Medical equipment and rehabilitation software information system based on the Internet of Things

Abstract

The present invention discloses a medical device and rehabilitation software information system based on the Internet of Things, specifically relating to the field of communication transmission management technology. By constructing event-anchored linear alignment on the edge side, taking the real-time channel quality vector as input, generating a compression strategy through a Pareto multi-arm bandit machine, and using the dynamic compression offset index as a security gate, an adaptive compression decision that takes both distortion and delay into account is achieved, solving the technical problem that multimodal physiological signals are difficult to simultaneously ensure timing consistency and diagnostic integrity under conditions of clock drift and bandwidth limitation; by setting a cross-modal alignment quality monitoring module, it is used to identify and suppress cross-modal misalignment pseudo-correlation in real time, thereby ensuring that rehabilitation assessment and short-cycle diagnosis are based on a reliable timing correspondence, solving the misalignment problem between electrocardiogram signals and motion signals.

Description

Technical Field

[0001] The present invention relates to the technical field of communication transmission management, and more specifically, to a medical device and rehabilitation software information system based on the Internet of Things. Background Art

[0002] In IoT medical and rehabilitation scenarios, patients usually wear multiple types of sensors such as electrocardiogram, electrophysiology, and inertial sensors at the same time. The data streams of each channel collected by medical equipment transmit continuous physiological data to the edge gateway and cloud platform through the home-hospital hybrid network, so that doctors can remotely evaluate the quality of rehabilitation and adjust prescriptions in time. Due to the diverse data sources, different sampling frequencies and dynamic fluctuations in network bandwidth, the system must continuously complete time alignment, compressed transmission and diagnostic analysis between the end, edge and cloud. At the same time, physiological information is highly sensitive data. Any synchronization or compression process that exposes plaintext timestamps or diagnostic features may bring privacy leakage risks.

[0003] Existing solutions primarily rely on centralized clocks or public timestamps to synchronize devices, and use fixed compression ratios or single-metric thresholds to balance bandwidth and distortion. These solutions suffer from the following drawbacks:

[0004] First, the plaintext synchronization mode cannot prevent the reconstruction of personal physiological rhythms, and also lacks a computationally robust offset correction mechanism in the case of distortion or arrhythmia;

[0005] Second, static or single-indicator error compression strategies cannot simultaneously consider the signal diagnostic value and real-time link capacity, often resulting in excessive distortion of the diagnostic window or bandwidth waste.

[0006] The above defects directly restrict the comprehensive performance improvement of the rehabilitation Internet of Things system in terms of privacy compliance, network adaptability and real-time diagnostic accuracy. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art, the present invention provides a medical device and rehabilitation software information system based on the Internet of Things, which generates multi-scale rotation-invariant spectral fingerprints on the terminal device and uses zero-knowledge proof to complete privacy-verifiable decentralized synchronization, and adopts a Pareto multi-armed bandit adaptive selection compression strategy based on the real-time channel quality vector to solve the problems raised in the above-mentioned background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solutions: a medical device and rehabilitation software information system based on the Internet of Things, comprising:

[0009] The data alignment module collects raw data containing the user's multimodal physiological and motion signals through medical equipment, collects timestamps of each channel data stream, and uses physiological events as time anchors to achieve time calibration through a linear regression model;

[0010] The data feature extraction module divides the time-calibrated data stream of each channel into time windows based on physiological events as time anchor points, and calculates the channel quality vector including signal entropy, spectral sparsity, error sensitivity, maximum tolerable delay, and link available bandwidth;

[0011] The compression strategy selection module uses the channel quality vector as input to construct a compression strategy space consisting of lossless, lossy, rarefaction, and variable sampling rate algorithms, as well as multiple compression ratios. It uses a Pareto multi-armed bandit algorithm with distortion and end-to-end delay as two-dimensional payoffs to select compression strategies online. Before selecting a compression strategy, it calculates the dynamic compression offset index. If the compression offset index exceeds a safety threshold and the proposed algorithm is a high-distortion solution, it automatically falls back to a low-distortion strategy.

[0012] The compression strategy execution and feedback module executes the selected compression strategy at the edge, writes the workload label, compression strategy identifier, and distortion estimate into the packet header, and then encrypts and transmits it to the cloud. The cloud decompresses and restores the actual distortion and total latency and provides feedback, and regularly updates the compression strategy.

[0013] Preferably, the data alignment module includes a basic alignment unit, which collects the timestamps of the data streams of each channel, calculates the time offset between the data streams of each channel through a linear regression model, generates a correction coefficient, and realizes global timing synchronization through time drift calibration.

[0014] Preferably, the data alignment module includes a time anchor sampling frequency control unit, which is used to adjust the time anchor sampling frequency to balance synchronization accuracy and power consumption, including:

[0015] Physiological events are used as time anchor points for ECG signals within the time window. Time drift calibration is performed through linear regression. The drift stability coefficient is constructed by the ratio of the drift slope to the regression residual. The width of the next round of synchronization window is dynamically adjusted based on the drift stability coefficient.

[0016] When the drift stability coefficient exceeds the upper threshold, it indicates that the current clock drift rate is greater than the allowable range and the fitting residual is within the preset range. The clock is rapidly deviating from the reference, so the time anchor sampling frequency should be increased. When the drift stability coefficient is below the lower threshold, it indicates that the clock drift rate is small and the noise ratio is outside the preset range. The clock is in an acceptably stable state, so the time anchor sampling frequency should be reduced.

[0017] Preferably, the distribution of the drift stability coefficient is statistically analyzed in the latest N synchronization cycles, and the 95th percentile corresponding to the target synchronization error upper limit is taken as the upper limit of the threshold; the minimum sampling frequency is solved according to the device power consumption budget on the same data set, and the maximum stability coefficient S_max acceptable under the minimum sampling frequency condition is calculated in reverse, and the corresponding S_max is taken as the lower limit of the threshold.

[0018] Preferably, the compression offset index is obtained as follows:

[0019] Error drift history modeling records the deviation between the predicted error and the actual error of each data channel before and after compression to form an error drift sequence. This sequence is then used to construct the compression error offset entropy to characterize the stability of the error offset distribution.

[0020] Compression drift response sensitivity modeling combines the error drift growth rate (the relative change between the current and previous error offsets) with the channel error sensitivity index through a nonlinear combination to construct a function that determines the sensitivity of the compression strategy to the error change trend.

[0021] The dynamic compression offset index is obtained by combining the compression error offset entropy and the sensitivity response function;

[0022] Based on the dynamic compression offset index to assist decision-making, when the dynamic compression offset index exceeds the preset threshold, the compression strategy switching or compression ratio adjustment mechanism is triggered to balance compression error and transmission delay, improving the robustness and adaptability of the overall data management of the system;

[0023] When the dynamic compression deviation index exceeds the preset threshold, the compression ratio is automatically reduced or switched to lossless compression. When the dynamic compression deviation index does not exceed the preset threshold, the original compression ratio is restored to suppress the rapid growth of compression error while ensuring the transmission delay constraint.

