Remote monitoring and health management system for fall risk in stroke patients

Abstract

The present application relates to the field of intelligent medical treatment and rehabilitation, in particular to a fall risk remote monitoring and health management system for stroke patients; containing pathological benchmark modeling, risk factor simulation generation, double-path difference extraction and feature form similarity judgment unit; the system constructs a personalized hemiplegia steady-state model based on historical rehabilitation data and defines a safety benchmark; the core is to convert the pathological mechanism into a disturbance factor to generate a theoretical waveform, and to determine the risk by comparing the morphological similarity of the real and theoretical residual characteristics; the present application establishes a personalized pathological reference benchmark, discards the general threshold determination, automatically adapts to the differentiated motion mode, and effectively reduces the false alarm caused by the uncoordinated movement of the patient.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare and rehabilitation, specifically to a remote monitoring and health management system for fall risk in stroke patients. Background Technology

[0002] With the increasing demand for stroke rehabilitation, remote monitoring and health management for hemiplegic patients have become increasingly important. Stroke patients often have motor dysfunction, making them at much higher risk of falls than the general population, requiring continuous safety monitoring. Currently, existing fall detection technologies generally use a judgment logic based on a universal threshold. This involves collecting motion data through sensors and using the gait characteristics of healthy individuals as a reference. When the monitored acceleration or angular velocity data exceeds a preset universal safety threshold, the system triggers a fall alarm. However, stroke hemiplegic patients have extremely unique pathological gait characteristics, such as circling and dragging movements. Traditional universal threshold methods are difficult to adapt to these highly individualized pathological gaits, easily misinterpreting active exertion or body swaying during rehabilitation training as a premonition of a fall, resulting in a very high false alarm rate. At the same time, for special fall patterns with low acceleration, such as flaccid paralysis and slow slips, traditional methods often fail to identify them because the signal fluctuations do not reach the threshold, leading to missed alarms and failing to achieve a balance between high sensitivity and high specificity.

[0003] Therefore, how to automatically adapt to the patient's individualized pathological gait and accurately capture various fall risks while eliminating interference from rehabilitation movements has become an urgent problem to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to provide a remote monitoring and health management system for fall risk in stroke patients, addressing the following technical problems:

[0005] The purpose of this invention is to provide a remote monitoring and health management system for fall risk in stroke patients. It can automatically adapt to the vastly different pathological gait of different patients, effectively reduce the false alarm rate caused by the patient's own incoordination or rehabilitation training, and actively capture low-acceleration fall patterns such as flaccid paralysis and slow slips, achieving a balance between high sensitivity and high specificity.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A remote monitoring and health management system for fall risk in stroke patients, including:

[0008] Pathological benchmark modeling unit: Collects real-time pathological motion data streams, trains nonlinear dynamic models based on historical rehabilitation data; constructs personalized hemiplegic steady-state models, and defines hemiplegic safety benchmark data that includes patient-specific gait characteristics;

[0009] Risk factor simulation generation unit: calls a pre-stored set of pathological mechanism parameters to convert the pathological mechanism of stroke into risk disturbance factors; injects the risk disturbance factors into the simulation synthesis module and superimposes them onto the personalized hemiplegic steady-state model to generate theoretical fall precursor data waveforms;

[0010] Dual-path differential extraction unit: performs subtraction operation between the real-time pathological motion data stream and the personalized hemiplegic steady-state model to extract the actual residual feature vector; performs subtraction operation between the theoretical fall precursor data waveform and the personalized hemiplegic steady-state model to extract the theoretical residual feature vector;

[0011] Feature morphological similarity judgment unit: Constructs a phase space, calculates the morphological similarity value between the actual residual feature vector and the theoretical residual feature vector; compares the morphological similarity value with a preset similarity threshold to generate a risk judgment result; configured as follows: if the morphological similarity value is higher than the preset similarity threshold, triggers a graded intervention instruction; if the morphological similarity value is lower than or equal to the preset similarity threshold, performs an environmental noise filtering operation.

[0012] As a further aspect of the present invention, the method for constructing the personalized hemiplegic steady-state model includes:

[0013] Historical inertial measurement data of patients during periods without falls were obtained as a training set;

[0014] An autoregressive neural network was used to learn the pathological gait fluctuation patterns of patients and predict the theoretical movement trajectory at the next moment.

[0015] Periodic swaying that conforms to the characteristics of hemiplegia is defined as fluctuation within the baseline, a personalized pathological reference baseline is established, and a personalized hemiplegic steady-state model is generated.

[0016] As a further aspect of the present invention: the method of converting the pathological mechanism of stroke into a risk disturbance factor includes:

[0017] Read the digital pathological descriptions of muscle weakness, spasticity, and central instability from the pre-stored set of pathological mechanism parameters;

[0018] The description of myasthenia gravis is transformed into a nonlinear decay coefficient, which is then injected into the ratio of the support phase time between the healthy and affected sides.

[0019] The spasm description is converted into high-frequency impulse noise and superimposed on the angular velocity data stream;

[0020] The instability of the center of gravity is described as a drift vector pointing outward, which is applied to the trajectory of the center of pressure on the plantar surface;

[0021] The risk disturbance factor is generated by integrating the nonlinear attenuation coefficient, the high-frequency pulse noise, and the drift vector.

[0022] As a further aspect of the present invention: the method for generating theoretical fall precursor data waveforms includes:

[0023] Start the simulation synthesis module and load the personalized hemiplegic steady-state model as the background carrier;

[0024] Based on the preset fault evolution logic, the risk disturbance factor is dynamically mapped to the corresponding dimension of the background carrier;

[0025] Synthesize data waveforms containing characteristics of specific pathological fall patterns as theoretical fall precursor data waveforms.

[0026] As a further aspect of the present invention: the method for extracting the actual residual feature vector and the theoretical residual feature vector includes:

[0027] In the dual-path differential extraction unit, the real-time pathological motion data stream and the theoretical fall precursor data waveform are input synchronously;

[0028] Using the personalized hemiplegic steady-state model as the common subtrahend, differential operations are performed respectively;

[0029] The pathological fluctuation background within the baseline is removed from the difference results, while the abnormal mutation components are retained;

[0030] The outputs are the real residual feature vector containing a mixture of true fall signals and environmental noise, and the theoretical residual feature vector containing pure pathological fall patterns, respectively.

[0031] As a further aspect of the present invention: the method of comparing the morphological similarity value with a preset similarity threshold to generate a risk assessment result includes:

[0032] A dynamic time warping algorithm is used to align the time axes of the actual residual feature vector and the theoretical residual feature vector in phase space;

[0033] Calculate the morphological similarity between the two.

[0034] If the morphological similarity value is higher than the preset similarity threshold, it is determined that the actual abnormal fluctuation conforms to the pathological fall pattern, and a true fall risk alarm signal is generated as the risk determination result.

[0035] If the morphological similarity value is lower than or equal to the preset similarity threshold, the abnormal fluctuation is determined to be a non-pathological interference, and a rehabilitation action or life noise label is generated as the risk assessment result.

[0036] As a further aspect of the present invention: the method of triggering the graded intervention command or performing the environmental noise filtering operation includes:

[0037] In response to the true fall risk alarm signal, the audible and visual alarm of the remote monitoring terminal is immediately activated, and an emergency data package containing the fall type is pushed to the cloud management platform;

[0038] In response to the rehabilitation action or living noise tag, the alarm trigger is suppressed, and the corresponding real residual feature vector is marked as a safe sample;

[0039] The safety sample is fed back to the pathological benchmark modeling unit to update the historical rehabilitation data and optimize the robustness of the personalized hemiplegic steady-state model.

