Stroke fall early warning method and related device based on flexible sensor

Abstract

This invention discloses a stroke fall early warning method and related device based on flexible sensors. The method includes: calculating the combined acceleration, combined angular velocity, and combined pressure time series collected by a lumbar flexible sensor and each ankle flexible sensor based on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography (EMG) data; calculating the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence, as well as the balance coefficient between the left and right ankle acceleration sequences, using a multi-scale DTW algorithm; fusing the stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and EMG data collected by the flexible sensors to form a multimodal feature vector; and inputting the multimodal feature vector into an ANFIS fall model to predict the patient's fall risk probability. This invention improves the accuracy of patient fall risk prediction by using fuzzy logic and neural networks for multimodal data fusion and gait analysis.

Description

Technical Field

[0001] This invention relates to the field of medical assistive technology, specifically to a stroke fall early warning method and device, and a computing device based on a flexible sensor. Background Technology

[0002] Stroke patients often face problems such as balance dysfunction, decreased muscle strength, and sensory abnormalities, making them more prone to falls.

[0003] Traditional fall risk assessment relies mainly on doctors' subjective judgment or simple clinical tests (such as the stand-up-walk test), which has drawbacks such as high subjectivity, low efficiency, and inability to conduct real-time monitoring, making it difficult to meet clinical needs.

[0004] With the development of artificial intelligence technology, some studies have begun to attempt to use machine learning or deep learning models to predict fall risk by analyzing sensor data. However, most of these studies rely on data from a single sensor (such as accelerometer) or a single feature (such as walking speed) for fall detection, ignoring the complexity of falls and failing to comprehensively reflect a patient's fall risk.

[0005] To address the aforementioned issues, this invention proposes a stroke fall early warning method based on flexible sensors. By using fuzzy logic and neural networks to fuse multimodal data and perform gait analysis, the method aims to improve the accuracy of predicting patient fall risk. Summary of the Invention

[0006] In view of the above problems, the present invention provides a stroke fall early warning method and device, and a computing device based on a flexible sensor.

[0007] According to one aspect of the present invention, a stroke fall early warning method based on a flexible sensor is provided, comprising:

[0008] The system collects real-time data on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography in three-dimensional space using multiple flexible sensors. These flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor, and a flexible electromyography sensor. The sensors are wearable and attached to multiple preset locations on the patient's body, including the waist, ankle, elbow, and knee joints.

[0009] Based on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography data in three-dimensional space, the combined acceleration, combined angular velocity, and combined pressure time series collected by the lumbar flexible sensor and each ankle flexible sensor are calculated; the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence, as well as the balance coefficient between the left and right ankle acceleration sequences, are calculated using the multi-scale DTW algorithm.

[0010] By integrating stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and electromyography data collected by flexible sensors, a multimodal feature vector is formed.

[0011] The multimodal feature vectors are input into the ANFIS fall model to predict the patient's fall risk probability; wherein, the ANFIS fall model includes an input layer, a stacked hourglass prediction layer, a network fuzzing layer, a rule layer, a normalization layer, a consequence layer, and an output layer.

[0012] In an alternative embodiment, the calculation of the resultant acceleration, resultant angular velocity, and resultant pressure time series acquired by the lumbar flexible sensor and each ankle flexible sensor further includes:

[0013] The resultant acceleration is calculated once based on the three components of the acceleration at each time point, and the resultant acceleration time series is obtained.

[0014] The resultant angular velocity is calculated once based on the three components of the angular velocity at each time point, resulting in a resultant angular velocity time series;

[0015] For a multi-region discrete pressure sensor, the pressure values ​​at each time point are summed to obtain the combined pressure time series; for a multi-region continuous pressure sensor, the average pressure value of each pixel region at each time point is calculated to obtain the combined pressure time series.

[0016] The flexible pressure sensor includes a single-point pressure sensor, a multi-region discrete pressure sensor, and a multi-region continuous pressure sensor. The multi-region discrete pressure sensor contains multiple independent pressure-sensitive regions, each of which provides a pressure value. The multi-region continuous pressure sensor provides a pressure value for each pixel region in the form of a matrix.

[0017] In an alternative approach, the calculation of the stability coefficients between the waist acceleration sequence and the normal gait template sequence, as well as the balance coefficients between the left and right ankle acceleration sequences, based on the multi-scale DTW algorithm, further includes:

[0018] Multiscale smoothing was performed on the combined acceleration time series of the waist and ankle.

[0019] The stability coefficient for each gait cycle is calculated based on the MS-DTW distance between the combined lumbar acceleration time series and the normal gait template sequence.

[0020] The balance coefficient for each gait cycle is calculated based on the MS-DTW distance of the combined acceleration time series of the left and right ankles.

[0021] In one alternative approach, the fusion of stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and electromyography data collected by the flexible sensor to form a multimodal feature vector further includes:

[0022] The equilibrium state is calculated based on the stability coefficient and the equilibrium coefficient; the maximum pressure, average pressure and pressure center location are obtained based on the pressure distribution data; the heart rate variability is calculated based on the heart rate data; and the activity intensity is calculated based on the electromyography data.

[0023] Based on the set activity intensity weight, equilibrium state weight, and heart rate variability weight, a multimodal feature vector is obtained.