[0024] Preferably, the data alignment module also includes a privacy-preserving alignment unit, which is deployed at the edge and is used to perform short-time Fourier transform and continuous wavelet transform on the collected multimodal physiological signals to extract multi-scale time-frequency features, and use a preset sparse Gaussian random projection matrix to map the features into a fingerprint vector with a length of one hundred and twenty-eight; the privacy-preserving alignment unit also uses a zero-knowledge time offset proof based on Pedersen commitment and bullet proof to complete decentralized synchronization without exposing the plaintext timestamp, thereby realizing time drift calibration between devices.

[0025] Preferably, the system further includes a cross-modal alignment quality monitoring module, which is used to evaluate the reliability of the timing correspondence between the ECG signal and the motion signal in real time during the data compression and decompression process and output gating instructions to the compression strategy and diagnosis process. The operation process of the cross-modal alignment quality monitoring module includes:

[0026] Step 1: Peak event extraction: Within the synchronization window normalized by the R-peak time anchor point, perform sliding window standard deviation peak detection on the ECG signal and motion signal respectively to obtain the ECG peak sequence and motion peak sequence;

[0027] Step 2: Candidate alignment generation: Based on the sampling sequence number and timestamp in the same window, the peak events in the two sequences are preliminarily matched according to the nearest neighbor principle. The search window width for candidate pairing of ECG signal and motion signal peak sequences is adaptively set based on the distortion estimate in the compressed packet header. The ECG signal and motion signal peak sequences are extracted and a peak event candidate alignment set is generated.

[0028] Step 3: Based on the candidate alignment set, a ranking-time difference consistency matrix is ​​constructed and expanded into a mixed tensor. Robust principal component decomposition is used to obtain low-rank components and sparse residual components, and the corresponding anomaly score is calculated. The anomaly score is defined as the absolute value or normalized value of the candidate peak pairing event in the sparse residual, reflecting the degree to which the pairing deviates from the normal alignment pattern.

[0029] Step 4: Map the sparse residual components into graph node weights and input them into the graph neural network to generate an alignment confidence heat map. Apply the online change point detection method on the time series to detect misalignment change points in real time and output the topological consistency.

[0030] Step 5: Fusion of topological consistency and anomaly score to obtain global alignment confidence, and adjustment of rehabilitation assessment based on the global alignment confidence.

[0031] Preferably, the abnormality score is obtained in the following manner:

[0032] Extract all paired time differences from the candidate alignment set and sort them in chronological order or order of size to form a time difference sequence; use the time difference sequence to construct a time difference consistency matrix;

[0033] The time difference consistency matrix is ​​expanded into a hybrid tensor to fuse multi-dimensional information and improve robustness to misalignment patterns. After the tensor is expanded, Wasserstein distance regularization is added to strengthen the constraints on the offset distribution. Top-k sparse gating is used to retain high-risk peak pairings, reduce false detections, and highlight true anomalies.

[0034] Robust principal component decomposition is applied to the constructed mixed tensor to separate the normal alignment pattern and the abnormal misalignment components. The decomposition results in a low-rank matrix and a sparse matrix. The low-rank matrix represents the robust alignment pattern in which the time difference between the ECG peak and the action peak is generally consistent, which is recorded as the low-rank component. The sparse matrix represents the sparse residual component, which captures the abnormal pairing that deviates from the main pattern.

[0035] The non-zero elements in the sparse matrix are mapped back to the original candidate pairing sequence according to the position of the corresponding peak event. The abnormality score is calculated for each candidate peak pairing event based on the absolute value or normalized value of the candidate peak pairing event in the sparse residual, reflecting the degree to which the pairing deviates from the normal alignment pattern.

[0036] Preferably, the cross-modal alignment quality monitoring module feeds back the global alignment confidence to the compression strategy selection module to adaptively adjust the compression strategy; the adjustment measures include: lowering the compression ratio or switching to a lossless compression strategy to reduce distortion, increasing the transmission frequency of key frames or synchronization signals to correct cross-modal time synchronization, or adjusting the sensor data acquisition rate and transmission bandwidth to improve data alignment accuracy; after the adjustment is completed and the global confidence of the data alignment is restored to above the threshold, the rehabilitation assessment process and the blocked diagnosis module can be restarted to continue normal operation.

[0037] Preferably, the data alignment module includes a multi-anchor alignment unit, which is used for data alignment compensation in arrhythmia scenarios. When arrhythmia is detected, the P peak is used if the P wave is detectable, otherwise the R peak, P-peak and QRS complex form are used as multiple time anchors, and Bayesian posterior fusion is used to determine the global synchronization reference. The self-attention codec is used to perform weight distribution on the multiple time anchors, and the uncertainty of the entropy-temperature adjustment is output as the gating threshold; and the compression ratio is gated to ensure data integrity during the abnormal period.

[0038] The technical effects and advantages of the present invention are as follows:

[0039] (1) The present invention constructs event-anchored linear alignment on the edge side, takes the real-time channel quality vector as input, generates a compression strategy through a Pareto multi-arm bandit machine, and uses a dynamic compression offset index as a security gate to achieve an adaptive compression decision that takes both distortion and delay into account. This solves the technical problem that it is difficult to simultaneously ensure timing consistency and diagnostic integrity of multimodal physiological signals under clock drift and bandwidth constraints.

[0040] (2) The present invention replaces the basic alignment unit with a privacy protection alignment unit, which is suitable for home rehabilitation or cross-institutional data sharing, and it is necessary to strictly avoid timestamp leakage scenarios; by setting a cross-modal alignment quality monitoring module, it is used to identify and suppress cross-modal misalignment pseudo-correlation in real time, thereby ensuring that rehabilitation assessment and short-cycle diagnosis are based on a reliable time series correspondence, and solving the misalignment problem between ECG signals and motion signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a structural block diagram of the data alignment and compression management of each channel of the present invention.

[0042] Figure 2 This is a block diagram of the information management structure of the medical equipment and rehabilitation software of the present invention.

[0043] Figure 3 This is a flow chart of the cross-modal misalignment artifact suppression of the present invention. DETAILED DESCRIPTION

[0044] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0045] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0046] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0047] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0048] Example 1, see Figure 1 The present invention provides a block diagram of the data alignment and compression management structure of each channel. Figure 1 The medical equipment and rehabilitation software information system based on the Internet of Things shown includes:

[0049] The data alignment module collects raw data containing the user's multimodal physiological and motion signals through medical devices, collects timestamps from each channel's data stream, and uses a linear regression model to achieve time alignment using physiological events as time anchors. The aligned data is then integrated onto the same timeline, fusing different modal data (such as electromyography, kinematics, and virtual scene data) as needed to generate a unified multidimensional data stream and construct the temporal context of the multimodal signals.