[0040] As a further aspect of the present invention, the system further includes:

[0041] The multimodal data acquisition front end is equipped with an inertial measurement unit and a pressure sensor to acquire the patient's triaxial acceleration, angular velocity and plantar pressure center trajectory, and perform spatiotemporal alignment processing to generate the real-time pathological motion data stream.

[0042] The beneficial effects of this invention are:

[0043] 1. This invention obtains historical inertial measurement data of patients during periods without falls using a pathological benchmark modeling unit as a training set, and employs an autoregressive neural network to learn the pathological gait fluctuation patterns of patients; the system defines periodic swaying that conforms to hemiplegic characteristics as fluctuations within the benchmark, establishes a personalized pathological reference benchmark, and thus constructs a personalized hemiplegic steady-state model that defines hemiplegic safety benchmark data containing patient-specific gait characteristics; this mechanism enables the system to adapt to individual pathological gait differences through a nonlinear dynamic model;

[0044] 2. This invention uses a risk factor simulation generation unit to read digital pathological descriptions of muscle weakness, spasticity, and instability from a set of pathological mechanism parameters. These descriptions are then converted into nonlinear attenuation coefficients, high-frequency impulse noise, and drift vectors, and integrated into risk disturbance factors. These factors are then superimposed onto a personalized hemiplegic steady-state model to generate theoretical fall precursor data waveforms. A phase space is constructed using a feature morphological similarity judgment unit. A dynamic time warping algorithm is employed to align the time axis in the phase space, and the morphological similarity between the actual residual feature vector and the theoretical residual feature vector is calculated. Risk is determined through morphological comparison.

[0045] 3. This invention utilizes a dual-path differential extraction unit, with the personalized hemiplegic steady-state model as the common subtractor, to perform differential operations on the synchronously input real-time pathological motion data stream and the theoretical fall precursor data waveforms. The differential results are used to remove the pathological fluctuation background within the baseline, outputting real-world residual feature vectors and theoretical residual feature vectors that retain abnormal mutation components. Simultaneously, when generating rehabilitation actions or everyday noise labels, the corresponding real-world residual feature vectors are marked as safe samples and fed back to the pathological baseline modeling unit to update historical rehabilitation data, thereby optimizing the robustness of the personalized hemiplegic steady-state model. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the remote monitoring and health management system for fall risk in stroke patients according to the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figure 1 This invention is a remote monitoring and health management system for fall risk in stroke patients, comprising:

[0049] Pathological baseline modeling unit: configured to collect real-time pathological motion data streams, train a nonlinear dynamic model based on historical rehabilitation data; construct a personalized hemiplegic steady-state model, and define hemiplegic safety baseline data containing patient-specific gait characteristics; Risk factor simulation generation unit: configured to call a pre-stored set of pathological mechanism parameters, transform the stroke pathological mechanism into risk perturbation factors; inject the risk perturbation factors into the simulation synthesis module, superimpose them onto the personalized hemiplegic steady-state model, and generate theoretical fall precursor data waveforms;

[0050] Dual-path differential extraction unit: configured to perform subtraction operations between real-time pathological motion data stream and personalized hemiplegic steady-state model to extract real residual feature vector; and to perform subtraction operations between theoretical fall precursor data waveform and personalized hemiplegic steady-state model to extract theoretical residual feature vector;

[0051] Feature morphological similarity judgment unit: configured to construct phase space, calculate the morphological similarity value between the actual residual feature vector and the theoretical residual feature vector; compare the morphological similarity value with a preset similarity threshold to generate a risk judgment result; configured to: trigger a graded intervention instruction in response to the morphological similarity value being higher than the preset similarity threshold; and perform an environmental noise filtering operation in response to the morphological similarity value being lower than or equal to the preset similarity threshold.

[0052] This embodiment discloses a remote monitoring and health management system for fall risk in stroke patients; the system breaks away from the traditional logic based on general threshold judgment and adopts a dual-track differential and topological coupling architecture; the system mainly includes the following core processing units: the pathological benchmark modeling unit aims to establish a personalized pathological reference benchmark;

[0053] In this embodiment, the unit is configured to collect real-time pathological motion data streams from the patient and train them using long short-term memory networks or other nonlinear dynamic algorithms based on rehabilitation data from periods when the patient had no falls. The personalized hemiplegic steady-state model constructed by this unit no longer uses the gait of healthy individuals as the standard, but instead defines hemiplegic safety baseline data that includes the patient's unique gait characteristics. The hemiplegic safety baseline data is defined as a structured tensor sequence in the system storage. ,in, Represents the normalized time index within the gait cycle. This represents the total number of sampling points in a single step state period;

[0054] The sequence contains the mean vector and covariance matrix of triaxial acceleration, angular velocity, and the theoretical trajectory of the center of pressure during the gait cycle; the set of pathological mechanism parameters is configured as a lookup table in key-value pair format, storing digital pathological features including the muscle weakness attenuation coefficient, spasticity frequency distribution parameters, and the center of gravity drift vector limit.

[0055] The risk factor simulation generation unit aims to generate theoretical risk simulation objects. This unit calls a pre-stored set of pathological mechanism parameters to transform abstract medical pathological mechanisms, such as dystonia, into mathematical risk perturbation factors. These factors are then injected into the simulation synthesis module and superimposed on the aforementioned personalized hemiplegic steady-state model to generate theoretical fall precursor data waveforms. This represents the form that the sensor data should present if the patient is about to fall due to a specific pathological cause.

[0056] The dual-path difference extraction unit is designed to remove background noise; this unit performs two parallel subtraction operations: the first track calculates the difference between the real-time pathological motion data stream and the personalized hemiplegic steady-state model to obtain the actual residual feature vector; the second track calculates the difference between the theoretical fall precursor data waveform and the personalized hemiplegic steady-state model to obtain the theoretical residual feature vector; the feature morphology similarity decision unit is designed to perform pattern matching.

[0057] This unit constructs a phase space and calculates the morphological similarity value between the actual residual and the theoretical residual. If the similarity is higher than the preset similarity threshold, it indicates that the abnormal fluctuation in reality conforms to a certain pathological fall pattern, triggering a graded intervention instruction; otherwise, it is judged as environmental noise, such as active exertion in rehabilitation training, and a filtering operation is performed.

[0058] This embodiment, through the collaborative work of constructing a personalized hemiplegic steady-state model and a risk factor simulation generation unit, can automatically adapt to the vastly different pathological gait of different patients in the context of stroke patient monitoring, effectively reducing the false alarm rate caused by the patient's own incoordination. At the same time, the system actively captures low-acceleration fall patterns such as flaccid paralysis and slow slips, solving the problem of missed detection in such high-risk scenarios by the traditional threshold method, and achieving a balance between high sensitivity and high specificity.

[0059] The personalized hemiplegic steady-state model is stored in memory as a matrix containing neural network weights. Error distribution characteristics and reference time series tensor The structured object; its physical meaning is to describe the individual dynamic attractor of the patient in a risk-free state, which is a set of expected motion vectors that can dynamically evolve with the patient's gait phase.