[0024] In one alternative approach, the fuzzification layer includes multiple input nodes, a fuzzy rule layer, and an output node; wherein each input node is connected to at least one membership function; the membership function includes a Gaussian function, a bell function, and a trapezoidal function; and the fuzzy rule layer calculates the trigger strength of each rule according to the fuzzy rules.

[0025] The stacked hourglass prediction layer is used for joint pose prediction and supplementing missing joint data. It includes an initial convolutional layer, multiple hourglass modules, and an end convolutional layer. Each hourglass module includes a downsampled path encoding layer, a low-resolution feature map network layer, and an upsampled path decoding layer.

[0026] The consequence layer calculates the fall risk probability based on the trigger strength and the consequent parameters of the fuzzy rule.

[0027] In one alternative approach, the pressure distribution data includes pressure center offset, plantar pressure distribution symmetry, pressure peak position within the gait cycle, and dynamic pressure distribution range.

[0028] The heart rate data features include heart rate variability indicators, heart rate acceleration / deceleration rate, and frequency of abnormal heart rate fluctuations;

[0029] The electromyographic data includes muscle activation intensity, muscle synergistic activation level, and muscle fatigue indicators.

[0030] In an alternative approach, calculating the fall risk probability based on the trigger strength and the consequent parameters of the fuzzy rule further includes:

[0031] Define the state set S = {stable gait, slightly unstable gait, moderately unstable gait, high risk of falling, and fall occurred};

[0032] Define the observation symbol V = [p1,p2,p3,p4,h1,h2,h3,e1,e2,e3,st,b]; where p1 is the pressure center offset; p2 is the symmetry of plantar pressure distribution; p3 is the position of the pressure peak within the gait cycle; p4 ​​is the dynamic pressure distribution range; h1 is the heart rate variability index; h2 is the heart rate acceleration / deceleration rate; h3 is the frequency of abnormal heart rate fluctuations; e1 is the muscle activation intensity; e2 is the degree of muscle synergistic activation; e3 is the muscle fatigue index; st is the stability coefficient; and b is the balance coefficient.

[0033] Define the transition probability matrix A, the initial state probability distribution π, and the observation probability matrix B for each state set S;

[0034] The probability of each state from the first observation to the current observation is calculated using the forward algorithm, and the probability of each state from the current observation to the last observation is calculated using the backward algorithm.

[0035] The trigger strength is adjusted according to the probability of each state, and the probability of each state is incorporated into the consequent parameter of the fuzzy rule.

[0036] Calculate the probability of falling based on the adjusted trigger strength and rule consequent parameters.

[0037] In one alternative approach, the formula for calculating the probability of falling is:

[0038]

[0039] Among them, P i α is the consequent parameter of the i-th rule; i_adjusted Let α be the trigger strength of the i-th rule; i_adjusted =α i ·∑ s∈S P(q t =s|O)·w s ; xi is the i-th input variable, L xi Let be the membership value of xi; P(q) t =s|O) is the probability that, given the observation sequence O, the state is s at time t; w s Let S be the weight of state s; S is the set of states.

[0040] According to another aspect of the present invention, a stroke fall warning device based on a flexible sensor is provided, comprising:

[0041] The data acquisition module is used to collect the patient's acceleration, angular velocity, pressure, heart rate and electromyography data in three-dimensional space in real time based on multiple flexible sensors. The flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor and a flexible electromyography sensor, which can be worn and attached to multiple preset positions on the patient's body, including the waist, ankle, elbow joint and knee joint.

[0042] The dynamic time warping feature extraction module is used to calculate the resultant acceleration, resultant angular velocity, and resultant pressure time series collected by the lumbar flexible sensor and each ankle flexible sensor based on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography data in three-dimensional space; and to calculate the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence and the balance coefficient between the left and right ankle acceleration sequences based on the multi-scale DTW algorithm.

[0043] The multimodal data fusion module is used to fuse stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and electromyography data collected by flexible sensors to form a multimodal feature vector;

[0044] The fall prediction module is used to input the multimodal feature vector into the ANFIS fall model to predict the patient's fall risk probability; wherein, the ANFIS fall model includes an input layer, a fuzzification layer, a rule layer, a normalization layer, a consequence layer, and an output layer.

[0045] According to another aspect of the present invention, a computing device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus;

[0046] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described stroke fall early warning method based on flexible sensors.

[0047] According to the solution provided by the present invention, multiple flexible sensors are used to collect real-time acceleration, angular velocity, pressure, heart rate, and electromyography (EMG) data of a patient in three-dimensional space. The flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor, and a flexible EMG sensor, and are wearable and attached to multiple preset locations on the patient's body, including the waist, ankle, elbow, and knee joints. Based on the patient's acceleration, angular velocity, pressure, heart rate, and EMG data in three-dimensional space, the combined acceleration collected by the lumbar flexible sensor and each ankle flexible sensor is calculated. The invention utilizes time series data of angular velocity and pressure; calculates the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence, as well as the balance coefficient between the left and right ankle acceleration sequences, based on the multi-scale DTW algorithm; fuses the stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and electromyography data collected by flexible sensors to form a multimodal feature vector; inputs the multimodal feature vector into the ANFIS fall model to predict the patient's fall risk probability; wherein, the ANFIS fall model includes an input layer, a stacked hourglass prediction layer, a network fuzzification layer, a rule layer, a normalization layer, a consequence layer, and an output layer. This invention improves the accuracy of patient fall risk prediction through multimodal data fusion and gait analysis using fuzzy logic and neural networks.

