly person's historical normal performance. The early warning module is used to determine the abnormal state before falling and issue an early warning based on the pre-fall risk value of each motion state segment and / or the pre-fall risk values ​​of multiple consecutive motion state segments.

Images