[0060] In a preferred embodiment of the present invention, the method for constructing a personalized hemiplegic steady-state model includes:

[0061] Historical inertial measurement data of patients during periods without falls were acquired as a training set; an autoregressive neural network was used to learn the pathological gait fluctuation patterns of patients and predict the theoretical motion trajectory at the next moment; periodic swaying that conforms to the characteristics of hemiplegia was defined as fluctuation within the benchmark, a personalized pathological reference benchmark was established, and a personalized hemiplegic steady-state model was generated.

[0062] This embodiment further defines the specific implementation logic for constructing a personalized hemiplegic steady-state model; to accurately capture the patient's unique gait patterns, this embodiment uses an autoregressive neural network to learn the intrinsic dynamic characteristics of time-series data; [Definition] for The inertial measurement state vector at time t, including triaxial acceleration, triaxial angular velocity, and coordinates of the two-axis plantar pressure center. Axis, total Each channel has a component; the model predicts the theoretical state at the next moment using historical window data. Its prediction model is as follows:

[0063]

[0064] in, The theoretically predicted state vector, derived from model calculations, physically represents the system's expected motion state at the next moment. The nonlinear mapping function originates from fitting a multilayer long short-term memory network, but can also employ other autoregressive neural network structures such as gated recurrent units and recurrent neural networks.

[0065] Input layer dimension set to Among them, 8 corresponds to the number of channels for triaxial acceleration, triaxial angular velocity, and biaxial plantar pressure center. The hidden layer uses a two-layer stacked structure with 128 and 64 nodes respectively, and is configured with activation functions to capture long-term temporal dependencies; to clarify the nonlinear mapping function... In this embodiment, the LSTM unit performs the specific calculations at time steps. The internal update logic is as follows:

[0066] definition The input vector at the current moment is a slice of historical window data. This is the hidden state from the previous moment. The cell state at the previous time step; forget gate The calculation formula is:

[0067]

[0068] in, Here is the weight matrix for the forget gate. This is the bias vector for the forget gate;

[0069] Input gate The calculation formula is:

[0070]

[0071] Candidate cell state The calculation formula is:

[0072]

[0073] in, It is a sigmoid activation function. This represents the Hadamard product, multiplied element by element. and Network weight parameters A subset; the final predicted output From a hidden state The result obtained after mapping through a fully connected layer is:

[0074]

[0075] in, and These are the weight matrix and bias vector of the output fully connected layer, respectively, both belonging to the global parameter set. Part of;

[0076] The output layer is a fully connected layer that maps back to the 8-dimensional sensor space; the network is trained using an optimizer with an initial learning rate set to 0.001 and a loss function defined as the mean square error between the predicted and true values. The length of the time window is derived from the patient's gait cycle setting, and its physical meaning is the minimum amount of historical data required to capture gait characteristics; Network weight parameters are derived from training set optimization and physically represent a memory matrix storing patients' personalized exercise habits.

[0077] In this embodiment, the fluctuation within the benchmark is defined as the prediction error. The case where the value is less than the statistical significance boundary; among which, represent The actual observation vector acquired by the sensor at any given moment. The theoretical prediction vector represents the output of the model; the system calculates the prediction error sequence for all time steps in the training set, and calculates the mean error vector for each of the eight sensor dimensions. with the standard deviation vector of the error ;

[0078] The statistical significance boundary is set as a vector. This refers to the three-standard-deviation criterion for each dimension; this criterion is based on the empirical rule of the normal distribution, and is established by setting... The boundary, theoretically, can cover The normal gait fluctuation samples are used to significantly reduce false alarms.

[0079] This means that the condition for determining fluctuations within the baseline is that the absolute value of the prediction error for each dimension of the real-time data is less than the vector. The corresponding component value;

[0080] As long as the patient's real-time movements conform to their previous hemiplegic gait pattern, even if that pattern is medically abnormal, the model prediction value will be acceptable. It can keep up with real-time values This causes the difference result to fall within the boundary and approach zero;

[0081] This embodiment successfully establishes a personalized pathological reference benchmark by defining periodic swaying that conforms to the characteristics of hemiplegia as fluctuations within the benchmark. This makes the system extremely robust when facing rehabilitation training scenarios, and can automatically ignore non-fall swaying caused by the patient's weak limb control ability, thereby focusing on detecting danger signals that truly deviate from their inherent movement patterns, avoiding the inherent detection logic bias of traditional algorithms that misjudge pathological gait as a premonition of a fall.

[0082] In a preferred embodiment of the present invention, the method of converting the pathological mechanism of stroke into a risk perturbation factor includes:

[0083] Read the digital pathological descriptions of myasthenia gravis, spasticity, and instability from the pre-stored set of pathological mechanism parameters; convert the myasthenia gravis description into a nonlinear decay coefficient and apply it to the support phase time ratio parameter between the healthy and affected sides; convert the spasticity description into high-frequency impulse noise and superimpose it onto the angular velocity data stream.

[0084] The instability of the center of gravity is described as a drift vector pointing outward, which is applied to the trajectory of the center of plantar pressure; the nonlinear attenuation coefficient, high-frequency impulse noise and drift vector are integrated to generate a risk disturbance factor.

[0085] This embodiment details the specific digital process of transforming the pathological mechanisms of stroke into risk perturbation factors; this is the core step of twin simulation, aiming to transform qualitative medical descriptions into quantitative mathematical operators; the system's pre-stored set of pathological mechanism parameters includes digital descriptions of myasthenia gravis, spasticity, and instability of the center of gravity, and the specific transformation logic is as follows: parameterization for myasthenia gravis; myasthenia gravis is mainly manifested as a decrease in the support capacity of the affected limb;

[0086] This embodiment introduces a nonlinear attenuation coefficient. Inject it into the support phase time ratio parameter; let For the duration of support on the healthy side, The affected side support time, and the affected side support time after perturbation. The calculation is as follows:

[0087]

[0088] in, Index representing the cumulative number of steps taken in the current continuous walking. This ensures that the fatigue effect increases monotonically with the number of steps. The amplitude of muscle strength decline is a dimensionless proportionality coefficient derived from patient muscle strength grading data. The specific mapping logic is as follows: it maps the muscle strength decline amplitude from the clinical Loft grading system... Level linear mapping to The space is calculated using the following formula:

[0089]

[0090] in, This refers to the patient's clinically assessed muscle strength grade; its physical meaning is the degree of loss of muscle contraction ability; to ensure the non-negativity of physical time, the range of values ​​for this coefficient is strictly limited to [value missing]. Within the range, avoid excessive attenuation. The physical paradox of negative values; Fatigue accumulation rate, dimensionless , originating from a preset constant; among which, Refers to the count of each independent gait cycle of the patient's footsteps.

[0091] In this embodiment, the preset constant The value range is set to 0.005 to 0.05; the specific value is determined according to the patient's recovery stage. For example, the value is close to 0.05 for patients in the acute phase to simulate rapid fatigue, while the value is close to 0.005 for patients in the recovery phase. The physical meaning is the rate at which muscle strength decreases during walking. Step count index, unit: This comes from an internal event counter within the software; this counter is reset to zero each time the system transitions from standby mode to activity monitoring mode. As can be seen from the formula, the exponential term The result is a dimensionless number, which ensures the physical rigor of the mathematical calculation;

[0092] This formula accurately simulates the cumulative effect of fatigue: when That is, at the beginning of walking, the decay term is 0, and the support time is normal; as... As the term increases, the decay term gradually approaches the mean. This leads to a significant reduction in the support time on the affected side, which is consistent with pathophysiological patterns.