[0048] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0049] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0050] Figure 1 A flowchart illustrating a stroke fall early warning method based on a flexible sensor according to an embodiment of the present invention is shown.

[0051] Figure 2 A schematic diagram illustrating the calculation process of the stability coefficient and equilibrium coefficient according to an embodiment of the present invention is shown;

[0052] Figure 3 A schematic diagram of the ANFIS fall model according to an embodiment of the present invention is shown;

[0053] Figure 4 A schematic diagram of the fall risk probability calculation process according to an embodiment of the present invention is shown;

[0054] Figure 5 A 2D-3D pose diagram of an embodiment of the present invention is shown;

[0055] Figure 6 A schematic diagram illustrating posture missing supplementation according to an embodiment of the present invention is shown;

[0056] Figure 7 This invention presents a schematic diagram of the framework of the device for stroke fall early warning method based on flexible sensors, according to an embodiment of the present invention.

[0057] Figure 8 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. Detailed Implementation

[0058] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0059] Figure 1 A flowchart illustrating a stroke fall early warning method based on a flexible sensor, according to an embodiment of the present invention, is shown. Specifically, as... Figure 1 As shown, it includes the following steps:

[0060] Step S101: Real-time acquisition of the patient's acceleration, angular velocity, pressure, heart rate, and electromyography data in three-dimensional space using multiple flexible sensors. The flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor, and a flexible electromyography sensor, which can be worn and attached to multiple preset positions on the patient's body, including the waist, ankle, elbow, and knee joints.

[0061] In this embodiment, a flexible triaxial accelerometer measures the patient's acceleration in three axes (forward / backward, left / right, and up / down) to reflect the patient's body movement and posture changes. A flexible triaxial gyroscope measures the patient's angular velocity in three axes, i.e., the speed of body rotation, to reflect the patient's balance and posture changes. A flexible pressure sensor measures the pressure between the patient's body and the contact surface (e.g., a plantar pressure sensor measures the pressure distribution on the sole of the foot when the patient is walking or standing). A flexible heart rate sensor measures the patient's heart rate. A flexible electromyography (EMG) sensor measures the electrical activity of the patient's muscles; the EMG signal reflects the muscle activation intensity and fatigue level.

[0062] To ensure convenient and comfortable long-term use of the sensor, this invention employs a flexible sensor integrated into a wearable device, transmitting data via a wireless network, for example:

[0063] Strap / armband type: Flexible accelerometers, gyroscopes, electromyography sensors, etc. are integrated into adjustable straps or armbands and fixed at the waist, ankle, elbow and knee joints.

[0064] Insole type: Flexible pressure sensors are embedded in the insole for easy monitoring of foot pressure.

[0065] Patch type: The heart rate sensor is attached to the chest or wrist in a patch-like manner.

[0066] Step S102: Based on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography data in three-dimensional space, calculate the combined acceleration, combined angular velocity, and combined pressure time series collected by the lumbar flexible sensor and each ankle flexible sensor; calculate the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence, as well as the balance coefficient between the left and right ankle acceleration sequences, according to the multi-scale DTW algorithm.

[0067] In this embodiment, as Figure 2 As shown, wearable sensors are used to collect human motion data, gait features are extracted through data processing, and the stability and balance of the gait are evaluated using a multi-scale dynamic time warping (DTW) algorithm. The multi-scale DTW algorithm provides quantitative indicators of gait stability and balance, facilitating objective evaluation and tracking. Personalized gait analysis is achieved by comparing the data with a normal gait template. The resultant acceleration is calculated by taking the square root of the sum of the squares of the acceleration values ​​along each axis, and the resultant pressure can be obtained by averaging the data from each pressure sensor or using other weighting methods.

[0068] Specifically, gait data of normal people are collected in advance, and their combined acceleration time series is extracted as a template for normal gait.

[0069] Multi-scale decomposition was performed on the waist acceleration sequence and the normal gait template sequence. For example, wavelet decomposition was used to decompose the data into components of different frequencies (approximate components and detail components).

[0070] At each scale, the distance between the waist acceleration sequence and the normal gait template sequence is calculated using the DTW algorithm. The DTW algorithm can handle sequences of different lengths and rhythms and allows for local time warps between sequences.

[0071] The stability coefficient is defined as the reciprocal of the DTW distance or a normalized value. The higher the stability coefficient, the closer the gait is to the normal gait template.

[0072] Calculate the multi-scale DTW distance between the resultant acceleration sequences of the left and right ankles. The balance coefficient is defined as the reciprocal of the DTW distance or calculated based on the offset of the DTW distance. The higher the balance coefficient, the more symmetrical the movement of the left and right ankles.

[0073] In an alternative embodiment, the calculation of the resultant acceleration, resultant angular velocity, and resultant pressure time series acquired by the lumbar flexible sensor and each ankle flexible sensor further includes:

[0074] The resultant acceleration is calculated once based on the three components of the acceleration at each time point, and the resultant acceleration time series is obtained.

[0075] The resultant angular velocity is calculated once based on the three components of the angular velocity at each time point, resulting in a resultant angular velocity time series;

[0076] For a multi-region discrete pressure sensor, the pressure values ​​at each time point are summed to obtain the combined pressure time series; for a multi-region continuous pressure sensor, the average pressure value of each pixel region at each time point is calculated to obtain the combined pressure time series.