[0050] Furthermore, when any critical signal changes suddenly or enters a critical diagnostic window, an adaptive compression process is triggered. Certain physiological signal changes (such as sudden arrhythmias or abnormal rehabilitation movements) in physiological or motion signals are clinically valuable information nodes and are marked as critical signals. Critical signals refer to physiological signals that are strongly associated with health and require higher fidelity and lower latency. Dynamic compression control is triggered through signal mutation detection or diagnostic window identification (such as the analysis window for remote ECG monitoring), saving bandwidth overhead in normal conditions while ensuring high-quality communication at critical moments.

[0051] The data feature extraction module divides the time-calibrated data streams of each channel into time windows using physiological events as time anchor points. It calculates the channel quality vector including signal entropy, spectrum sparsity, error sensitivity, maximum tolerable delay, and link available bandwidth, and detects key signal mutations or diagnostic windows in real time.

[0052] Explanation: Signal entropy reflects information density, helping to determine lossy compression. Spectral sparsity is useful for selecting compression strategies (such as DCT and wavelet). Error sensitivity assesses the impact of reconstruction errors in diagnosis. Maximum tolerable delay is used to balance compression duration and upload time. These metrics together form a quality vector, enabling subsequent decisions to make intelligent compression strategies based on data content rather than fixed configurations.

[0053] The compression strategy selection module uses the channel quality vector as input to construct a compression strategy space consisting of lossless, lossy, rarefaction, and variable sampling rate algorithms, as well as multiple compression ratios. It uses a Pareto multi-armed bandit algorithm with distortion and end-to-end delay as two-dimensional payoffs to select compression strategies online. Before selecting a compression strategy, it calculates the dynamic compression offset index. If the compression offset index exceeds a safety threshold and the proposed algorithm is a high-distortion solution, it automatically falls back to a low-distortion strategy.

[0054] Explanation: In a preferred embodiment, the Pareto multi-arm bandit online compression strategy includes 12 arms, which are raw-1.0, zlib-0.3, zlib-0.5, zstd-0.4, zstd-0.6, zstd-0.8, lz4-0.3, lz4-0.5, lz4-0.7, brotli-0.4, brotli-0.6, brotli-0.8 according to the algorithm name-target maximum compression ratio format; the arm profit is calculated in each round. The method is as follows: first, divide the compression distortion (percentage) of the arm by the 5% upper limit and subtract the result from 1 to obtain the distortion benefit. Then, divide the end-to-end delay (seconds) by the 1-second upper limit and subtract the result from 1 to obtain the delay benefit. Then, weighted sum is calculated with a ratio of 0.6:0.4 to obtain the total benefit. The algorithm adopts the ε-greedy strategy, with ε initially 0.2 and decaying as ε(t)=0.2·e^(-0.001t). The mean benefit is updated using an exponential sliding average with a step size of 0.1. The variance of the probability of selecting each arm is less than 1×10⁻ in 500 consecutive rounds. 4 After 300,000 rounds of simulation verification, this parameter combination can achieve the optimal average profit under the constraints of 5% distortion and 1 second delay.

[0055] Background: During edge compression and asynchronous network transmission, ECG and motion peaks may be temporally misaligned due to reordering, packet loss, or compression thinning. Traditional methods that align only by timestamps can misidentify unrelated events as the same physiological event, leading to false positives in ECG-motion correlation analysis. This unit uses a four-stage process—error-driven search, consistency decomposition, graph topology heat, and confidence fusion—to identify and block cross-modal misalignment artifacts in real time during the compression process, ensuring that the diagnostic algorithm operates only on truly synchronized data windows.

[0056] The compression strategy execution and feedback module executes the selected compression strategy at the edge, writes the workload label, compression strategy identifier, and distortion estimate into the packet header, and then encrypts and transmits it to the cloud. The cloud decompresses and restores the actual distortion and total latency and provides feedback. The compression strategy is regularly updated through federated averaging and knowledge distillation, and a lightweight model is distributed to achieve a self-evolving closed loop of synchronization, compression, and diagnosis.

[0057] It should be further explained in the embodiment of the present invention that the data alignment module includes a basic alignment unit, which collects the timestamps of the data streams of each channel, calculates the time offset between the data streams of each channel through a linear regression model, generates a correction coefficient, and achieves global timing synchronization through time drift correction;

[0058] It is necessary to further explain in the embodiment of the present invention that the specific implementation of the time drift calibration in the basic alignment unit includes:

[0059] To overcome data alignment errors caused by local clock drift and asynchronous sampling in multiple channels and devices, and to ensure the timing consistency and high-precision fusion of distributed physiological signals, the data alignment module performs clock anchoring based on physiological signals. The operation process includes:

[0060] After receiving the raw data stream, the R-peak in the ECG signal is automatically identified through the R-peak detection algorithm. Each time an R-peak is identified, a corresponding global time reference is generated to achieve synchronous triggering based on physiological events.

[0061] After completing the data stream acquisition of each channel, the local clock drift of each data channel is estimated and the calibration coefficient is calculated using the generated R-peak global time base through a real-time linear regression model, thereby completing the timing calibration of the original data stream and outputting the time-calibrated synchronized data;

[0062] Receive the calibrated data and establish a data state synchronization channel between the network edge and the cloud based on a lightweight communication protocol. The calibrated data is serialized and transmitted according to the R-peak time anchor sequence to achieve timing consistency alignment and global data fusion between multiple devices.

[0063] It should be further explained in the embodiment of the present invention that the data alignment module includes a time anchor sampling frequency control unit, which is used to adjust the time anchor sampling frequency to balance synchronization accuracy and power consumption, including:

[0064] Within the time window, physiological events (such as R peaks) are used as time anchor points for ECG signals. Time drift is calibrated through linear regression. The drift stability coefficient is constructed from the ratio of the drift slope to the regression residual. The width of the next synchronization window is dynamically adjusted based on the drift stability coefficient.

[0065] When the drift stability coefficient exceeds the upper threshold, it indicates that the current clock drift rate is greater than the allowable range and the fitting residual is within the preset range. The clock is rapidly deviating from the reference, so the time anchor sampling frequency should be increased. When the drift stability coefficient is below the lower threshold, it indicates that the clock drift rate is small and the noise ratio is outside the preset range. The clock is in an acceptably stable state, so the time anchor sampling frequency should be reduced.

[0066] In one possible embodiment, the distribution of the drift stability coefficient is statistically analyzed within the most recent N synchronization cycles, and the 95th percentile corresponding to the target synchronization error upper limit is taken as the upper threshold limit; the minimum sampling frequency is solved according to the device power consumption budget on the same data set, and the maximum stability coefficient S_max acceptable under the minimum sampling frequency condition is calculated in reverse, and the corresponding S_max is taken as the lower threshold limit.

[0067] It needs to be further explained in the embodiment of the present invention that the compression offset index is obtained as follows:

[0068] Error drift history modeling records the deviation between the predicted error and the actual error of each data channel before and after compression to form an error drift sequence. This sequence is then used to construct the compression error offset entropy to characterize the stability of the error offset distribution.