[0093] Parameterization for spasticity; spasticity is characterized by sudden, involuntary muscle contractions; this embodiment converts it into high-frequency impulse noise. Superimposed on the angular velocity data stream middle:

[0094]

[0095] in, The symbol for summation; This represents the total number of spasms that occurred within the simulation window; The physical unit for the equivalent pulse intensity of spasm is set to radians; Represents the Dirac function; to simulate the instantaneous change in velocity at the kinematic level, the system directly uses... With the original angular velocity sequence Perform algebraic superposition;

[0096] The values ​​were obtained through Monte Carlo sampling, and their mean values ​​were positively correlated with the envelope amplitude of abnormal discharge segments in the patient's historical electromyography; here, displacement was used. Multiply by the unit impulse function The form is designed to simulate instantaneous velocity changes at the kinematic level, rather than torque impulses at the dynamic level;

[0097] The intensity of the spasmodic impulse under an angular velocity field is defined in radians. This physical quantity represents the cumulative amount of angular change over an extremely short period of time; given the Dirac function... In time-domain analysis, it has dimensional properties, i.e. The product operation of the two The generating unit is The physical quantity, thus obtaining the relationship with angular velocity. Consistent dimensions ;

[0098] During the simulation synthesis process, since there is no real-time electromyography (EMG) signal input, the system is configured to generate this parameter using Monte Carlo sampling: a Weiber distribution is fitted based on the peak statistical data of the patient's historical EMG, which includes shape parameters. Describes the central tendency of spasm intensity, dimensionless and scale parameter. The characteristic amplitudes describing the intensity of spasms, measured in radians, are randomly sampled from these amplitudes to obtain the current simulation result. Value; the specific Monte Carlo sampling logic is as follows: The Weiber probability density function is obtained by fitting historical data:

[0099]

[0100] During the simulation generation phase, a simulation is generated. Uniformly distributed random numbers in an interval The current spasm intensity is calculated using the inverse transform sampling method. :

[0101]

[0102] This formula ensures that the intensity of the generated spasm strictly follows the patient-specific pathological statistical distribution; this approach not only solves the problem of missing input in the simulation environment, but also covers a variety of pathological scenarios from mild twitching to tonic spasm by introducing a probability distribution.

[0103] The timing of spasms is randomly generated. To conform to pathophysiological characteristics, the random generation is not completely disordered, but rather set to follow a uniformly distributed sampling pattern within the swaying phase of the gait cycle, i.e., the 60%-100% stage of the gait cycle, or based on a Poisson process to generate time intervals, physically signifying the unpredictability of spasm occurrence. To clarify the mathematical form of high-frequency impulse noise, this embodiment... Specifically defined as a discretized Gaussian pulse approximating the Dirac function:

[0104]

[0105] in, Pi is a constant. The base of the natural logarithm; the denominator in the formula This represents the square root operation; The pulse width parameter is set to a range of 10ms to 50ms to simulate the instantaneous discharge characteristics of neural electrical signals. Specifically, considering the discrete sampling characteristics of digital systems, the aforementioned continuous function is used in actual calculations. The sampled sequence is mapped to a discrete-time grid and energy normalization is performed to ensure that the sampling period is within a certain range. The following satisfies To prevent pulse energy loss due to insufficient sampling rate;

[0106] Parameterization for center of gravity instability; center of gravity instability manifests as the body's inability to maintain balance along the midline; this embodiment defines a drift vector. The trajectory of the center of pressure applied to the sole of the foot superior:

[0107]

[0108] in, It is a monotonically increasing vector pointing outwards from the affected side, simulating the process of the body gradually losing balance control; in this embodiment, this vector is concretized as a restricted linear drift model:

[0109]

[0110] in, : This is the relative time variable after the simulation event is triggered, with the physical unit being seconds, and its domain is... ,in, This is a preset time window for warning signs of a fall, such as 2.0 seconds; this definition avoids issues if global time is used. The logic error that caused the drift to increase indefinitely ensured that the simulation only modeled the instant before the fall due to the failure of balance control;

[0111] : is the drift rate constant, with a value range of to ; The maximum permissible drift amplitude, for example, 15cm, is derived from the physical range boundary of the plantar pressure sensor or the human body's stability limit; this cutoff function is introduced. The purpose is to prevent the virtual data generated by the mathematical model from exceeding the physical measurement range of the sensor, which could lead to numerical overflow or non-physical infinite drift in subsequent simulations. : is a unit direction vector pointing towards the affected side, explicitly defined in the sensor coordinate system as:

[0112]

[0113] Among them, parameters This is a lateral symbol. This parameter is a hemiplegic side indicator pre-stored in the patient's file; 1 represents right hemiplegia, and -1 represents left hemiplegia, ensuring the mathematical and logical uniqueness and correctness of the drift direction; integrated generation; the above. , and Together, they constitute the risk disturbance factors.

[0114] In a preferred embodiment of the present invention, the method for generating theoretical fall precursor data waveforms includes:

[0115] The simulation synthesis module is activated, and a personalized hemiplegic steady-state model is loaded as the background carrier. Based on the preset fault evolution logic, the risk disturbance factors are dynamically mapped to the corresponding dimensions of the background carrier. Data waveforms containing specific pathological fall pattern characteristics are synthesized as theoretical fall precursor data waveforms.

[0116] This embodiment describes the process of generating theoretical fall precursor data waveforms, i.e., the synthesis step; after starting the simulation synthesis module, the system loads a personalized hemiplegic steady-state model as a background carrier; this carrier represents the patient's current safety state, and its corresponding time series data is uniformly denoted as […]. That is, the steady-state model prediction output, this sequence This refers to the hemiplegic safety baseline data tensor defined in the aforementioned invention. The specific instantiation behavior within the current simulation time window ensures the consistency of symbol definitions within the system;

[0117] Based on the preset fault evolution logic, for example: muscle weakness occurs first, leading to center of gravity shift and inducing spasms, and the generated risk disturbance factors are dynamically mapped to the corresponding dimension of the background carrier;

[0118] Considering that risk factors include time-dimensional parameters, such as changes in the support phase time ratio caused by myasthenia gravis, and amplitude-dimensional parameters, such as spastic noise, this embodiment refines the synthesis formula into a spatiotemporal two-level cascaded mapping process to ensure the closed loop of physical logic:

[0119] Level 1: Time-dimensional mapping; non-linear decay coefficient for myasthenia gravis. The system constructs a time warping function; to achieve the identification of support phase intervals in the background carrier, this embodiment adopts an automatic segmentation algorithm based on pressure threshold: traversing the background carrier. The total plantar pressure channel To satisfy all The time index set is marked as the supporting phase;

[0120] Among them, threshold In this embodiment, the physical boundary for determining effective ground contact is derived from the signal-to-noise ratio characteristics of the plantar pressure sensor, typically set to a fixed value of 30N to 50N, or preferably set to 5% of the patient's body weight. This proportion is based on biomechanical statistics, as the spurious pressure generated by insole bending during the swing phase of normal walking is usually less than the patient's body weight. Therefore, set It can be effectively used as a ground contact threshold, thereby effectively eliminating false pressure signals caused by insole bending during the swing phase;

[0121] Interpolation scaling is applied; to strictly match the nonlinear fatigue accumulation model and prevent timing logic errors caused by simple linear scaling, this embodiment explicitly applies a nonlinear time mapping here: making the duration of this interval... Duration after perturbation :

[0122]

[0123] in, This represents the current simulation step number. To clarify the specific logic for obtaining this step number index during continuous simulation, this embodiment specifies that the counter is initialized at the simulation start time. ; in traversing the background carrier In order to eliminate logical loops, the system pre-processes the background carrier before starting the simulation. Perform full-cycle phase pre-labeling; systematically traverse undeformed... Extract all potential touchdown index sequences ;

[0124] When performing the first-level time mapping, the current step index... By directly referencing this pre-labeled sequence, it is ensured that the system has determined the gait cycle number to which it belongs before the duration of the support phase shortens; The waveform change, whenever a condition is detected that satisfies The counter executes at the rising edge, i.e., when a new heel strike event occurs. The cumulative operation ensures that the fatigue accumulation effect is strictly synchronized with the actual walking cycle in the background carrier.