[0077] The flexible pressure sensor includes a single-point pressure sensor, a multi-region discrete pressure sensor, and a multi-region continuous pressure sensor. The multi-region discrete pressure sensor contains multiple independent pressure-sensitive regions, each of which provides a pressure value. The multi-region continuous pressure sensor provides a pressure value for each pixel region in the form of a matrix.

[0078] In this embodiment, different resultant pressure calculation methods are adopted for different types of pressure sensors (single-point, multi-region discrete, and multi-region continuous), and the resultant acceleration and resultant angular velocity are calculated using standard vector magnitude, ultimately forming a time series.

[0079] Specifically, for each time point t, the three components of acceleration are Ax(t), Ay(t), and Az(t), and the resultant acceleration is A_mag(t): Arrange the resultant acceleration values ​​at each time point in chronological order to obtain the resultant acceleration time series, and similarly obtain the resultant angular velocity time series.

[0080] For a single-point pressure sensor, at each time point t, the pressure sensor provides a pressure value P(t), which is directly used as the resultant pressure at that time point.

[0081] For a multi-region discrete pressure sensor, the sensor has n independent pressure-sensitive regions, each providing pressure values ​​P1(t), P2(t), ..., Pn(t). The resultant pressure P_mag(t) is then: P1(t) + P2(t) + ... + Pn(t). Arranging the resultant pressure values ​​at each time point in chronological order yields the resultant pressure time series.

[0082] For a multi-region continuous pressure sensor, the sensor has m×n pixel regions, and the pressure value of each region is Pij(t), where i represents the row and j represents the column. The resultant pressure P_mag(t) is the average pressure value of all pixel regions. Arranging the resultant pressure values ​​at each time point in chronological order yields the resultant pressure time series.

[0083] Arrange the resultant acceleration values ​​at each time point in chronological order to obtain the resultant acceleration time series.

[0084] In an alternative approach, the calculation of the stability coefficients between the waist acceleration sequence and the normal gait template sequence, as well as the balance coefficients between the left and right ankle acceleration sequences, based on the multi-scale DTW algorithm, further includes:

[0085] Multiscale smoothing was performed on the combined acceleration time series of the waist and ankle.

[0086] The stability coefficient for each gait cycle is calculated based on the MS-DTW distance between the combined lumbar acceleration time series and the normal gait template sequence.

[0087] The balance coefficient for each gait cycle is calculated based on the MS-DTW distance of the combined acceleration time series of the left and right ankles.

[0088] In this embodiment, traditional DTW is highly sensitive to local distortions in time series, while MS-DTW, through smoothing at multiple scales, is more tolerant of noise, individual differences, and variations in stride speed in gait data. By comparing the similarity of time series at multiple scales, more comprehensive gait features are captured, thereby more accurately calculating the differences between the gait and the normal template, as well as the coordination of left and right ankle movements. Furthermore, multi-scale smoothing extracts gait features at different scales; for example, the coarse scale reflects the overall smoothness of the gait, while the fine scale reflects subtle changes in stride details. If the MS-DTW distance for a certain gait cycle is 0.2, while the standard for a normal person is 0.05, it indicates that the movements of the left and right ankles are uncoordinated and the balance is poor during that gait cycle.

[0089] Step S103: The stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and electromyography data collected by the flexible sensor are fused to form a multimodal feature vector.

[0090] In one alternative approach, the fusion of stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and electromyography data collected by the flexible sensor to form a multimodal feature vector further includes:

[0091] The equilibrium state is calculated based on the stability coefficient and the equilibrium coefficient; the maximum pressure, average pressure and pressure center location are obtained based on the pressure distribution data; the heart rate variability is calculated based on the heart rate data; and the activity intensity is calculated based on the electromyography data.

[0092] Based on the set activity intensity weight, equilibrium state weight, and heart rate variability weight, a multimodal feature vector is obtained.

[0093] In this embodiment, multimodal fusion can capture subtle changes that are difficult to detect with a single modality. For example, by using pressure distribution and electromyography data, the relationship between muscle activity and plantar pressure can be analyzed more precisely.

[0094] For example, the stability coefficient is 0.4, the balance coefficient is 0.3, and the pressure distribution extracts the pressure peak (left foot: 1.2 kPa, right foot: 1.1 kPa), the dispersion of the pressure center trajectory (0.05 cm), and the symmetry of the pressure distribution (0.1). Heart rate: average heart rate 90 bpm, heart rate variability 40 ms. Electromyography: the average activation amplitude of the quadriceps femoris is 0.5 mV, and the activation time of the tibialis anterior muscle is 0.2 seconds. All extracted features are concatenated into a vector to form a multimodal feature vector: [0.4, 0.3, 1.2, 1.1, 0.05, 0.1, 90, 40, 0.5, 0.2].

[0095] In one alternative approach, the pressure distribution data includes pressure center offset, plantar pressure distribution symmetry, pressure peak position within the gait cycle, and dynamic pressure distribution range.

[0096] The heart rate data features include heart rate variability indicators, heart rate acceleration / deceleration rate, and frequency of abnormal heart rate fluctuations;

[0097] The electromyographic data includes muscle activation intensity, muscle synergistic activation level, and muscle fatigue indicators.