[0069] Compression drift response sensitivity modeling combines the error drift growth rate (the relative change between the current and previous error offsets) with the channel error sensitivity index through a nonlinear combination to construct a function that determines the sensitivity of the compression strategy to the error change trend.

[0070] The dynamic compression offset index is obtained by combining the compression error offset entropy and the sensitivity response function;

[0071] Based on the dynamic compression offset index to assist decision-making, when the dynamic compression offset index exceeds the preset threshold, the compression strategy switching or compression ratio adjustment mechanism is triggered to balance compression error and transmission delay, improving the robustness and adaptability of the overall data management of the system;

[0072] When the dynamic compression deviation index exceeds the preset threshold, the compression ratio is automatically reduced or switched to lossless compression. When the dynamic compression deviation index does not exceed the preset threshold, the original compression ratio is restored to suppress the rapid growth of compression error while ensuring the transmission delay constraint.

[0073] Furthermore, the calculation formula of the compression error offset entropy is:

[0074]

[0075] Among them, i represents the error sequence index, n represents the total amount of error, Offset ratio after off-table normalization, To avoid division by zero, the value is set to 0.01.

[0076] Furthermore, the calculation formula of the sensitive response function is:

[0077]

[0078] in, They are used to adjust the nonlinear amplification of the three items respectively; represents the error drift growth rate, Indicates error sensitivity, which indicates the immediate impact of a single-bit error in the channel on the key diagnostic feature value. In the offline stage of the model, perturbations can be injected into the channel bit by bit, the diagnostic feature gradient can be measured, and the absolute value of the gradient can be normalized and stored in a lookup table structure for online reading. The currently measured one-way queuing delay (in milliseconds), represents the delay reference threshold; the three items use inverse hyperbolic sine, hyperbolic tangent, and sigmoid mapping respectively to amplify the difference between the middle and high ranges and suppress low-end noise; multiplying them together gives the comprehensive sensitivity.

[0079] In a preferred embodiment, the dynamic compression offset index DCI is expressed as

[0080]

[0081] in, are the weight coefficients of each item, ; It is the single cycle drift rate; when DCI>0.65, the high compression ratio arm is selected; otherwise, the low compression ratio arm is used.

[0082] In summary, the embodiment of the present invention uses a basic alignment unit to complete clock drift calibration, cooperates with the time anchor sampling frequency control unit to optimize energy consumption; adopts a dynamic compression offset exponential closed-loop adjustment of the compression part; reflects the basic functions of the system and ensures performance-power consumption balance.

[0083] Example 2, see Figure 2 The structural diagram of the medical equipment and rehabilitation software information management is different from that of Example 1 in that:

[0084] The data alignment module includes a privacy protection alignment unit, and the basic alignment unit and the privacy protection alignment unit belong to parallel entries of the data alignment module;

[0085] What needs to be further explained in the embodiment of the present invention is that the data alignment module also includes a privacy-preserving alignment unit, which is deployed at the edge and is used to perform short-time Fourier transform and continuous wavelet transform on the collected multimodal physiological signals to extract multi-scale time-frequency features, and use a preset sparse Gaussian random projection matrix to map the features into a fingerprint vector with a length of one hundred and twenty-eight; the privacy-preserving alignment unit also uses zero-knowledge time offset proof based on Pedersen commitment and bullet proof to complete decentralized synchronization without exposing the plaintext timestamp, thereby realizing time drift calibration between devices.

[0086] In a preferred embodiment, the edge node first performs a 256-point short-time Fourier transform and a continuous wavelet transform on the ECG (250Hz) and inertial measurement signals (100Hz), respectively, to obtain a 128-dimensional time-frequency signature. A fingerprint vector is then generated using a Gaussian random projection matrix with 128 rows and 1,000 columns, N(0,1) distribution elements, and L2-normalized column vectors. Each device generates a ciphertext with a time offset using a Pedersen commitment and performs zero-knowledge verification with its peer using bullet proofs. In a test involving ten edge nodes and 10,000 interactions, the average verification latency was 3.1ms, with a standard deviation of 0.4ms, and the root mean square error of inter-device time drift was 4.7ms, meeting the medical synchronization accuracy requirements.

[0087] It needs to be further explained in the embodiment of the present invention that the decentralized synchronization includes:

[0088] Matching spectrum fingerprints generated by different devices to obtain a time offset, constructing an encrypted commitment for the time offset, and using a zero-knowledge proof to prove that the time offset is within a preset tolerance interval;

[0089] The hash value of the zero-knowledge proof and the device ID are written into the decentralized consensus drift table. Each device reads and verifies the proof before adjusting its local clock to maintain time consistency.

[0090] When the offset statistic in the drift table reaches the first global threshold, the data compression strength is reduced or switched to lossless compression. When the offset statistic drops below the second global threshold, the original compression ratio is restored, and the expired drift record is cleared after synchronization is completed. The first global threshold is obtained by statistically analyzing the distribution of the drift variance σ² during the last N synchronization processes, and selecting the variance value corresponding to the 90th percentile higher than σ² as the first global threshold to trigger compression degradation. The second global threshold is obtained by calculating the minimum acceptable compression ratio β_min based on the device power consumption budget, and calculating the upper limit σ²_min of the drift variance when maintaining β_min from the historical data, and setting σ²_min as the second global threshold to restore the original compression ratio. The purpose of the privacy protection alignment unit is to provide a unified and reliable timing reference for subsequent channel quality assessment and adaptive compression through privacy-safe time base alignment.

[0091] Explanation: Spectral fingerprint refers to extracting the energy spectrum vector at each scale after multi-scale time-frequency transformation of multimodal physiological signals. Then, through fixed random projection mapping, a fixed-length numerical sequence that remains robust to rotation and time shift is obtained to uniquely characterize the time-frequency characteristics of the signal segment.

[0092] Decentralized synchronization means that medical devices do not need to rely on a central clock or public timestamp, but instead achieve local clock consistency among all devices in the entire network by exchanging encrypted fingerprints and verifying time offset proofs. Zero-knowledge proof is a cryptographic protocol that allows the prover to prove to the verifier that the parameter satisfies a certain property (such as falling within a given interval) without revealing the specific value of the parameter to be proved. Any third party can only be sure that the property holds and cannot deduce the original parameter.

[0093] The difference between this embodiment of the present invention and embodiment 1 is that the system further includes a cross-modal alignment quality monitoring module, which is used to evaluate the reliability of the timing correspondence between the ECG signal and the motion signal in real time during the data compression and decompression process and output gating instructions to the compression strategy and diagnosis process, and includes a peak event candidate generation unit, a consistency robust decomposition unit, a topology confidence scoring unit, and a confidence fusion control unit;

[0094] The peak event candidate generation unit is used to extract ECG peak trains and motion peak trains, and generate a cross-modal peak pairing set based on temporal proximity;

[0095] The consistency robust decomposition unit is used to construct the time difference tensor for the peak pair set, robustly decompose it to obtain low-rank components and sparse residual components, and calculate the corresponding anomaly score;

[0096] A topology confidence scoring unit is used to map the sparse residual components into graph node weights and input them into a graph neural network to generate an alignment confidence heat map. It also applies an online change point detection method to the time series to detect misaligned change points in real time and output the topology consistency.

