A multi-modal dynamic physiological monitoring method based on event-driven and riemannian manifold

Abstract

The present application relates to multi-modal physiological state monitoring technology, and discloses a kind of multi-modal dynamic physiological monitoring method based on event-driven and riemannian manifold.The method collects multiple physiological signals and carries out pre-conditioning, quality assessment and differential modulation event processing, and is based on asynchronous event tuple incrementally updates covariance matrix and carries out positive definite constraint;Calculate the manifold distance of current covariance matrix and each physiological scene long-term base point, obtain scene posteriori confidence and reverse adjust event trigger threshold;According to the confidence, construct dynamic base point, carry out tangent space mapping to current covariance matrix, obtain offset amplitude and mode contribution;Based on the offset result, carry out abnormal offset classification, update long-term base point using normal stable sample, and write abnormal sample into isolated record set.The present application can reduce redundant sampling and scene switching false alarm, avoid abnormal sample pollution long-term baseline, improve the stability and interpretability of continuous monitoring.

Description

Technical Field

[0001] This invention relates to the field of multimodal physiological state monitoring technology, and discloses a multimodal dynamic physiological monitoring method based on event-driven and Riemannian manifold. Background Technology

[0002] With the development of wearable sensors, remote health management and continuous physiological state monitoring technologies, state recognition methods based on multi-channel physiological signals have been widely applied in scenarios such as outpatient follow-up, rehabilitation monitoring, sports health and long-term vital sign observation.

[0003] Patent CN110494078A discloses a technical solution for collecting and detecting multiple parameters using a wearable physiological monitor. This monitor can collect physiological parameters such as cough, wheezing, heart rate, skin temperature, activity, respiratory rate, skin impedance, electrocardiogram data, blood pressure, and skin conductance. It can also present the analysis results from data from one or more wearable devices to the user or healthcare provider after appropriate processing. The patent also discloses that the wearable device can be equipped with multiple sensors, and that the sensor data can be processed or transmitted through a preprocessor, buffer, and main processor structure, thereby achieving continuous monitoring of the human physiological state.

[0004] The aforementioned existing technologies demonstrate that multi-parameter wearable physiological monitoring has a certain technological foundation. However, they focus more on the collection, recording, display, transmission, and general analysis of various physiological parameters by wearable devices, and do not fully address the issue of reliable fusion between multimodal physiological signals in complex dynamic scenarios. Specifically, existing multimodal physiological monitoring methods typically still use fixed thresholds, simple weighted aggregation, or fixed frame rate sampling for state judgment. When the monitored object is in different physiological scenarios such as resting, active, sleeping, or rehabilitation exercises, the normal fluctuation range of each modality signal and the coupling relationship between channels are not the same. If a unified baseline or fixed threshold is still used for judgment, false alarms or missed alarms are easily generated in situations such as scene switching, body movement interference, and changes in sensor contact status.

[0005] Meanwhile, different physiological signals such as ECG, PPG, blood pressure estimation, respiratory impedance, and acceleration exhibit significant differences in sampling frequency, physical dimensions, noise sources, and signal quality. Existing methods that directly perform synchronous resampling or feature stitching on multiple signals introduce a large amount of redundant data, increasing power consumption and computational burden on the device. Furthermore, they struggle to preserve the coupling relationships between modalities such as ECG-pulse, respiration-blood pressure, and body movement-pulse. Particularly in low-power wearable devices, continuous sampling at a fixed frame rate can result in invalid sampling during deep sleep or stable states, while short-term abrupt events can cause response delays due to sampling intervals or window waiting.

[0006] Furthermore, existing dynamic baseline update methods typically lack a strict mechanism for isolating anomalous samples. When anomalous offset samples, poor sensor contact samples, or transitional samples from scene switching are directly included in the baseline update, the long-term baseline can easily become contaminated, thereby reducing the stability of subsequent state recognition. Existing technologies also lack a closed-loop mechanism that uses backend scene recognition results to adjust the frontend event triggering strategy, leading to a disconnect between frontend sampling density, backend scene judgment, and anomalous offset recognition.