[0098] In this embodiment, the pressure center offset reflects the movement trajectory and stability of the pressure center during gait; a large offset indicates gait instability or poor center of gravity control. The symmetry of plantar pressure distribution reflects the degree of symmetry in weight-bearing and pressure distribution between the feet. The location of the pressure peak within the gait cycle reflects the area and time of pressure concentration. The dynamic pressure distribution range reflects the change in the range of force on the plantar surface during gait, revealing motor control and muscle coordination. The heart rate acceleration / deceleration rate reflects how quickly the heart rate changes over time. The frequency of abnormal heart rate fluctuations reflects the frequency of irregular heart rate fluctuations. The degree of muscle synergistic activation reflects the coordinated activity of different muscles during movement. Muscle fatigue indicators reflect the degree of muscle fatigue and can be assessed using specific parameters (median frequency) in electromyography (EMG) signals.

[0099] Step S104: Input the multimodal feature vector into the ANFIS fall model to predict the patient's fall risk probability; wherein, the ANFIS fall model includes an input layer, a stacked hourglass prediction layer, a network fuzzing layer, a rule layer, a normalization layer, a consequence layer, and an output layer.

[0100] In this embodiment, the ANFIS model combines the self-learning ability of neural networks with the uncertainty handling capabilities of fuzzy logic. The neural network part learns patterns in the data, while the fuzzy logic part handles the fuzziness and uncertainty in the data. Unlike black-box neural networks, the ANFIS model has a degree of interpretability; fuzzy rules reveal the relationship between features and fall risk.

[0101] The model is structured as follows: The input layer receives multimodal feature vectors, with the number of input neurons equal to the dimension of the feature vectors. The stacked hourglass prediction layer enhances the model's feature learning ability; the stacked hourglass structure includes encoder and decoder components. The network fuzzification layer transforms input features into fuzzy membership degrees, mapping each input feature to a different fuzzy set through a predefined membership function (such as a Gaussian function or a trigonometric function). The rule layer defines a set of fuzzy rules, each representing the relationship between a combination of input features and the risk of falling. The consequence layer calculates the contribution of each rule to the output based on its activation strength and consequence parameters. The output layer performs a weighted sum of the contributions from all rules to obtain the final probability of falling.

[0102] In one alternative approach, the fuzzification layer includes multiple input nodes, a fuzzy rule layer, and an output node; wherein each input node is connected to at least one membership function; the membership function includes a Gaussian function, a bell function, and a trapezoidal function; and the fuzzy rule layer calculates the trigger strength of each rule according to the fuzzy rules.

[0103] The stacked hourglass prediction layer is used for joint pose prediction and supplementing missing joint data. It includes an initial convolutional layer, multiple hourglass modules, and an end convolutional layer. Each hourglass module includes a downsampled path encoding layer, a low-resolution feature map network layer, and an upsampled path decoding layer.

[0104] The consequence layer calculates the fall risk probability based on the trigger strength and the consequent parameters of the fuzzy rule.

[0105] In this embodiment, as Figure 3 As shown, each input node can be connected to multiple membership functions. An input feature expresses multiple fuzzy concepts during the fuzzification process, thus more precisely characterizing the complex relationship between features and fall risk. For example, a Gaussian function smoothly represents the center value, while a trapezoidal function represents the interval value. Using stacked hourglass prediction layers more effectively learns human joint posture information and completes missing joint data, learning information representations with more information content than the original features. Through downsampling and upsampling operations, the hourglass module learns features at different scales, thereby better capturing global and local information about joint posture. The low-resolution feature map network layer performs convolution and activation operations on the low-resolution feature map. The final convolutional layer convolves the output of the last hourglass module to obtain the final joint posture feature representation, and performs joint posture prediction and completes missing joint data (such as...). Figure 5 , Figure 6 (As shown). The consequence layer calculates the fall risk probability based on the trigger strength of the rule and the consequent parameters of the rule.

[0106] In an alternative approach, calculating the fall risk probability based on the trigger strength and the consequent parameters of the fuzzy rule further includes:

[0107] Define the state set S = {stable gait, slightly unstable gait, moderately unstable gait, high risk of falling, and fall occurred};

[0108] Define the observation symbol V = [p1,p2,p3,p4,h1,h2,h3,e1,e2,e3,st,b]; where p1 is the pressure center offset; p2 is the symmetry of plantar pressure distribution; p3 is the position of the pressure peak within the gait cycle; p4 ​​is the dynamic pressure distribution range; h1 is the heart rate variability index; h2 is the heart rate acceleration / deceleration rate; h3 is the frequency of abnormal heart rate fluctuations; e1 is the muscle activation intensity; e2 is the degree of muscle synergistic activation; e3 is the muscle fatigue index; st is the stability coefficient; and b is the balance coefficient.

[0109] Define the transition probability matrix A, the initial state probability distribution π, and the observation probability matrix B for each state set S;

[0110] The probability of each state from the first observation to the current observation is calculated using the forward algorithm, and the probability of each state from the current observation to the last observation is calculated using the backward algorithm.

[0111] The trigger strength is adjusted according to the probability of each state, and the probability of each state is incorporated into the consequent parameter of the fuzzy rule.

[0112] Calculate the probability of falling based on the adjusted trigger strength and rule consequent parameters.