[0097] The confidence fusion control unit is used to fuse the topological consistency and the anomaly score to obtain the global alignment confidence, and adjust the rehabilitation assessment based on the global alignment confidence.

[0098] Mutual Relationship: The candidate generation unit provides a pairing set to the robust decomposition unit; the robust decomposition unit provides sparse residual components to the topology scoring unit; the topology scoring unit and the robust residual are jointly input into the confidence fusion control unit; the confidence fusion control unit feeds the results back to the compression strategy and diagnosis process.

[0099] It is necessary to further explain in the embodiments of the present invention that Figure 3 The cross-modal misalignment artifact suppression flow chart shows that the operation process of the cross-modal alignment quality monitoring module includes:

[0100] Step 1: Peak event extraction: Within the synchronization window normalized by the R-peak time anchor point, perform sliding window standard deviation peak detection on the ECG signal and motion signal respectively to obtain the ECG peak sequence and motion peak sequence;

[0101] The specific implementation method includes: performing peak detection on the ECG signal and the motion signal respectively; applying pre-processing filtering to the ECG signal (such as band-pass filtering to remove baseline drift and high-frequency noise), and detecting the ECG R wave peak using threshold judgment or differential threshold method; performing smoothing filtering on the motion signal (such as accelerometer or gyroscope data) and detecting significant motion peaks (such as extreme points of the motion cycle), and storing the detected ECG peak sequence and motion peak sequence in ascending order by timestamp, with each peak containing attributes such as timestamp and peak amplitude;

[0102] Step 2: Candidate alignment generation: Based on the sampling sequence number and timestamp in the same window, the peak events in the two sequences are preliminarily matched according to the nearest neighbor principle. The search window width for candidate pairing of ECG signal and motion signal peak sequences is adaptively set based on the distortion estimate in the compressed packet header. The ECG signal and motion signal peak sequences are extracted and a peak event candidate alignment set is generated.

[0103] The specific implementation includes: reading the compressed packet header metadata provided by the data compression module, obtaining a distortion estimate (e.g., a predicted value of a quantization error or a reconstruction error) under the current compression strategy, and using the distortion estimate as a parameter to adjust the search window width for cross-modal peak pairing; when the distortion estimate is large, increasing the pairing search window width; and when the distortion estimate is small, reducing the search window width;

[0104] Perform cross-modal peak event matching within the set time search window: traverse each ECG peak in the ECG peak sequence, and according to its timestamp, search for all action peaks within the range of ±W of the timestamp in the action peak sequence as candidate matches; each pair of ECG peaks and action peaks that meet the time proximity condition constitutes a peak event candidate pairing. All such pairing records are added to the peak event candidate alignment set data structure; each entry in the candidate alignment set contains: ECG peak identifier and its timestamp t_e, action peak identifier and its timestamp t_m, the time difference between the two, and possible peak amplitude information and other metadata; if an ECG peak corresponds to multiple action peaks in the window or vice versa, multiple candidate pairing records will be generated. This process ensures that all possible cross-modal peak pairs that may correspond in time sequence can still be collected in the presence of signal compression distortion, providing a complete set of candidate events for subsequent consistency analysis; the input of this step includes the decompressed ECG and action signal data sequence and the distortion parameters of the compressed packet header, and the output is the ECG peak sequence, action peak sequence and peak event candidate alignment set obtained by adaptive window matching;

[0105] Step 3: Based on the candidate alignment set, a ranking-time difference consistency matrix is ​​constructed and expanded into a mixed tensor. Robust principal component decomposition is used to obtain low-rank components and sparse residual components, and the corresponding anomaly scores are calculated.

[0106] The specific implementation method includes: extracting all paired time difference values ​​from the candidate alignment set, and sorting them in time order or size order to form a time difference sequence; constructing a time difference consistency matrix using the time difference sequence;

[0107] The time difference consistency matrix is ​​expanded into a hybrid tensor to integrate multi-dimensional information and improve robustness to misaligned patterns. The hybrid tensor is constructed by adding new dimensions to the time difference consistency matrix and integrating the corresponding relevant features into the tensor. For example, the time index dimension of the peak event or the modal category dimension is introduced to form a three-dimensional tensor. Through tensor expansion, the data representation simultaneously covers the consistency of the time difference sequence (reflected on the main plane of the matrix) and the cross-modal peak event characteristics (reflected in the additional dimension), thereby forming a hybrid tensor that comprehensively reflects the peak alignment pattern.

[0108] Furthermore, after tensor expansion, Wasserstein distance regularization is added to strengthen the constraints on the offset distribution; Top-k sparse gating is used to retain high-risk peak pairings, reduce false detections, and highlight true anomalies;

[0109] Explanation: After the time difference consistency matrix is ​​expanded and flattened into a tensor, a regularization term based on the Wasserstein distance is introduced; the regularization term calculates the optimal transmission distance between the current residual distribution and the zero-centered reference distribution, and is weighted by the coefficient λ_W in the objective function of the robust principal component decomposition. By minimizing the transmission distance, pseudo outliers introduced by compression and rarefaction or link jitter can be suppressed, ensuring that the truly abnormal residuals play a dominant role in the decomposition process, thereby improving the detection accuracy of abnormal pairing; the sparse residual matrix obtained by decomposition is sorted in descending order of absolute value, retaining only the top k residual elements and setting the remaining elements to zero; this gating mechanism removes low-amplitude noise points while retaining high-risk peak pairings, which not only reduces false detections but also significantly compresses the input scale and computational overhead of subsequent graph neural networks, making real-time monitoring more efficient and robust;

[0110] Robust principal component decomposition is applied to the constructed mixed tensor to separate normal alignment patterns and abnormal misalignment components. Robust principal component decomposition can be performed by appropriately flattening the tensor or converting it into a matrix form piece by piece (for example, expanding the tensor into a matrix in the fragment dimension), and solving it using the principal component chasing algorithm or the alternating Lagrange multiplier method. By minimizing the low-rank term weighted by the nuclear norm plus the L1 norm of the sparse term, a low-rank matrix and a sparse matrix are decomposed. The low-rank matrix represents a robust alignment pattern in which the time difference between the ECG peak and the action peak is generally consistent (for example, an approximately constant or slowly drifting time delay), which is recorded as a low-rank component, and the sparse matrix represents a sparse The sparse residual component captures abnormal pairings that deviate from the main pattern (i.e., possible misplaced peak correspondences or missing pairings); the non-zero elements in the sparse matrix are mapped back to the original candidate pairing sequence according to the position of the corresponding peak event, and an anomaly score is calculated for each candidate peak pairing event based on the absolute value or normalized value of the candidate peak pairing event in the sparse residual, reflecting the degree to which the pairing deviates from the normal alignment pattern; for a candidate peak pairing event, if the time difference is significantly abnormal compared with most events, a larger value will appear in the sparse residual component, giving it a high anomaly score; conversely, if it belongs to a robust alignment pattern, its residual is close to zero, and the anomaly score is low;

[0111] Step 4: Map the sparse residual components into graph node weights and input them into the graph neural network to generate an alignment confidence heat map. Apply the online change point detection method on the time series to detect misalignment change points in real time and output the topological consistency.