[0007] Therefore, there is an urgent need for a new multimodal dynamic physiological monitoring method that can transform multiple asynchronous physiological signals into a joint state representation that can be incrementally updated under low power conditions, and maintain independent base points based on different physiological scenarios, forming a closed loop between scene recognition, event triggering, offset measurement and abnormal sample isolation, so as to improve the stability and reliability of continuous physiological state monitoring in complex dynamic scenarios. Summary of the Invention

[0008] This study addresses the problems in existing multimodal dynamic physiological monitoring methods, such as the inability of fixed thresholds to adapt to individual differences, lack of physiological scene differentiation, lack of reliable dynamic fusion between multiple physiological signals, susceptibility of dynamic baselines to contamination by abnormal samples, and lack of closed-loop feedback between scene recognition results and front-end sampling strategies.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a multimodal dynamic physiological monitoring method based on event-driven and Riemannian manifolds, comprising the following steps:

[0010] S1: Collect multiple physiological signals from the monitored object, pre-condition and quality assessment of each physiological signal, and obtain the conditioning signal and the corresponding quality score;

[0011] S2 performs differential modulation event processing on each conditioning signal and uniformly stamps the trigger events based on the shared clock. When the difference obtained from the differential modulation event processing reaches the event trigger threshold of the corresponding physiological mode, an asynchronous event tuple is generated.

[0012] S3, based on the asynchronous event tuple and quality score, perform event-driven incremental update and positive definite constraint on the covariance matrix of the multi-path physiological signals to obtain the current covariance matrix on the symmetric positive definite manifold;

[0013] S4, calculate the manifold distance between the current covariance matrix and the long-term reference points corresponding to different physiological scenarios, the long-term reference points Indicates the first The long-term normal reference center of the multimodal covariance matrix under each physiological scenario is determined, the current scenario and its posterior confidence are determined, and the event triggering threshold in step S2 is adjusted in reverse according to the posterior confidence.

[0014] S5. Construct a dynamic base point for manifold mapping based on the posterior confidence level, and perform tangent space mapping on the current covariance matrix with the dynamic base point as a reference to obtain the offset magnitude and modal contribution of each physiological mode.

[0015] S6. Based on the offset amplitude, modal contribution, posterior confidence and quality score, the current physiological state is classified as abnormal. Normal samples that meet the stability conditions are used to update the long-term baseline of the corresponding physiological scenario. Abnormal offset samples are written into the isolation record set and are prohibited from participating in the update of the long-term baseline.

[0016] The beneficial effects of this invention are as follows:

[0017] 1. This invention uses differential modulation event processing to output events only when the physiological signal changes effectively relative to the reference level, avoiding the generation of a large amount of redundant data in the steady state by traditional fixed frame rate sampling. It transforms the reference level and quality score corresponding to multiple physiological signals into a multimodal weighted covariance matrix, and processes this covariance matrix as a point on a symmetric positive definite manifold, thereby simultaneously expressing the fluctuation of each mode and the coupling relationship between different modes.

[0018] 2. This invention maintains long-term baselines for different physiological scenarios and constructs dynamic baselines when switching scenarios, so that the offset measurement benchmark of the current state changes smoothly with the posterior confidence of the scenario, avoiding false offsets caused by abrupt changes in baselines. The event triggering threshold is modulated in reverse by the posterior confidence of the scenario, so that the front-end event generation strategy can be dynamically adjusted according to the current scenario and its stability, thereby improving the monitoring stability in complex activity scenarios.

[0019] 3. This invention separates normal stable samples from abnormal offset samples by using a stability window criterion and an isolated record set. Normal stable samples are used to update long-term baselines, while abnormal offset samples do not participate in long-term baseline updates, thus structurally preventing abnormal samples from contaminating the long-term baseline.

[0020] 4. This invention enables the system to maintain continuous monitoring capability even when subjected to body motion interference, changes in sensor contact state, or a decline in the quality of some modes through quality scoring, temporary modal order reduction comparison, dynamic base points, and anomaly isolation mechanisms. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention;

[0022] Figure 2 This is a schematic diagram of the SPD Riemannian manifold characterization and dynamic basis point evolution of the present invention;

[0023] Figure 3 This invention relates to a self-evolving closed-loop system with three interlocking rings.