[0125] This means that as the simulation progresses, the support time on the affected side will show an accelerated shortening trend according to the fatigue curve, rather than a fixed percentage shortening, thus ensuring the authenticity of the simulation data.

[0126] The duration of the oscillation phase remains constant. To eliminate the discontinuity at the splicing points caused by time-domain deformation, i.e., the Gibbs phenomenon, which may introduce spurious high-frequency noise, this embodiment specifically introduces a boundary smoothing constraint mechanism. In the compressed support phase... With constant oscillation phase At the connection point, a length of [length] is defined. For example, in a 50ms transition period, weighted cross-fade-in and fade-out processing is performed within this region:

[0127]

[0128] in, The smoothing weights generated for the Sigmoid function ensure that the data stream remains not only numerically continuous at the splice point, but also that the first derivative, i.e., the velocity, remains continuous. To ensure that the weights smoothly transition from 0 to 1 in the transition region, this embodiment defines a normalized time variable. Mapping to actual transition time ; Specific smoothing weight function The structure is as follows:

[0129]

[0130] in, Represented by the natural constant An exponential function with base 0; This represents the relative time position of the current point within the transition zone. For the corresponding steepness coefficient, this embodiment preferably uses... To ensure and This achieves near-perfect seamless stitching; cubic spline interpolation is used to resample the deformed and smoothed waveform to the original sampling rate, generating an intermediate waveform with temporal pathological characteristics. ;

[0131] Level 2: Spatial amplitude mapping; targeting spasm noise With drift vector To address the dimensionality mismatch issue and accurately simulate the response differences of multi-axis sensors, the system constructs a sparse perturbation vector aligned with all dimensions of the sensor. ,in, It includes triaxial acceleration, triaxial angular velocity, and biaxial pressure center;

[0132] Connection matrix The cross-connected elements in the code possess physical dimension transformation properties; specifically, the element that maps rotational angular velocity to a velocity channel has dimensions of... The element that maps the center of pressure to the velocity channel has dimensions of After matrix mapping, the dimensions of each channel are compared with the background carrier. Maintaining consistency eliminates dimensional inconsistencies caused by direct algebraic addition across physical domains;

[0133] The specific operation is as follows: Map to the dimension index corresponding to the angular velocity, such as index 4-6, and... The components are mapped to the dimension indices corresponding to the pressure center, such as indices 7-8, while the remaining acceleration dimensions are set to zero; the sensitivity coupling matrix is ​​applied. After weighted superposition, the synthesis formula is modified as follows:

[0134]

[0135] in, The theoretical fall precursor data waveform is derived from cascaded calculations, and its physical meaning is the sensor signal assumed to occur when a specific pathological fall occurs. Intermediate data after time mapping originates from the first-level processing. :for The diagonal sensitivity matrix, whose diagonal elements Representing the The response coefficient of each sensor channel to a specific pathological disturbance; It is a sparse perturbation vector containing risk perturbation factors;

[0136] In this embodiment, And the channel index is agreed upon. The mapping relationship with the physical sensor channels is shown in the table below to ensure that the physical orientation of the matrix operations is clear: X-axis acceleration ; Y-axis acceleration ; Z-axis acceleration ; Angular velocity along the X-axis ; Angular velocity along the Y-axis ; Z-axis angular velocity ; X-axis pressure center coordinates ; Y-axis pressure center coordinates Therefore, the sensitivity matrix is ​​defined. diagonal elements Each of the above physical channels corresponds one-to-one;

[0137] To overcome the parameter uncertainties caused by relying on a general biomechanical model, this embodiment pre-defines a pathology-sensor mapping lookup table to determine matrix elements. The specific content of this lookup table is derived from the statistical mapping of biomechanical simulation and clinical data:

[0138] A hemiplegic human model was constructed using open-source biomechanical simulation software and other dynamic simulation software. Standardized pathological perturbations, such as spastic torques of specific Newton-meters, were applied, and the response amplitude ratios at each sensor installation location were observed. Based on the simulation results, the energy transfer coefficients between the main response axis and the coupling axis were determined.

[0139] The specific configuration is as follows: For the spasm factor, focusing on angular velocity mutations, the diagonal elements of the corresponding angular velocity dimension are set to... ,Right now The acceleration dimension is set to ,Right now Simulates flutter propagation, with all others set to zero; for the center of gravity instability factor, primarily based on pressure center drift, the diagonal elements corresponding to the CoP dimension are set to... The acceleration dimension is set to Simulates torso tilt, with the rest set to zero;

[0140] Regarding myasthenic factor, since its main characteristics have already been reflected in the first-level time mapping, here... The matrix is ​​set to either an identity matrix or a zero matrix; this explicit parameter mapping ensures the reproducibility of the synthesized waveform; through the above two-level mapping, this embodiment ensures that the synthesized waveform conforms to the real pathological evolution pattern in both rhythm and morphology.

[0141] In a preferred embodiment of the present invention, the method for extracting the actual residual feature vector and the theoretical residual feature vector includes:

[0142] In the dual-path differential extraction unit, real-time pathological motion data streams and theoretical fall precursor data waveforms are input synchronously; differential operations are performed separately using the personalized hemiplegic steady-state model as the common subtraction; the pathological fluctuation background within the baseline is removed from the differential results, while the abnormal mutation components are retained; and the real residual feature vector containing a mixture of true fall signals and environmental noise, and the theoretical residual feature vector containing pure pathological fall patterns are output separately.

[0143] This embodiment details the working mechanism of the dual-path differential extraction unit; this unit simultaneously receives two signals: a real-time pathological motion data stream from a sensor, defined here as... And theoretical fall precursor data waveforms from the simulation unit The unit uses the predicted output of a personalized hemiplegic steady-state model. As a common subtrahend, a difference operation is performed; the difference operation is strictly defined as point-by-point vector subtraction under time synchronization, that is, at any given time... Subtract the baseline value of the corresponding dimension of the steady-state model from each component of the real-time data vector; extract the real-world residual feature vector. :

[0144]

[0145] In this operation, It effectively counteracts pathological fluctuations in the patient's daily gait, such as dragging, making... Only the anomalous mutation component is retained, which could be a genuine fall or an environmental collision; theoretical residual feature vectors are extracted. :

[0146]

[0147] because It is by Therefore, it is generated by superimposing risk factors. In fact, it purely represents the characteristic fingerprint of the pathological fall pattern, that is, a pure fault signal;

[0148] This embodiment uses a dual-track differential mechanism and a feature decoupling strategy to transform the complex anomaly detection problem into a relatively simple signal matching problem. The system does not need to search for fall features in the raw, noisy sensor data, but instead compares whether the actual anomaly matches the theoretical fault in the residual domain after removing background noise. This processing method greatly improves the signal-to-noise ratio, and can still maintain high-precision detection capability, especially in complex pathological scenarios where patients have severe tremors.