[0112] The specific implementation method is as follows: a peak event graph structure representing the cross-modal peak alignment relationship is constructed based on candidate pairing events; the implementation method is to regard each peak event candidate pairing as a node in the graph, and the node attributes include the timestamp of the event and its anomaly score (obtained by sparse residual assignment); in the peak event graph structure, in order to reflect the temporal and topological relationship between paired events, the nodes are connected into a chain or adjacency structure according to time sequence: for example, according to the time sequence, an edge is established between each two adjacent candidate pairing nodes, and the edge weight can be set to the absolute difference of their time difference change or simply set to 1 to represent a sequential connection; if there are different candidate pairs that share ECG peaks or shared action peaks (indicating potential conflicts or substitution relationships), connecting edges are introduced between these nodes to capture the competitive relationship between the pairs; thus, a peak event graph structure that integrates time sequence and candidate correspondence is obtained; the initial value of the node weight of the peak event graph structure is set to the node's anomaly score, indicating the degree to which each candidate pairing is considered abnormal;

[0113] The above peak event graph structure is input into the pre-trained graph neural network (GNN) model for topological consistency analysis;

[0114] The reliability of each candidate pairing is evaluated globally by integrating the characteristic information of each node and its neighbors through iterative message passing. During the calculation process, each node exchanges information with its connected neighboring nodes, allowing the anomaly score to propagate and smooth between temporally adjacent or topologically related peak events. For example, if a node (pairing event) has a high anomaly score, but its adjacent pairs are normal (low anomaly), the confidence of the node is reduced, because the isolated anomaly may be transient noise. Conversely, if multiple connected nodes show anomalies, the anomaly signal of the local structure is strengthened, and it is considered that a persistent dislocation has occurred.

[0115] After several layers of graph convolution / updates, the graph neural network outputs an alignment confidence value for each node. The alignment confidence ranges from 0 to 1 (or a relative score), indicating the degree of confidence that the corresponding peak pairing belongs to the correct alignment topology. A higher alignment confidence value indicates a more reliable local ECG-motion signal alignment relationship, while a lower value indicates potential misalignment.

[0116] Arranging the alignment confidence of all nodes in chronological order can form a one-dimensional heat map of alignment confidence that changes over time (i.e., a confidence time series, which is often visualized as a heat map with color changing over time).

[0117] In order to timely perceive the sudden change of alignment status, the online change point detection method is applied to the generated alignment confidence time series to find the misalignment change points in real time; online change point detection detects the moment when the statistical characteristics change significantly through dynamic monitoring; for example, sliding window statistics or cumulative sum control chart algorithms can be used: maintain the mean of the alignment confidence in a recent period of time, compare it with the mean of an earlier period, and when the difference between the two exceeds a set threshold, it is determined that a change point has occurred at that time point; or calculate the cumulative sum deviation of the confidence sequence, and trigger a change point signal when the deviation exceeds a certain limit; in this way, the moment when the alignment confidence changes significantly can be identified online, which corresponds to the time point when the ECG signal and the motion signal are misaligned;

[0118] Each time a misalignment change point event is detected, the time of occurrence and the magnitude of the change are recorded, and the topological consistency is updated accordingly. Topological consistency is defined as a measure of the stability of the cross-modal peak correspondence relationship within the entire current observation interval. For example, it can be characterized by the number of misalignment change points detected per unit time or the average interval: the fewer the change points and the more stable the alignment relationship, the higher the topological consistency; the more frequent the occurrence of change points, the lower the topological consistency. In one possible embodiment, the topological consistency is calculated by taking the completely consistent state without change points as 100% (or 1.0), and decaying based on the observed frequency of misalignment change point events.

[0119] Step 5: The topology consistency and anomaly score are combined to obtain a global alignment confidence, and the rehabilitation assessment is adjusted based on the global alignment confidence. For example, when the global alignment confidence falls below a first global threshold, the corresponding short-term diagnosis is blocked. When it falls below a second global threshold, the entire rehabilitation assessment is paused. The sparse residual component is then fed back to the data compression management module to adjust the subsequent compression strategy.

[0120] Explanation: The anomaly score and topological consistency reflect the degree of anomaly in local pairing events and the stability of the overall topological alignment, respectively. To comprehensively evaluate the reliability of cross-modal data alignment, the two are fused to calculate the global alignment confidence. The global alignment confidence is obtained through weighted summation or nonlinear fusion function. During the fusion process, low topological consistency (indicating multiple misalignments) or high anomaly score (significant anomalous pairings) will lead to a decrease in global alignment confidence.

[0121] Two confidence thresholds are pre-set and rolled out, namely the first global threshold and the second global threshold (the second global threshold is less than the first global threshold), which are used to trigger different levels of response measures. When the calculated global alignment confidence is lower than the first global threshold, it indicates that the reliability of cross-modal data alignment within the current short time range is poor, but not completely lost. At this time, the cross-modal alignment quality monitoring module sends a signal to the upper-level control logic to block the corresponding short-cycle diagnosis work. The so-called short-cycle diagnosis refers to a functional component that performs rapid analysis and diagnosis based on a short time window (such as a few heartbeats or a few movement cycles), such as modules such as instant arrhythmia detection or instant movement amplitude assessment. Blocking short-cycle diagnosis means temporarily stopping or ignoring the output of this module to prevent misjudgment due to data misalignment. Short-cycle diagnosis automatically resumes operation after the alignment confidence returns to normal.

[0122] Furthermore, when the global alignment confidence continues to decline and falls below the second global threshold, it indicates that the cross-modal data alignment relationship has seriously failed, affecting the reliability of the entire rehabilitation assessment process; the entire rehabilitation assessment is immediately suspended, that is, the data analysis and feedback of the entire current assessment session or stage are suspended; during the suspension period, the clinician or rehabilitation personnel are notified of the data abnormality and an inspection of the data path is initiated.

[0123] Furthermore, the cross-modal alignment quality monitoring module feeds back the global alignment confidence to the compression strategy selection module to adaptively adjust the compression strategy; the adjustment measures include: lowering the compression ratio or switching to a lossless compression strategy to reduce distortion, increasing the transmission frequency of key frames or synchronization signals to correct cross-modal time synchronization, or adjusting the sensor data acquisition rate and transmission bandwidth to improve data alignment accuracy; once the adjustment is completed and the global confidence of the data alignment returns to above the threshold, the rehabilitation assessment process and the blocked diagnostic module can be restarted to continue normal operation.