[0024] Figure 4 This is a comparison curve showing the sensitivity of ECG-related structural abnormality shifts under motion artifact signal-to-noise ratio in this invention.

[0025] Figure 5 This is a comparison chart of the performance matrix and overall average ranking of this invention;

[0026] Figure 6 This is a comparison chart of the logarithmic scale of the false positive rate (FPR) across different scenarios in this invention.

[0027] Figure 7 This is a radar chart comparing the decoupling of the five physiological modal characteristic loads and the purity of the present invention.

[0028] Figure 8 This is a waterfall plot showing the evolutionary attribution of the average FPR stepwise ablation across different scenarios in this invention. Detailed Implementation

[0029] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] like Figures 1-8 As shown, this embodiment provides a multimodal dynamic physiological monitoring method based on event-driven and Riemannian manifolds, including the following steps:

[0031] S1: Collect multiple physiological signals from the monitored object, pre-condition and quality assessment of each physiological signal, and obtain the conditioning signal and the corresponding quality score;

[0032] In this embodiment, such as Figure 1 As shown, the monitored subjects wear wearable multimodal physiological monitoring devices during the continuous monitoring period. The continuous monitoring period can be a preset period for postoperative follow-up, rehabilitation observation, or continuous health monitoring outside the hospital. The wearable multimodal physiological monitoring device is used to collect multiple heterogeneous physiological signals during the user's daily life, sleep, light activity, and rehabilitation activities, and converts the collected multiple physiological signals into an asynchronous event stream suitable for subsequent covariance estimation.

[0033] In a preferred embodiment, the multiple physiological signals include five signals: electrocardiogram (ECG), photoplethysmography (PPG), blood pressure estimation (BP), thoracic respiratory impedance, and wrist acceleration. The PPG signal can be acquired using a red / infrared dual-wavelength photoelectric channel and can be further used to estimate blood oxygen saturation. The blood pressure estimation signal BP can be estimated based on the pulse wave conduction time (PTT) between the R wave and PPG waveform in the ECG signal, and can be periodically calibrated using intermittent cuff blood pressure NIBP; the wrist acceleration signal can be represented by the sum of triaxial acceleration values ​​to characterize the body's dynamic state.

[0034] The input for this step is a multi-channel raw physiological signal:

[0035] ;in, Indicates the first Original continuous physiological signals of the road This indicates the number of channels participating in the data collection. In this embodiment, The five signals are ECG voltage, PPG light intensity, blood pressure estimate, respiratory impedance value, and acceleration amplitude.

[0036] The output of this step includes two types of data: the first type is the set of event tuples within the sliding time window. Used for subsequent covariance estimation steps; the second type is the quality score vector. It is used as a sample weight in subsequent covariance estimation and as a stability window criterion in the scene baseline maintenance and anomaly isolation steps.

[0037] in, , Indicates the first Road mode at time Quality rating This indicates the transpose of a matrix or vector.

[0038] Traditional constant frame rate sampling architectures typically force multiple signals to be synchronized to a higher sampling rate channel, causing redundant sampling to continue even during stable phases such as deep sleep and rest. Furthermore, response delays may occur during short-term abrupt events due to fixed sampling intervals or window-wide waiting. This implementation uses differential modulation event processing, ensuring that each channel outputs an event only when the signal changes sufficiently relative to a reference level. This allows the data output rate to be correlated with the activity of physiological signals, rather than being tied to a fixed sampling rate.

[0039] At the level of covariance estimation, the estimation accuracy of traditional fixed frame rate sampling can be characterized by the following relationship:

[0040] ;

[0041] in, This represents the variance of the covariance estimator. This represents the number of samples within the sliding time window. Indicates the sampling interval. This represents the covariance estimate. This relationship indicates that within a fixed sliding time window, if the sampling interval increases or the number of effective samples decreases, the stability of the covariance estimate decreases; conversely, increasing the number of samples by maintaining a fixed high frame rate increases power consumption and redundant computation. This implementation employs an event-driven front-end, causing the effective sample rate to vary with the event occurrence rate, thereby reducing redundant sampling during stable phases and increasing event output density during signal abrupt changes.