[0149] In a preferred embodiment of the present invention, the method for comparing the morphological similarity value with a preset similarity threshold to generate a risk assessment result includes:

[0150] A dynamic time warping algorithm is used to align the time axes of the actual residual feature vector and the theoretical residual feature vector in phase space; the morphological similarity between the two is calculated.

[0151] If the morphological similarity value is higher than the preset similarity threshold, it is determined that the actual abnormal fluctuation conforms to the pathological fall pattern, and a true fall risk alarm signal is generated as the risk assessment result.

[0152] If the morphological similarity value is lower than or equal to the preset similarity threshold, the abnormal fluctuation is determined to be a non-pathological interference, and a rehabilitation action or life noise label is generated as the risk assessment result.

[0153] This embodiment illustrates the specific algorithm implementation of the feature morphological similarity decision unit; for comparison and Considering the similarity between the two, and taking into account the potential nonlinear scaling on the time axis (e.g., the speed at which a fall occurs in reality is faster than in the theoretical model), this embodiment employs a dynamic time warping algorithm. A phase space is constructed, defined here as a high-dimensional Euclidean space, where each point represents the complete dynamic state of the system at a certain moment. Specifically, this embodiment uses the delayed coordinate embedding method to reconstruct the trajectory in this phase space from a one-dimensional or low-dimensional sensor time series, in order to reveal the implicit nonlinear dynamic attractor structure.

[0154] This embodiment employs a multivariate shared embedding strategy to reconstruct the phase space, preserving the coupling relationship between different sensor channels: for multidimensional residual signals Its dimensions are , here Consistent with the aforementioned number of sensor channels, dimensional normalization processing as described in the multimodal data acquisition front-end must be performed, using the reciprocal of the historical statistical standard deviation as a weighting coefficient to normalize the acceleration. angular velocity and pressure center A unified mapping is performed to dimensionless standardized amplitudes, making them dimensionless vectors to eliminate numerical weight biases caused by different physical units for acceleration, angular velocity, and pressure center; its Euclidean norm sequence is then calculated. A scalar sequence representing the total energy of the signal is obtained;

[0155] For this scalar sequence The first minimum point of the average mutual information function is calculated to determine the uniform time delay parameter. Specifically, the sequence The value is dynamically divided into Histogram intervals, here The total number of intervals, with the preferred value range. to The aim is to balance the accuracy and computational complexity of mutual information computation, calculating data points falling into the first... The probability of each interval and delay point pairs Falling into the range joint probability Based on the formula:

[0156]

[0157] in, Indicates the delay time as Interactive information; This indicates that the data point falls into the first... The probability of each interval. This indicates that the data point falls into the first... The probability of each interval; Indicates the time delay point pair falling into the interval and interval The joint probability;

[0158] Calculate the mutual information sequence and select... The number of lag steps when the value first drops to a local minimum is used as To ensure the nonlinear independence of the reconstructed coordinates; the embedding dimension is determined using the spurious nearest neighbor algorithm. To reduce the complexity of high-dimensional space computations and ensure the topological consistency of multidimensional data, this embodiment explicitly implements a structural homomorphism strategy: that is, the scalar sequence obtained from the aforementioned calculations... That is, the total energy trajectory of the signal, which serves as the input object for the false nearest neighbor algorithm;

[0159] Calculation in In 3D space The distance evolution between the reconstructed vector and its nearest neighbor, when the distance change rate exceeds a threshold When a node is identified as a false neighbor, this threshold is applied. The typical value range is: to In this embodiment, the empirical values ​​recommended by Kennel are preferred. The determined optimal dimension This will be used as a unified structural parameter for all original sensor channels. Phase space reconstruction;

[0160] This is based on a physical assumption: that multimodal data generated by the same human motion system, including acceleration, angular velocity, and pressure, are controlled by the same underlying dynamic attractor and therefore share the same topological dimension; whereby... The macroscopic size of the attractor is derived from the statistical characteristics of the overall distribution of the input signal; the specific calculation formula is as follows:

[0161]

[0162] in, Time series Total length, This is the arithmetic mean of the sequence; here... Characterizing the macroscopic scale of phase space attractors in Euclidean space; introducing The physical significance of the normalization factor is to eliminate the influence of signal amplitude scaling on the embedding dimension determination, and to ensure that the algorithm maintains a consistent criterion for determining false neighbors even when the signal is weak, such as during a quiescent period, or strong, such as during a fall, thus preventing the dimension estimation from being too high due to small-scale distance fluctuations dominated by noise.

[0163] Based on certainty and For the original multidimensional signal Perform stacking reconstruction to construct a high-dimensional phase space state vector. The dimension of this vector is This preserves the nonlinear dynamic topology of the signal completely;

[0164] Then, the morphological similarity between the two is calculated. To achieve the technical effect of focusing on morphological evolution rather than amplitude, this embodiment uses cosine distance as a local cost function in the dynamic time warping algorithm; any two phase space state vectors Local distance between Defined as:

[0165]

[0166] in, This represents the vector dot product operation. The L2 norm of a vector is used to represent its Euclidean norm. To prevent small, numerically stable terms with a denominator of zero, for example Based on this local distance, the dynamic time warping algorithm plans an optimal curved path. This results in the cumulative cost along the path. Minimum;

[0167] The specific optimal path planning is achieved by constructing a cumulative cost matrix. Implementation; Setting The sequence length is , The sequence length is any point in the matrix The cumulative cost follows the following dynamic programming recursive equation:

[0168]

[0169] in, This indicates taking the minimum cumulative cost of three adjacent nodes; the boundary conditions are set as follows: The remaining boundaries are set to infinity; the cumulative cost of the final optimal path is... This step ensures a globally optimal match under nonlinear time warping.

[0170] To avoid interference from differences in time series lengths in similarity determination, i.e., to prevent similar waveforms over long periods from being misclassified as different classes due to large cumulative distances, this embodiment must perform path length normalization on the cumulative cost; a normalized morphological distance is defined. as follows:

[0171]

[0172] in, This is the step length of the optimal path; this step ensures that the similarity metric is only related to the waveform topology and decoupled from the event duration; the final morphological similarity value... Defined as:

[0173]

[0174] Regarding the preset similarity threshold Specific configuration logic: This embodiment adopts an adaptive calibration method based on the subject operating characteristic curve; during the initialization phase, the system uses a validation set containing simulated fall samples and daily high-intensity rehabilitation action samples for testing, traversing... The threshold for the interval is generated based on the receiver operating characteristic curve, and the Youden index is calculated. The calculation formula is:

[0175]

[0176] in, Indicates the Yoden Index; Indicates sensitivity; Indicates specificity;

[0177] The optimal value is selected from the points corresponding to the maximum Youden index. To achieve the best balance between sensitivity and specificity; simulated fall samples consist of virtual positive samples automatically generated in batches by the risk factor simulation generation unit based on the patient's current homeostasis model; while daily high-intensity rehabilitation action samples are extracted from large-amplitude safe movements in the patient's historical data;

[0178] This threshold calibration method based on digital twins solves the cold-start problem of traditional methods, which struggle to obtain real patient fall data, thus ensuring the accuracy of the threshold. The settings are based on sufficient statistical evidence and are practically feasible;

[0179] The risk assessment logic is as follows: in response to That is, by setting a preset similarity threshold, the system determines that the actual abnormal fluctuations match the pathological fall pattern and generates a true fall risk alarm signal; in response to The system determines that although the abnormal fluctuations in reality are large, they do not conform to any known pathological fall mechanism in terms of form. It may be a sudden physical collision in the environment or a large-scale movement during rehabilitation training, and generates rehabilitation movement or life noise tags.