[0124] Summary: The embodiment of the present invention replaces the basic alignment unit with a privacy-preserving alignment unit, which is suitable for home rehabilitation or cross-institutional data sharing, and needs to strictly avoid timestamp leakage scenarios; in view of the time misalignment between ECG signals and motion signals during the compression-decompression and reconstruction process, which may cause the two-modal features to be incorrectly paired and trigger false positive diagnosis, a cross-modal alignment quality monitoring module is set to identify and suppress cross-modal misalignment pseudo-correlation in real time, thereby ensuring that rehabilitation assessment and short-term diagnosis are based on a reliable time series correspondence, and solving the misalignment problem between ECG signals and motion signals; the global alignment confidence is calculated based on the topological consistency and the anomaly score, and the rehabilitation assessment is adjusted based on the global alignment confidence; and the global alignment confidence is fed back to the compression strategy selection module to adaptively adjust the compression strategy.

[0125] The difference between Example 3 and Examples 1 and 2 is that:

[0126] The data alignment module includes a multi-anchor alignment unit. When an arrhythmia is detected, the P peak is used if the P wave is detectable. Otherwise, the R peak, P-peak, and QRS complex morphology are used as multiple time anchors. Bayesian posterior fusion is used to determine the global synchronization reference, and the compression ratio is gated to ensure data integrity during abnormal periods.

[0127] The multi-anchor point alignment unit is used for data alignment compensation in arrhythmia scenarios, and the specific implementation includes:

[0128] Low-confidence R-peak screening: The ECG signals of patients with arrhythmias are marked as ECG signals to be synchronized. The ECG signal window is obtained, and continuous wavelet-convolutional network joint detection is performed on each candidate R-peak. The morphological confidence C_R is calculated. When the morphological confidence C_R is lower than the preset requirement, the window is marked as a low-confidence event and the corresponding original waveform segment is saved.

[0129] For example, compare C_R with the first threshold θ1; when the morphological confidence C_R < the first threshold θ1, mark the window as a low-confidence event and save the corresponding original waveform segment; the first threshold θ1 is the dynamic lower limit determined by taking the 5th percentile of the confidence distribution in the most recent N time windows;

[0130] Candidate time anchor point construction: For the low-confidence events, fast Fourier transform and adaptive wavelet decomposition are used to extract the P-peak, QRS onset, and T-peak, respectively, to generate a candidate anchor set A = {a_i}, where each time anchor point a_i includes a timestamp t_i, a peak type identifier s_i, and a morphology vector f_i (composed of amplitude, width, and slope). Set A is passed along with the confidence level C_R to the confidence fusion step.

[0131] Confidence fusion: Using the pre-trained ECG event transition probability dynamic Bayesian network, confidence fusion is performed on the candidate anchor set A to obtain the posterior probability of each time anchor point P_i = P(a_i|A, C_R); if the posterior probability is not lower than the preset requirement, the corresponding time anchor point is used as the global time reference;

[0132] In one possible embodiment, the second threshold θ2 is set to the 95th percentile of the posterior probability when the synchronization error is ≤ E_max in the training set; a single time anchor point or a combination of anchors weighted by P_i that satisfies P_i ≥ θ2 is selected as the current global time reference, and the time deviation Δt of the reference relative to the original R peak is output;

[0133] Drift calibration: The device's local clock drift slope is updated based on the time deviation Δt, and the lossy compression ratio threshold is dynamically lowered to the compression strategy module based on the posterior probability P_i. When the time deviation converges for k consecutive windows and the posterior probability exceeds the second threshold, the original threshold is automatically restored. At the same time, the posterior probability statistics are uploaded to the cloud for federated averaging to regularly fine-tune the continuous wavelet-convolutional network, thereby forming a closed-loop improved synchronization-compression collaborative control.

[0134] In one possible embodiment, a self-attention codec is used to assign weights to multiple time anchor points and output the uncertainty of entropy-temperature adjustment as a gating threshold.

[0135] Explanation: The Transformer architecture is used to perform self-attention encoding on multiple time anchor points. The decoding end outputs normalized weights for each event, which is used to dynamically evaluate the contribution of each type of anchor point to the synchronization accuracy. The information entropy of the output weight distribution is taken and scaled by temperature T to obtain numerical uncertainty. The higher the uncertainty, the lower the quality of the anchor point. The uncertainty is directly used as the Bayesian fusion gating threshold to automatically suppress the intervention of low-confidence anchor points. For example, the P peak, R peak, QRS start point, QRS end point and T peak detected in this ECG segment are arranged in chronological order to obtain a multi-event sequence, which is then sent to the Transformer structure. The encoder first calculates the correlation between each anchor point and other anchor points; the decoder then combines these correlations and outputs a set of normalized weights, the sum of which equals one, indicating the importance of each anchor point to the subsequent synchronization accuracy; to judge the overall anchor point quality, the "information entropy" of this set of weights is first calculated, that is, observing whether the weights are concentrated in a few anchor points or dispersed across all anchor points; then an adjustable temperature coefficient is used to amplify or reduce the entropy value to obtain the final uncertainty value. A high uncertainty indicates that the weight distribution is dispersed and there is no obvious reliable main anchor; a low uncertainty indicates that the weights are concentrated and some anchor points are highly credible.

[0136] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. The medical equipment and rehabilitation software information system based on the Internet of Things is characterized by: include: The data alignment module collects raw data containing the user's multimodal physiological and motion signals through medical equipment, collects timestamps of each channel data stream, and uses physiological events as time anchors to achieve time calibration through a linear regression model; The data feature extraction module divides the time window of each channel data stream after time calibration with physiological events as the time anchor point and calculates the channel quality vector; The compression strategy selection module uses the channel quality vector as input to construct a compression strategy space consisting of lossless, lossy, rarefaction, and variable sampling rate algorithms, as well as multiple compression ratios. It uses a Pareto multi-armed bandit algorithm with distortion and end-to-end delay as two-dimensional benefits to select compression strategies online. Before selecting a compression strategy, it calculates the dynamic compression offset index. If the compression offset index exceeds the safety threshold and the proposed algorithm is a high-distortion solution, it automatically falls back to the low-distortion strategy. The compression offset index is obtained as follows: Error drift history modeling records the deviation between the predicted error and the actual error of each data channel before and after compression to form an error drift sequence. This sequence is then used to construct the compression error offset entropy to characterize the stability of the error offset distribution. Compression drift response sensitivity modeling combines the error drift growth rate with the channel error sensitivity index through nonlinear combination to construct a function that determines the sensitive response of the compression strategy to the error change trend. The dynamic compression offset index is obtained by combining the compression error offset entropy and the sensitivity response function; Based on the dynamic compression deviation index to assist decision-making, when the dynamic compression deviation index exceeds the preset threshold, the compression strategy switching or compression ratio adjustment mechanism is triggered to balance the compression error and transmission delay; When the dynamic compression deviation index exceeds the preset threshold, the compression ratio is automatically reduced or switched to lossless compression. When the dynamic compression deviation index does not exceed the preset threshold, the original compression ratio is restored to suppress the rapid growth of compression error while ensuring transmission delay constraints; The compression strategy execution and feedback module executes the selected compression strategy at the edge, writes the workload label, compression strategy identifier, and distortion estimate into the packet header, and then encrypts and transmits it to the cloud. The cloud decompresses and restores the actual distortion and total latency and provides feedback, and regularly updates the compression strategy.