[0042] S1.1 Multi-channel pretreatment and quality assessment:

[0043] The reliability of event-based processing depends on the signal quality before it enters the differential modulator. If baseline drift, power line interference, motion artifacts, poor electrode contact, or unstable photoelectric sensor contact are not suppressed, these interferences may be misidentified as valid events by the differential modulator, thus affecting the subsequent covariance matrix. Therefore, before performing event-based processing, the raw physiological signals from each channel are pre-conditioned and their quality assessed in parallel.

[0044] For the Path primitive physiological signals The conditioned signal is obtained after processing by the corresponding pre-conditioning filter:

[0045] ;in, Indicates the first Physiological signals after pathway regulation Indicates the first The impulse response of the bandpass conditioning filter corresponding to the path mode. This represents the convolution operation.

[0046] In this implementation, different conditioning strategies are used for different modalities. For the ECG channel, the filter passband covers the main energy frequency band of the QRS complex to preserve the R wave spike; for the PPG channel, the filter passband covers the main harmonics of the pulse wave to preserve the ascending segment of systole; for the blood pressure estimation channel, the filter passband preserves the cardiac cycle and respiratory oscillation-related fluctuations; for the thoracic respiratory impedance channel, the filter passband covers the normal respiratory frequency band; for the acceleration channel, the filtering is mainly used to suppress high-frequency sensor noise and preserve low-frequency or mid-to-low-frequency body movement characteristics such as body position switching and gait start-stop.

[0047] Quality scores are obtained using the following formula:

[0048] ;

[0049] in, Indicates the first Road physiological mode at time Quality rating; This represents the normalized short-time signal-to-noise ratio; This indicates the signal saturation ratio within the current window; This represents the normalized sensor contact state index; , and These represent the signal-to-noise ratio weight, unsaturation weight, and contact state weight corresponding to the k-th physiological modality, respectively. This restricts the calculation results to between 0 and 1. The weights are set according to the following formula:

[0050] ;

[0051] In one exemplary implementation, the 5th and 95th percentiles of the short-time signal-to-noise ratio (SNR) are first calculated within a calibration window free of significant anomalies, and the SNR is then linearly normalized to 0 to 1. The signal saturation ratio is determined by the ratio of the number of sampling points reaching the upper and lower limits of analog-to-digital conversion to the total number of sampling points within the window. The sensor contact status index is determined by electrode impedance, photoelectric DC component stability, or wearing fixation status. Different physiological modalities can be assigned different weights, and each weight remains unchanged after device calibration or is only updated during the offline calibration phase.

[0052] The quality score mapping function The short-time signal-to-noise ratio (SNR), unsaturation ratio, and contact condition indicators show a monotonically increasing relationship. A higher SNR, lower saturation ratio, and better contact condition result in a higher quality score; conversely, a lower SNR, higher saturation ratio, or poorer contact condition leads to a lower quality score. For example, when turning over at night causes pressure on the fingertip PPG channel, the PPG channel quality score... The ECG channel quality score may decrease temporarily; when the ECG electrode becomes detached or has poor contact, the score may decrease. It can approach 0.

[0053] The quality score is used in at least three subsequent stages: first, as a weighting coefficient for each mode in the covariance increment update to reduce the impact of low-quality modes on the covariance matrix; second, as a basis for judging whether a certain mode is temporarily disabled in the temporary mode reduction comparison; and third, as a hard threshold for the stability window criterion when normal stable samples are fed back to update the long-term base point, thereby preventing low-quality samples from contaminating the long-term base point.

[0054] S2 performs differential modulation event processing on each conditioning signal and uniformly stamps the trigger events based on the shared clock to generate asynchronous event tuples;

[0055] S2.1 Signal eventification based on differential modulation:

[0056] The pre-conditioned signal then enters the differential modulation process. The basic method of differential modulation is to maintain a reference level for each mode. And continuously compare the current conditioning signals The difference between the reference level and the reference level. An event is only output when this difference crosses the trigger threshold of the corresponding mode; after the event is output, the reference level steps by a threshold step along the direction of signal change.