[0180] In a preferred embodiment of the present invention, the methods for triggering graded intervention commands or performing environmental noise filtering operations include:

[0181] In response to a genuine fall risk alarm signal, the remote monitoring terminal immediately activates its audible and visual alarms and pushes a fall-related emergency data package to the cloud management platform.

[0182] In response to rehabilitation movements or noise tags, alarm triggering is suppressed, and the corresponding real-world residual feature vectors are marked as safe samples. The safe samples are fed back to the pathological baseline modeling unit to update historical rehabilitation data and optimize the robustness of the personalized hemiplegic steady-state model.

[0183] This embodiment describes the system's graded intervention and closed-loop feedback mechanism; alarm response mechanism; in response to the system generating a true fall risk alarm signal, the remote monitoring terminal's audible and visual alarm is immediately activated, and an emergency data package containing the fall type, such as left-sided spastic fall, is pushed to the cloud management platform;

[0184] This provides emergency responders with crucial pathological information within the golden rescue time; negative sample feedback and model optimization mechanism; in response to the system generating rehabilitation actions or life noise labels, it not only suppresses alarms, but also marks the corresponding real-world residual feature vectors as safe samples;

[0185] Before marking a sample as safe, the system performs a second confidence check: calculating the peak characteristics of the actual residual vector. ,like , As shown in Example 2, even if the morphological similarity is low, the system still suspends the sample and prompts for manual review instead of directly entering it into the training set. This is to prevent serious fall precursors from being missed by the algorithm and contaminating the baseline model. The safe sample is fed back to the pathological baseline modeling unit as new training data to update the historical rehabilitation data. The update logic is as follows:

[0186]

[0187] in, The global parameter set of the steady-state model, and the weight parameters of the neural network. The symbols remain consistent, originating from the previous round of training, and their physical meaning represents the system's current cognitive model.

[0188] The learning rate, derived from preset hyperparameters, physically represents the speed at which the model adapts to new data. For gradient operators, it represents the partial derivative of the loss function with respect to the global parameter set; Safety sample data, derived from filtered noisy data, physically represents the patient's new safe action pattern; here, the loss function... The mean square error (MSE) is explicitly defined as the difference between the predicted trajectory and the actual safe trajectory.

[0189]

[0190] in, This represents the length of the safe sample sequence used in this training, which is different from the number of sensor channels. , The vector of real sensor data that has been collected and labeled as safe is used; the function aims to minimize the model’s prediction residuals for the newly labeled safe samples, thereby incorporating the new action patterns into the steady-state baseline.

[0191] This embodiment utilizes an online learning mechanism, enabling the system to continuously learn about the patient's new rehabilitation movements and lifestyle habits over time. This self-evolutionary capability makes the personalized hemiplegic steady-state model increasingly robust, with the false alarm rate decreasing exponentially over time, thereby achieving dynamic synchronization between the system and the patient's rehabilitation process.

[0192] In a preferred embodiment of the present invention, the system further includes: a multimodal data acquisition front end, configured with an inertial measurement unit and a pressure sensor, for acquiring the patient's triaxial acceleration, angular velocity and plantar pressure center trajectory, and performing spatiotemporal alignment processing to generate a real-time pathological motion data stream.

[0193] This embodiment specifies the hardware configuration of the system's data acquisition front-end; the system is equipped with a multimodal data acquisition front-end, including: an inertial measurement unit (IMU), worn on the patient's waist or affected limb, for acquiring triaxial acceleration. and angular velocity A pressure sensor, built into the smart insole, is used to collect the plantar pressure distribution and calculate the center of gravity (CoP) trajectory of the plantar pressure. The CoP calculation specifically uses the weighted centroid method.

[0194]

[0195] in, Indicates all of the soles of the feet The values ​​from each pressure sensor are summed. The total number of pressure sensors inside the insole, for example , For the first Real-time readings from each pressure sensor. Let the sensor be the fixed position vector in the insole coordinate system, containing Coordinates; to avoid calculation singularities caused by the sum of pressures being zero in the oscillating phase, a small regularization term is introduced into the denominator of the formula. In this embodiment, it is set to This value is slightly higher than the sensor's noise floor, ensuring stable calculations under no-load conditions.

[0196] Meanwhile, the system is configured with a stress activation threshold. ,when At that time, directly Set as the geometric center Based on this, spatiotemporal alignment and dimensional normalization are performed; due to the different sampling rates of different sensors, resampling and synchronization are required.

[0197] Meanwhile, given the acceleration dimension Angular velocity dimension Dimensions of the trajectory of the center of plantar pressure Directly concatenating physical quantities with different dimensions can lead to an imbalance in numerical weights during subsequent calculations. Therefore, the system introduces a normalization coefficient to map each physical quantity to a unified dimensionless space. The corrected data flow generation formula is as follows:

[0198]

[0199] in, Real-time pathological motion data stream originates from multi-sensor fusion and normalization processing, and its physical meaning is a comprehensive motion state vector containing attitude and balance information; The delay compensation amount comes from hardware calibration, and its physical meaning is to eliminate the time misalignment caused by transmission delay; Normalized weighted coefficients are derived from the statistical characteristics of patients' historical data;

[0200] Specifically, these coefficients are taken as the reciprocal of the standard deviation of the corresponding physical quantity in historical rehabilitation data; this is to prevent the standard deviation from being affected during periods of patient rest or low activity. The coefficient approaches zero. Numerical divergence occurs, i.e., noise amplification effect. This embodiment introduces a background noise constraint mechanism, and the formula is modified as follows:

[0201]

[0202] in, The inherent physical noise basis of the sensor, such as accelerometers. Furthermore, in order to adapt to the long-term changes in patient recovery, the standard deviation... Instead of being a fixed value, it is dynamically updated using the Welford online algorithm or exponential moving average:

[0203]

[0204]

[0205] in, for Raw sensor data collected at all times A forgetting factor is defined to assign higher weights to recent data, enabling the model to adapt to non-stationary changes during the recovery process, such as... This ensures that the normalized parameters can smoothly track changes in the patient's motor ability; the generated This refers to real-time pathological motion data stream; forgetting factor. The design is based on the balance model's ability to track rehabilitation progress and its robustness to transient disturbances.