[0113] In this embodiment, as Figure 4 As shown, a series of fall risk-related states are defined, such as from "stable gait" to "fall occurrence," which can capture the dynamic changes in a patient's fall risk, rather than just a static assessment. The set of observation symbols V includes indicators from multiple dimensions, such as center of pressure, plantar pressure distribution, heart rate variability, and muscle activation, assessing the patient's fall risk from multiple perspectives. By inferring the patient's latent state (e.g., "moderately unstable gait") from the observed data, the evolution of fall risk can be captured, rather than simply analyzing independent observations. The state probability is calculated using a forward-backward algorithm and incorporated into the fuzzy rule consequent parameters of ANFIS. The consequent parameters of the rules are no longer static but change according to the patient's real-time state. The trigger strength is dynamically adjusted according to the state probability, enabling more accurate predictions based on the patient's current state.

[0114] For example, the observation symbol V uses only the pressure center offset p1 as the observation symbol, and the state set S: {stable gait, high risk of falling}.

[0115] Observation sequence: Patient p1 changes over time. [0.2, 0.3, 0.4, 0.7, 0.8, 0.9, 0.5, 0.4].

[0116] Fuzzy rule: IF p1 IS high THEN fall risk IS high; IF p1 IS low THEN fall risk IS low.

[0117] Initialize A, π, and B. Set the transition probability from a "stable gait" to a "high-risk fall state" to 0.2, and vice versa to 0.1. The initial probability of π's "stable gait" is 0.9. The B matrix can be initialized based on the training data; for example, under the "high-risk fall state," the mean value of p1 is relatively high.

[0118] The ANFIS parameters are trained using training data, and the membership function and consequent parameters of ANFIS are continuously updated through backpropagation and gradient descent. For the observation sequence [0.2, 0.3, 0.4, 0.7, 0.8, 0.9, 0.5, 0.4], the state probability at each time step is calculated using a forward-backward algorithm. For example, at p1 = 0.9, the probability of a "high-risk fall state" is relatively high. The consequent parameters of the rule are dynamically adjusted based on the state probabilities. For example, if the probability of a "high-risk fall state" is high at the current time step, the consequent parameter of the rule "IF p1 IS high THEN fall risk IS high" will be amplified. Based on the adjusted rule and trigger strength, the final fall risk probability is calculated.

[0119] In one alternative approach, the formula for calculating the probability of falling is:

[0120]

[0121] Among them, P i α is the consequent parameter of the i-th rule; i_adjusted Let α be the trigger strength of the i-th rule; i_adjusted =α i ·∑ s∈S P(q t =s|O)·w s ; xi is the i-th input variable, L xi Let be the membership value of xi; P(q) t =s|O) is the probability that, given the observation sequence O, the state is s at time t; w s Let S be the weight of state s; S is the set of states.

[0122] In this embodiment, the risk values ​​of multiple rules are weighted and averaged, with the weights determined by the rule trigger strength and the probability of the current state. This allows for personalized risk assessments for different groups of people. Fuzzy logic handles the uncertainty of input variables and rules, which is more in line with reality. The rule form has good interpretability, making it easy to understand the reasons for the risk assessment.

[0123] According to the solution provided by the present invention, multiple flexible sensors are used to collect real-time acceleration, angular velocity, pressure, heart rate, and electromyography (EMG) data of a patient in three-dimensional space. The flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor, and a flexible EMG sensor, and are wearable and attached to multiple preset locations on the patient's body, including the waist, ankle, elbow, and knee joints. Based on the patient's acceleration, angular velocity, pressure, heart rate, and EMG data in three-dimensional space, the combined acceleration collected by the lumbar flexible sensor and each ankle flexible sensor is calculated. The invention utilizes time series data of angular velocity and pressure; calculates the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence, as well as the balance coefficient between the left and right ankle acceleration sequences, based on the multi-scale DTW algorithm; fuses the stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and electromyography data collected by flexible sensors to form a multimodal feature vector; inputs the multimodal feature vector into the ANFIS fall model to predict the patient's fall risk probability; wherein, the ANFIS fall model includes an input layer, a stacked hourglass prediction layer, a network fuzzification layer, a rule layer, a normalization layer, a consequence layer, and an output layer. This invention improves the accuracy of patient fall risk prediction through multimodal data fusion and gait analysis using fuzzy logic and neural networks.

[0124] Figure 7 A schematic diagram of the framework of a stroke fall warning device based on a flexible sensor according to an embodiment of the present invention is shown. The stroke fall warning device based on a flexible sensor includes:

[0125] The data acquisition module 710 is used to collect the patient's acceleration, angular velocity, pressure, heart rate and electromyography data in three-dimensional space in real time based on multiple flexible sensors. The flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor and a flexible electromyography sensor, which can be worn and attached to multiple preset positions on the patient's body, including the waist, ankle, elbow joint and knee joint.

[0126] The dynamic time warping feature extraction module 720 is used to calculate the resultant acceleration, resultant angular velocity, and resultant pressure time series collected by the lumbar flexible sensor and each ankle flexible sensor based on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography data in three-dimensional space; and to calculate the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence and the balance coefficient between the left and right ankle acceleration sequences based on the multi-scale DTW algorithm.

[0127] The multimodal data fusion module 730 is used to fuse stability coefficient, balance coefficient and pressure distribution data, heart rate data and electromyography data collected by flexible sensors to form a multimodal feature vector;

[0128] The fall prediction module 740 is used to input the multimodal feature vector into the ANFIS fall model to predict the patient's fall risk probability; wherein, the ANFIS fall model includes an input layer, a fuzzification layer, a rule layer, a normalization layer, a consequence layer and an output layer.