[0057] For the Path mode, first The trigger time of an event can be determined by the following formula:

[0058] ;

[0059] in, Indicates the first Road mode 1 The trigger timestamp of each event; This indicates the trigger timestamp of the previous event in this modality; Indicates the first The refractory period of the path mode is used to shield against consecutive false triggers on the same channel for a short period after an event is triggered; Indicates the first Road conditioning signals; This indicates the updated reference level obtained after the previous event was triggered; Indicates the first The event trigger threshold of the road mode has the same dimensions as the corresponding conditioning signal; This indicates the earliest time when the condition is met.

[0060] In one implementation, the event trigger threshold Instead of being a fixed constant, it is inversely modulated by the posterior confidence of the scene output from subsequent scene recognition steps. In the initial stage, the basic threshold can be set according to the stability window statistics of each modal signal. For example, the basic threshold of the ECG channel can be taken as a lower quantile of the normal upward amplitude of the R wave, so that each QRS complex triggers at least one valid event near the R peak; the basic threshold of the PPG channel can be taken as a fraction of the typical upward amplitude of the systolic phase, to capture the rising edge of the pulse wave; the basic threshold of the blood pressure estimation channel can be taken as the quantile of the blood pressure fluctuation amplitude within the cardiac cycle; the basic threshold of the respiratory impedance channel can be taken as the impedance crossing at the transition between inspiration and expiration; and the basic threshold of the acceleration channel can be taken as the change in the combined amplitude corresponding to the boundary between resting and mild activity.

[0061] After the event is triggered, the system calculates the event polarity:

[0062] in, Indicates the event polarity; when the signal crosses a threshold relative to the previous reference level, When the signal crosses a threshold relative to the previous reference level, .

[0063] The reference level is updated stepwise using the following formula:

[0064] in, Indicates the first The reference level is updated after the current event is triggered. Since the reference level only changes by a step when the event is triggered and remains unchanged between two adjacent events, a piecewise constant reference level sequence is formed. This piecewise constant characteristic allows the subsequent covariance estimation step to directly perform algebraic incremental updates based on the event stream, without needing to reconstruct the event stream into a fixed frame rate waveform.

[0065] An event tuple is generated after each event is triggered:

[0066] ;

[0067] in, Indicates the first Road mode 1 event tuples, Indicates the channel index. Indicates the trigger timestamp. Indicates the polarity of the event. This represents the quality score at the time the event was triggered. By writing the quality score into the event tuple, the subsequent covariance estimation step can obtain the reliability information of the event simultaneously when it arrives.

[0068] Compared to fixed frame rate sampling, the event tuple has a smaller data volume. Channel indexes, timestamps, event polarities, and quality scores can all be represented with fewer bits, facilitating edge-side caching and low-bandwidth transmission between the edge and cloud. During deep sleep or resting phases, the event output rate decreases due to slow signal changes; however, when abnormal deviations occur in ECG-pulse coupling, respiratory rhythm, blood pressure-related abnormal deviations, or significant body movement changes, the event output rate of the corresponding modality increases, thereby improving the ability to capture mutation information.

[0069] S2.2 Cross-modal timestamp unification and event aggregation:

[0070] Because different physiological signals have different sampling circuits, physical change rates, and event frequencies, cross-modal time base drift may occur if each channel is stamped independently. For example, ECG events and PPG events have a clear temporal coupling relationship physiologically. If the two are misaligned due to clock drift, it will affect the expression of the subsequent ECG-pulse coupling relationship in the covariance matrix.

[0071] To address this issue, this implementation uses a shared timestamp bus. Event trigger pulses from each differential modulator are connected to the same reference clock. The first-order arbitrator stamps the events uniformly according to their arrival order. For modes with different response delays in the sampling circuits, the fixed delay of each channel can be measured through the standard calibration signal during the equipment calibration phase, and the delay compensation value is written into the storage unit; during operation, the arbitrator performs deterministic compensation for the event timestamps according to the preset channel delay calibration table.