2. The medical equipment and rehabilitation software information system based on the Internet of Things according to claim 1, characterized in that: The data alignment module includes a basic alignment unit, which collects the timestamps of the data streams of each channel, calculates the time offset between the data streams of each channel through a linear regression model, generates a correction coefficient, and realizes global timing synchronization through time drift calibration.

3. The medical equipment and rehabilitation software information system based on the Internet of Things according to claim 2, characterized in that: The data alignment module includes a time anchor sampling frequency control unit, which is used to adjust the time anchor sampling frequency to balance synchronization accuracy and power consumption, including: Physiological events are used as time anchor points for ECG signals within the time window. Time drift calibration is performed through linear regression. The drift stability coefficient is constructed by the ratio of the drift slope to the regression residual. The width of the next round of synchronization window is dynamically adjusted based on the drift stability coefficient. When the drift stability coefficient exceeds the upper threshold, it indicates that the current clock drift rate is greater than the allowable range and the fitting residual is within the preset range. The clock is rapidly deviating from the reference, so the time anchor sampling frequency should be increased. When the drift stability coefficient is below the lower threshold, it indicates that the clock is in an acceptably stable state, so the time anchor sampling frequency should be reduced.

4. The medical equipment and rehabilitation software information system based on the Internet of Things according to claim 3, characterized in that: The distribution of the drift stability coefficient in the last N synchronization cycles is statistically analyzed, and the 95th percentile corresponding to meeting the target synchronization error upper limit is taken as the upper threshold limit; On the same data set, the minimum sampling frequency is solved according to the device power consumption budget, and the maximum acceptable stability coefficient S_max under the minimum sampling frequency condition is calculated in reverse, and the corresponding S_max is taken as the lower limit of the threshold.

5. The medical equipment and rehabilitation software information system based on the Internet of Things according to claim 2, characterized in that: The data alignment module also includes a privacy-preserving alignment unit, which is deployed at the edge and is used to perform short-time Fourier transform and continuous wavelet transform on the collected multimodal physiological signals to extract multi-scale time-frequency features, and use a preset sparse Gaussian random projection matrix to map the features into a fingerprint vector of length 128; the privacy-preserving alignment unit also uses zero-knowledge time offset proof based on Pedersen commitment and bullet proof to complete decentralized synchronization without exposing the plaintext timestamp, thereby realizing time drift calibration between devices.

6. The medical equipment and rehabilitation software information system based on the Internet of Things according to any one of claims 1 to 5, characterized in that: The system also includes a cross-modal alignment quality monitoring module, which is used to evaluate the reliability of the timing correspondence between the ECG signal and the motion signal in real time during the data compression and decompression process and output gating instructions to the compression strategy and diagnosis process. The operation process of the cross-modal alignment quality monitoring module includes: Step 1: Peak event extraction: Within the synchronization window normalized by the R-peak time anchor point, perform sliding window standard deviation peak detection on the ECG signal and motion signal respectively to obtain the ECG peak sequence and motion peak sequence; Step 2: Candidate alignment generation: Based on the sampling sequence number and timestamp in the same window, the peak events in the two sequences are preliminarily matched according to the nearest neighbor principle. The search window width for candidate pairing of ECG signal and motion signal peak sequences is adaptively set based on the distortion estimate in the compressed packet header. The ECG signal and motion signal peak sequences are extracted and a peak event candidate alignment set is generated. Step 3: Based on the candidate alignment set, a ranking-time difference consistency matrix is ​​constructed and expanded into a mixed tensor. Robust principal component decomposition is used to obtain low-rank components and sparse residual components, and the corresponding anomaly score is calculated. The anomaly score is defined as the absolute value or normalized value of the peak event candidate pair in the sparse residual, reflecting the degree to which the pairing deviates from the normal alignment pattern. Step 4: Map the sparse residual components into graph node weights and input them into the graph neural network to generate an alignment confidence heat map. Apply the online change point detection method on the time series to detect misalignment change points in real time and output the topological consistency. Step 5: Fusion of topological consistency and anomaly score to obtain global alignment confidence, and adjustment of rehabilitation assessment based on the global alignment confidence.

7. The medical equipment and rehabilitation software information system based on the Internet of Things according to claim 6, characterized in that: The anomaly score is obtained as follows: Extract all paired time differences from the candidate alignment set and sort them in chronological order or order of size to form a time difference sequence; use the time difference sequence to construct a time difference consistency matrix; The time difference consistency matrix is ​​expanded into a hybrid tensor to fuse multi-dimensional information and improve robustness to misalignment patterns. After the tensor is expanded, Wasserstein distance regularization is added to strengthen the constraints on the offset distribution. Top-k sparse gating is used to retain high-risk peak pairings, reduce false detections, and highlight true anomalies. Robust principal component decomposition is applied to the constructed mixed tensor to separate the normal alignment pattern and the abnormal misalignment components. The decomposition results in a low-rank matrix and a sparse matrix. The low-rank matrix represents the robust alignment pattern in which the time difference between the ECG peak and the action peak is generally consistent, which is recorded as the low-rank component. The sparse matrix represents the sparse residual component, which captures the abnormal pairing that deviates from the main pattern. The non-zero elements in the sparse matrix are mapped back to the original candidate pairing sequence according to the position of the corresponding peak event. The abnormality score is calculated for each candidate peak pairing event based on the absolute value or normalized value of the candidate peak pairing event in the sparse residual, reflecting the degree to which the pairing deviates from the normal alignment pattern.

8. The medical equipment and rehabilitation software information system based on the Internet of Things according to claim 7, characterized in that: The cross-modal alignment quality monitoring module feeds back the global alignment confidence to the compression strategy selection module to adaptively adjust the compression strategy; Adjustment measures include: reducing the compression ratio or switching to a lossless compression strategy to reduce distortion, increasing the transmission frequency of key frames or synchronization signals to correct cross-modal time synchronization, or adjusting the sensor data acquisition rate and transmission bandwidth to improve data alignment accuracy; Once the adjustment is completed and the global confidence of the data alignment is restored to above the threshold, the rehabilitation assessment process and the blocked diagnostic module can be restarted and continue normal operation.

9. The medical equipment and rehabilitation software information system based on the Internet of Things according to claim 5, characterized in that: The data alignment module includes a multi-anchor alignment unit, which is used for data alignment compensation in arrhythmia scenarios. When arrhythmia is detected, the P peak is used if the P wave is detectable. Otherwise, the R peak, P-peak and QRS complex form are used as multiple time anchors. Bayesian posterior fusion is used to determine the global synchronization reference. The self-attention codec is used to weight the multiple time anchors and output the uncertainty of entropy-temperature adjustment as the gating threshold. The compression ratio is also gated to ensure data integrity during abnormal periods.

Images