[0129] Figure 8 The diagram shows a structural schematic of an embodiment of the computing device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computing device.

[0130] like Figure 5 As shown, the computing device may include: a processor 802, a communications interface 804, a memory 806, and a communications bus 808.

[0131] The processor 802, communication interface 804, and memory 806 communicate with each other via communication bus 808. Communication interface 804 is used to communicate with other network elements such as clients or other servers. The processor 802 executes program 810, specifically performing the relevant steps in the above-described embodiment of the stroke and fall warning method based on flexible sensors.

[0132] Specifically, program 810 may include program code that includes computer operation instructions.

[0133] Processor 802 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computing device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0134] Memory 806 is used to store program 810. Memory 806 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0135] According to the solution provided by the present invention, multiple flexible sensors are used to collect real-time acceleration, angular velocity, pressure, heart rate, and electromyography (EMG) data of a patient in three-dimensional space. The flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor, and a flexible EMG sensor, and are wearable and attached to multiple preset locations on the patient's body, including the waist, ankle, elbow, and knee joints. Based on the patient's acceleration, angular velocity, pressure, heart rate, and EMG data in three-dimensional space, the combined acceleration collected by the lumbar flexible sensor and each ankle flexible sensor is calculated. The invention utilizes time series data of angular velocity and pressure; calculates the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence, as well as the balance coefficient between the left and right ankle acceleration sequences, based on the multi-scale DTW algorithm; fuses the stability coefficient, balance coefficient, and pressure distribution data, heart rate data, and electromyography data collected by flexible sensors to form a multimodal feature vector; inputs the multimodal feature vector into the ANFIS fall model to predict the patient's fall risk probability; wherein, the ANFIS fall model includes an input layer, a stacked hourglass prediction layer, a network fuzzification layer, a rule layer, a normalization layer, a consequence layer, and an output layer. This invention improves the accuracy of patient fall risk prediction through multimodal data fusion and gait analysis using fuzzy logic and neural networks.

[0136] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination of all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed can be employed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose. Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

Claims

1. A stroke fall early warning method based on flexible sensors, characterized in that, include: The system collects real-time data on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography in three-dimensional space using multiple flexible sensors. These flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor, and a flexible electromyography sensor. The sensors are wearable and attached to multiple preset locations on the patient's body, including the waist, ankle, elbow, and knee joints. Based on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography data in three-dimensional space, the combined acceleration, combined angular velocity, and combined pressure time series collected by the lumbar flexible sensor and each ankle flexible sensor are calculated; the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence, as well as the balance coefficient between the left and right ankle acceleration sequences, are calculated using the multi-scale DTW algorithm. By integrating stability coefficients, balance coefficients, and pressure distribution data, heart rate data, and electromyographic data collected by flexible sensors, a multimodal feature vector is formed. The pressure distribution data includes pressure center offset, plantar pressure distribution symmetry, pressure peak position within the gait cycle, and dynamic pressure distribution range. The heart rate data features include heart rate variability indicators, heart rate acceleration / deceleration rate, and abnormal heart rate fluctuation frequency. The electromyographic data includes muscle activation intensity, muscle synergistic activation degree, and muscle fatigue indicators. The multimodal feature vectors are input into the ANFIS fall model to predict the patient's fall risk probability. The ANFIS fall model includes an input layer, a stacked hourglass prediction layer, a network fuzzification layer, a rule layer, a normalization layer, a consequence layer, and an output layer. The network fuzzification layer includes multiple input nodes, a fuzzy rule layer, and an output node. Each input node is connected to at least one membership function, including Gaussian, bell, and trapezoidal functions. The fuzzy rule layer calculates the trigger strength of each rule based on fuzzy rules. The stacked hourglass prediction layer, used for joint posture prediction and supplementing missing joint data, includes an initial convolutional layer, multiple hourglass modules, and an end convolutional layer. Each hourglass module includes a downsampled path encoding layer, a low-resolution feature map network layer, and an upsampled path decoding layer. The consequence layer calculates the fall risk probability based on the trigger strength and the consequent parameters of the fuzzy rules.

2. The stroke fall early warning method based on a flexible sensor according to claim 1, characterized in that, The calculation of the resultant acceleration, resultant angular velocity, and resultant pressure time series collected by the lumbar flexible sensor and each ankle flexible sensor further includes: The resultant acceleration is calculated once based on the three components of the acceleration at each time point, and the resultant acceleration time series is obtained. The resultant angular velocity is calculated once based on the three components of the angular velocity at each time point, resulting in a resultant angular velocity time series; For a multi-region discrete pressure sensor, the pressure values ​​at each time point are summed to obtain the combined pressure time series; for a multi-region continuous pressure sensor, the average pressure value of each pixel region at each time point is calculated to obtain the combined pressure time series. The flexible pressure sensor includes a single-point pressure sensor, a multi-region discrete pressure sensor, and a multi-region continuous pressure sensor. The multi-region discrete pressure sensor contains multiple independent pressure-sensitive regions, each of which provides a pressure value. The multi-region continuous pressure sensor provides a pressure value for each pixel region in the form of a matrix.