[0072] After processing by the shared timestamp bus, the timestamp difference between any two modal events is: Comparisons can be made under a unified clock reference, where and Represents different modal channels, and This indicates the corresponding event number. Therefore, cross-modal time coupling parameters such as PTT can be preserved, and the temporal relationships between ECG-pulse coupling, respiration-blood pressure coupling, and body motion-pulse changes can be incorporated into the subsequent covariance estimation step.

[0073] In the sliding time window Internally, the system maintains a set of event tuples through an event buffer queue:

[0074] ;in, This represents the union of all modal event tuples within the current sliding time window; This indicates the length of the sliding time window, which can be set according to the needs of short-term abnormal offset monitoring, medium-term fluctuation monitoring, or long-term baseline observation. In this embodiment, the number of modes participating in the monitoring is indicated. .

[0075] Meanwhile, each channel's quality scoring unit outputs a quality scoring vector in real time:

[0076] The event buffer queue and quality score vector are outputs of this step and directly fed into the subsequent covariance estimation step. Specifically, the event tuple set... The event-driven update timing used to determine the reference level vector, and the quality score vector. Used to construct the covariance weighted matrix and participate in the determination of stable samples.

[0077] In this embodiment, the event tuple set is not reconstructed into a fixed frame rate waveform through zero-order hold or linear interpolation. Waveform reconstruction refers to inverting or padding an asynchronous sparse event sequence into a continuously sampled sequence with equal intervals for computation by traditional timing algorithms. If waveform reconstruction is performed, redundant computation at the fixed frame rate will be reintroduced, weakening the low power consumption and low redundancy advantages of event-driven sampling. Therefore, this embodiment directly delivers the event stream to the covariance estimation step, with subsequent steps performing event-driven algebraic incremental updates based on the reference level vector, maintaining the efficiency advantage of the event-driven front-end in the data path.

[0078] Through step S2 described above, this invention completes the conversion from multiple continuous physiological signals to an asynchronous event tuple set and a quality score vector. This step achieves low-redundancy sampling and provides directly usable input for subsequent multimodal covariance estimation, symmetric positive definite manifold representation, scene posterior calculation, dynamic base point construction, and anomaly isolation loop closure.

[0079] S3, based on the asynchronous event tuple and quality score, perform event-driven incremental update and positive definite constraint on the covariance matrix of the multi-path physiological signals to obtain the current covariance matrix on the symmetric positive definite manifold;

[0080] In this implementation, the event-driven front-end outputs a set of asynchronous event tuples for each modality. However, the triggering time, signal dimensions, amplitude range, and frequency characteristics of different modal events are not consistent. If only one-dimensional fluctuations are calculated for each physiological signal separately, the coupling relationships between modalities such as ECG-pulse, blood pressure-heart rate, respiration-blood pressure, and body movement-pulse cannot be expressed. Therefore, this step converts the asynchronous event stream and quality score within the sliding time window into a comparable joint observation, namely, a multimodal weighted covariance matrix.

[0081] The input for this step includes: a set of event tuples within the sliding time window. Quality score vector and the length of the sliding time window In the five-signal implementation, the quality scoring vector is:

[0082] ;in, to These correspond to the quality scores of ECG, PPG, blood pressure estimation, respiratory impedance, and acceleration mode, respectively.

[0083] The output of this step is a fifth-order multimodal weighted covariance matrix:

[0084] ;in, Indicates all A symmetric positive definite manifold constructed from symmetric positive definite matrices. The diagonal elements of the covariance matrix represent the fluctuation intensity of each mode itself, while the off-diagonal elements represent the coupling fluctuation intensity between different modes. For example, Indicates electrocardiogram-related fluctuations. Indicates pulse wave-related fluctuations. This indicates fluctuations related to blood pressure estimation. This indicates fluctuations related to respiratory impedance. This indicates fluctuations related to activity intensity; This indicates the electrocardiogram-pulse coupling relationship. This indicates the coupling relationship between electrocardiogram and blood pressure estimation. This indicates the coupling relationship between respiration and blood pressure estimation. This indicates the electrocardiogram-body motion coupling relationship.