[0206] Furthermore, in the aforementioned CoP calculation formula, the regularization term... Set as Newton, activation threshold Set as patient weight This design ensures that sensor background noise does not cause problems when the foot is completely off the ground. Random shift of coordinates;

[0207] This embodiment preferably uses the exponential moving average method to adapt to non-stationary recovery processes, as shown in the aforementioned formula; as an alternative, if higher numerical stability is required, the Welford algorithm can be used, whose... mean of steps With auxiliary sum of squares The update logic is as follows:

[0208]

[0209] in, For the first Real-time sensor readings at each sampling point; the fusion and normalization of multimodal data provide precise and physically consistent dimensional support for the aforementioned pathological mechanism analysis, ensuring the model's ability to capture subtle signs such as center of gravity instability, thereby significantly improving the system's perception accuracy of early fall risk;

[0210] System initialization and real-time performance assurance; for newly admitted patients, the system enters a period of... A silent background data acquisition period of 24 hours; during this period, the system only collects data without triggering automatic decisions, using the observation period data to complete the initial convergence of the Long Short-Term Memory network model and Initial calibration;

[0211] To meet real-time requirements, the system employs sliding window parallel computation: Long Short-Term Memory network inference and simulation synthesis run in a high-priority background thread, and their computation time is limited to... Within a specified range; Dynamic time warping employs a path optimization algorithm based on band constraints to ensure a consistent end-to-end delay from anomaly occurrence to alarm triggering. This meets the needs of clinical emergency care.

[0212] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A remote monitoring and health management system for fall risk in stroke patients, characterized in that, include: Pathological baseline modeling unit: Collects real-time pathological motion data streams and trains a nonlinear dynamic model based on historical rehabilitation data; Construct a personalized hemiplegic steady-state model and define hemiplegic safety baseline data that includes patient-specific gait characteristics; Risk factor simulation generation unit: calls a pre-stored set of pathological mechanism parameters to convert the pathological mechanism of stroke into risk perturbation factors; The risk disturbance factor is injected into the simulation synthesis module and superimposed on the personalized hemiplegic steady-state model to generate theoretical fall precursor data waveforms; Dual-path differential extraction unit: performs subtraction operation between the real-time pathological motion data stream and the personalized hemiplegic steady-state model to extract the actual residual feature vector; performs subtraction operation between the theoretical fall precursor data waveform and the personalized hemiplegic steady-state model to extract the theoretical residual feature vector; Feature morphological similarity judgment unit: Constructs a phase space, calculates the morphological similarity value between the actual residual feature vector and the theoretical residual feature vector; compares the morphological similarity value with a preset similarity threshold to generate a risk judgment result; configured to: if the morphological similarity value is higher than the preset similarity threshold, trigger a graded intervention instruction; If the morphological similarity value is lower than or equal to the preset similarity threshold, an environmental noise filtering operation is performed. The methods for transforming stroke pathology mechanisms into risk perturbation factors include: Read the digital pathological descriptions of muscle weakness, spasticity, and central instability from the pre-stored set of pathological mechanism parameters; The description of myasthenia gravis is transformed into a nonlinear decay coefficient, which is then injected into the ratio of the support phase time between the healthy and affected sides. The spasm description is converted into high-frequency impulse noise and superimposed on the angular velocity data stream; The instability of the center of gravity is described as a drift vector pointing outward, which is applied to the trajectory of the center of pressure on the plantar surface; The risk disturbance factor is generated by integrating the nonlinear attenuation coefficient, the high-frequency pulse noise, and the drift vector. The healthy side support time is set to , the affected side support time is set to , and the affected side support time after the injection of the disturbance is calculated as follows: ； in, This represents the magnitude of muscle strength decline. This represents the rate of fatigue accumulation. The index representing the cumulative number of steps taken in the current continuous movement; High-frequency impulse noise characterizing the pathological attributes of spasticity is superimposed onto the angular velocity data stream: ； in, This represents the total number of spasms that occurred within the simulation window. The Dirac function was used to determine the equivalent impulse intensity for spasm. Characterizing the moment of spasm The sudden pulse; Apply the outward drift vector, which characterizes the pathological properties of center of gravity instability, to the trajectory of the plantar pressure center: ； This outward drift vector is further concretized into a restricted linear drift model: ； in, The relative time variable after the simulation event is triggered. The drift rate constant is The maximum allowable drift amplitude, This is a unit direction vector pointing towards the affected side.

2. The remote monitoring and health management system for fall risk in stroke patients according to claim 1, characterized in that, The methods for constructing personalized hemiplegic steady-state models include: Historical inertial measurement data of patients during periods without falls were obtained as a training set; An autoregressive neural network was used to learn the pathological gait fluctuation patterns of patients and predict the theoretical movement trajectory at the next moment. Periodic swaying that conforms to the characteristics of hemiplegia is defined as fluctuation within the baseline, a personalized pathological reference baseline is established, and a personalized hemiplegic steady-state model is generated.

3. The remote monitoring and health management system for fall risk in stroke patients according to claim 2, characterized in that, The methods for generating theoretical fall precursor data waveforms include: Start the simulation synthesis module and load the personalized hemiplegic steady-state model as the background carrier; Based on the preset fault evolution logic, the risk disturbance factor is dynamically mapped to the corresponding dimension of the background carrier; Synthesize data waveforms containing characteristics of specific pathological fall patterns as theoretical fall precursor data waveforms.

4. The remote monitoring and health management system for fall risk in stroke patients according to claim 1, characterized in that, The methods for extracting the actual residual feature vector and the theoretical residual feature vector include: In the dual-path differential extraction unit, the real-time pathological motion data stream and the theoretical fall precursor data waveform are input synchronously; Using the personalized hemiplegic steady-state model as the common subtrahend, differential operations are performed respectively; The pathological fluctuation background within the baseline is removed from the difference results, while the abnormal mutation components are retained; The outputs are the real residual feature vector containing a mixture of true fall signals and environmental noise, and the theoretical residual feature vector containing pure pathological fall patterns, respectively.

5. The remote monitoring and health management system for fall risk in stroke patients according to claim 1, characterized in that, The method of comparing the morphological similarity value with a preset similarity threshold to generate a risk assessment result includes: A dynamic time warping algorithm is used to align the time axes of the actual residual feature vector and the theoretical residual feature vector in phase space; Calculate the morphological similarity between the two. If the morphological similarity value is higher than the preset similarity threshold, it is determined that the actual abnormal fluctuation conforms to the pathological fall pattern, and a true fall risk alarm signal is generated as the risk determination result. If the morphological similarity value is lower than or equal to the preset similarity threshold, the abnormal fluctuation is determined to be a non-pathological interference, and a rehabilitation action or life noise label is generated as the risk assessment result.

6. The remote monitoring and health management system for fall risk in stroke patients according to claim 5, characterized in that, The methods for triggering graded intervention commands or performing environmental noise filtering operations include: In response to the true fall risk alarm signal, the audible and visual alarm of the remote monitoring terminal is immediately activated, and an emergency data package containing the fall type is pushed to the cloud management platform; In response to the rehabilitation action or living noise tag, the alarm trigger is suppressed, and the corresponding real residual feature vector is marked as a safe sample; The safety sample is fed back to the pathological benchmark modeling unit to update the historical rehabilitation data and optimize the robustness of the personalized hemiplegic steady-state model.

7. The remote monitoring and health management system for fall risk in stroke patients according to claim 1, characterized in that, The system also includes: The multimodal data acquisition front end is equipped with an inertial measurement unit and a pressure sensor to acquire the patient's triaxial acceleration, angular velocity and plantar pressure center trajectory, and perform spatiotemporal alignment processing to generate the real-time pathological motion data stream.

Images