3. The stroke fall early warning method based on a flexible sensor according to claim 1, characterized in that, The calculation of the stability coefficients between the waist acceleration sequence and the normal gait template sequence, as well as the balance coefficients between the left and right ankle acceleration sequences, based on the multi-scale DTW algorithm, further includes: Multiscale smoothing was performed on the combined acceleration time series of the waist and ankle. The stability coefficient for each gait cycle is calculated based on the MS-DTW distance between the combined lumbar acceleration time series and the normal gait template sequence. The balance coefficient for each gait cycle is calculated based on the MS-DTW distance of the combined acceleration time series of the left and right ankles.

4. The stroke fall early warning method based on a flexible sensor according to claim 1, characterized in that, The process of fusing stability coefficients, balance coefficients, and pressure distribution data, heart rate data, and electromyography data collected by flexible sensors to form a multimodal feature vector further includes: The equilibrium state is calculated based on the stability coefficient and the equilibrium coefficient; the maximum pressure, average pressure and pressure center location are obtained based on the pressure distribution data; the heart rate variability is calculated based on the heart rate data; and the activity intensity is calculated based on the electromyography data. Based on the set activity intensity weight, equilibrium state weight, and heart rate variability weight, a multimodal feature vector is obtained.

5. The stroke fall early warning method based on a flexible sensor according to claim 1, characterized in that, The calculation of the fall risk probability based on the trigger strength and the consequent parameters of the fuzzy rule further includes: Define the state set S = {stable gait, slightly unstable gait, moderately unstable gait, high risk of falling, and fall occurred}; Define the observation symbol V = [p1, p2, p3, p4, h1, h2, h3, e1, e2, e3, st, b]; where p1 is the pressure center offset; p2 is the symmetry of plantar pressure distribution; p3 is the position of the pressure peak within the gait cycle; p4 ​​is the dynamic pressure distribution range; h1 is the heart rate variability index; h2 is the heart rate acceleration / deceleration rate; h3 is the frequency of abnormal heart rate fluctuations; e1 is the muscle activation intensity; e2 is the degree of muscle synergistic activation; e3 is the muscle fatigue index; st is the stability coefficient; and b is the balance coefficient. Define the transition probability matrix A, the initial state probability distribution π, and the observation probability matrix B for each state set S; The probability of each state from the first observation to the current observation is calculated using the forward algorithm, and the probability of each state from the current observation to the last observation is calculated using the backward algorithm. The trigger strength is adjusted according to the probability of each state, and the probability of each state is incorporated into the consequent parameter of the fuzzy rule. Calculate the probability of falling based on the adjusted trigger strength and rule consequent parameters.

6. The stroke fall early warning method based on a flexible sensor according to claim 5, characterized in that, The formula for calculating the probability of falling is: ； in, This is the consequent parameter of the i-th rule; Let be the trigger strength of the i-th rule; ; , For the i-th input variable, for The membership degree value; Given an observation sequence O, let be the probability of being in state s at time t. The weights of state s; It is a set of states.

7. A stroke fall early warning device based on a flexible sensor, characterized in that, include: The data acquisition module is used to collect the patient's acceleration, angular velocity, pressure, heart rate and electromyography data in three-dimensional space in real time based on multiple flexible sensors. The flexible sensors include a flexible triaxial accelerometer, a flexible triaxial gyroscope, a flexible pressure sensor, a flexible heart rate sensor and a flexible electromyography sensor, which can be worn and attached to multiple preset positions on the patient's body, including the waist, ankle, elbow joint and knee joint. The dynamic time warping feature extraction module is used to calculate the resultant acceleration, resultant angular velocity, and resultant pressure time series collected by the lumbar flexible sensor and each ankle flexible sensor based on the patient's acceleration, angular velocity, pressure, heart rate, and electromyography data in three-dimensional space; and to calculate the stability coefficient between the lumbar acceleration sequence and the normal gait template sequence and the balance coefficient between the left and right ankle acceleration sequences based on the multi-scale DTW algorithm. A multimodal data fusion module is used to fuse stability coefficients, balance coefficients, and pressure distribution data, heart rate data, and electromyographic data collected by flexible sensors to form a multimodal feature vector. The pressure distribution data includes pressure center offset, plantar pressure distribution symmetry, pressure peak position within the gait cycle, and dynamic pressure distribution range. The heart rate data features include heart rate variability indicators, heart rate acceleration / deceleration rate, and abnormal heart rate fluctuation frequency. The electromyographic data includes muscle activation intensity, muscle synergistic activation degree, and muscle fatigue indicators. A fall prediction module is used to input the multimodal feature vector into the ANFIS fall model to predict the patient's fall risk probability. The ANFIS fall model includes an input layer, a stacked hourglass prediction layer, a network fuzzification layer, a rule layer, a normalization layer, a consequence layer, and an output layer. The network fuzzification layer includes multiple input nodes, a fuzzy rule layer, and an output node. Each input node is connected to at least one membership function, including Gaussian, bell, and trapezoidal functions. The fuzzy rule layer calculates the trigger strength of each rule based on fuzzy rules. The stacked hourglass prediction layer is used for joint posture prediction and supplementing missing joint data. It includes an initial convolutional layer, multiple hourglass modules, and an end convolutional layer. Each hourglass module includes a downsampled path encoding layer, a low-resolution feature map network layer, and an upsampled path decoding layer. The consequence layer calculates the fall risk probability based on the trigger strength and the consequent parameters of the fuzzy rules.

8. A computing device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to any one of the stroke fall warning methods based on flexible sensors in claims 1-6.

Images