[0085] S3.1 A reference level vector is formed from the event stream:

[0086] Because the differential modulation event processing in step S2 causes the reference level of each channel to be updated only when the event is triggered, therefore, at any given time... The system can form a reference level vector from the current reference levels of each channel:

[0087] ;in, Indicates the first Road mode at time The reference level, This indicates the transpose. The reference level vector can be viewed as an instantaneous joint representation of the current multimodal state within an event-driven framework.

[0088] When the Road mode occurrence event At that time, the reference level of the channel steps according to the following formula:

[0089] ;in, For the first Lu Di The triggering time of an event, For the polarity of the event, For the first The trigger threshold for the road mode. Let the step size be: In this case, the event only changes the first element in the reference level vector. One component remains unchanged, while the remaining modal components remain unchanged.

[0090] S3.2 Event-Driven Covariance Analysis Incremental Update:

[0091] To avoid reconstructing the asynchronous event stream as a fixed frame rate waveform, this step directly uses the reference level step caused by the event trigger to perform a local algebraic update on the covariance matrix. Let... For the first The nth standard basis vector, i.e., only the nth standard basis vector. If one component is 1 and the rest are 0, then the first component is 1. Lu Di The covariance increment matrix triggered by a single event can be represented as:

[0092] ;

[0093] in, Indicates the first The first physiological modality of the road The covariance increment matrix formed by asynchronous events; This indicates the reference level step caused by this asynchronous event; Indicates the first One standard basis vector; This represents the multimodal reference level vector prior to the arrival of this asynchronous event; This represents the weighted average vector before the arrival of this asynchronous event. Because... Except for the All components are 0 except for the first component. The covariance increment matrix is ​​only 0 in the first component. line, number List and No. Each diagonal element generates an increment.

[0094] Therefore, each time an event arrives, the system only needs to update the covariance matrix containing the event with respect to the first event. The first road mode related row and number The column remains unchanged while the rest of the sub-blocks remain the same, thus avoiding batch recalculation of the entire window.

[0095] Based on this, the covariance matrix is ​​updated using a weighted incremental method corresponding to the length of the rolling time window:

[0096] ;

[0097] ;in, This indicates the average event interval. This setting allows event-driven updates to have a time memory length that is approximately equivalent to traditional sliding window statistics.

[0098] Simultaneously, the weighted mean vector is updated synchronously using the following formula:

[0099] ;in, This represents the weighted mean vector before the event arrives. This represents the updated weighted mean vector.

[0100] S3.3 Quality scores participate in covariance weighting:

[0101] To avoid low-quality modes contaminating the covariance matrix, a quality score vector is introduced during covariance updates. Specifically, the quality score is constructed as a diagonal matrix:

[0102] ;

[0103] in, Let be the diagonal matrix of the quality scores. The quality-weighted covariance matrix can be expressed as:

[0104] ;

[0105] in, This represents the diagonal matrix of quality scores composed of quality scores from each physiological modality; This represents the covariance matrix after quality weighting and before positive definite constraints. Quality weighting is only used to form the current manifold observation matrix. When the next asynchronous event arrives, the event-driven covariance state saved at the previous time step is still used as the starting point for recursion to avoid the quality score being repeatedly multiplied.

[0106] When the quality score of a certain mode decreases, the row and column contributions corresponding to that mode are simultaneously weakened. For example, when the PPG channel is weakened due to loose fitting or pressure... When the PPG is reduced, the contributions of the second row and second column in the covariance matrix are automatically reduced; when the acceleration mode increases dramatically and is accompanied by a decrease in the quality of other modes, it can suppress the non-real coupling fluctuations caused by body motion artifacts from entering the long-term baseline.

[0107] S3.4 Positive definite regularization and adaptive condition number:

[0108] Because event-driven incremental updates may cause variance terms to decay in some modes when there are no events triggered for a long time, the covariance matrix may suffer from condition number deterioration or approach singularity. To ensure the stable execution of subsequent manifold distance calculation, scene base point maintenance, and tangent space mapping, a positive definite regularization is applied to the covariance matrix:

[0109] ;

[0110] in, This represents the current covariance matrix after positive definite constraints; express An identity matrix of order 1; Indicates time The positive definite regularity coefficients.

[0111] The time-varying positive definite canonical coefficients can be determined by the following formula:

[0112]

[0113] in, Indicates the basic regularity strength; The condition number represents the covariance matrix before positive definite constraints; This indicates the maximum number of preset conditions. According to trace mean to Multiplier setting, The accuracy can be set according to the numerical precision on the end side. to .

[0114] This process ensures the output matrix is ​​guaranteed. It is a strictly symmetric positive definite matrix, which can be used as input for subsequent scene recognition, dynamic base point construction and tangent space offset measurement.

[0115] S3.5 Modal Temporary Order Reduction Comparison:

[0116] In actual continuous monitoring, a particular mode may experience temporary inactivation. For example, the PPG channel may experience a decrease in quality score due to pressure from the wearer or sweating, the ECG channel may experience a quality score approaching 0 due to poor electrode contact, and the acceleration channel may generate strong artifacts due to vigorous movement. If a full five-dimensional covariance matrix is ​​forcibly used for comparison, low-quality modes may affect the overall judgment.

[0117] Therefore, when the quality score of a certain mode is lower than the preset mode disabling threshold... At that time, the system constructs a selection matrix. Only the available modes that meet the quality score requirements are retained. Let the number of available modes be... ,but: ;

[0118] By selecting a matrix to project the covariance matrix and the long-term basis points of the current scene onto the same-dimensional principal submatrix:

[0119] ; ;

[0120] in, This represents the reduced covariance matrix. This represents the long-term baseline of the current scenario after the order reduction. This represents the long-term baseline of the current scenario.

[0121] For example, when the PPG channel is temporarily disabled, the selection matrix is ​​used. The row corresponding to the second mode is deleted, and the original five-dimensional covariance matrix is ​​projected into a four-dimensional covariance submatrix. The system then... Manifold distance calculation and offset measurement continue on the submanifold. Once the PPG channel quality score is restored, the system re-emphasizes the full five-dimensional covariance matrix for comparison.

[0122] Since the principal submatrices of a symmetric positive definite matrix remain symmetric positive definite matrices, the aforementioned reduced-order projection does not violate the mathematical foundation of subsequent manifold metrics. This mechanism allows the system to continue operating even when some sensing modes are temporarily disabled without restarting the edge computing process or retraining the model.

[0123] S3.6 Multi-scale Block Diagonal Expansion:

[0124] In a further embodiment, to simultaneously characterize short-term anomalous offsets, medium-term state fluctuations, and long-term baseline drift, the covariance matrix can be extended into a multi-scale block diagonal structure. Specifically, three sliding time windows of different scales are set. These correspond to short-term, medium-term, and long-term time scales, respectively.

[0125] Perform the above event-driven incremental update for each sliding time window to obtain:

[0126] ];

[0127] in, .

[0128] The covariance matrices at the three scales are constructed as block diagonal matrices:

[0129] ;

[0130] in, The block diagonal matrix construction operator is represented. It is a multi-scale block diagonal covariance matrix.

[0131] In this structure, short-time window Used to capture sudden physiological shifts lasting from seconds to tens of seconds; intermediate window Used to capture fluctuations in the internal state of a scene; long-term window This method is used to capture intraday or multi-day baseline drift. Sub-blocks at different scales share the same event stream input but maintain their respective means, covariances, and forgetting factors. The diagonal block structure prevents interference between different scales, allowing subsequent migration metrics to read migration results at each scale separately, thus avoiding the masking effect between short-term abrupt changes and long-term drifts.

[0132] Through the above processing, this step will generate the asynchronous event tuple set. and quality rating vector Transform into the covariance matrix on a symmetric positive definite manifold or multi-scale covariance matrix The covariance matrix preserves both the individual fluctuations of each mode and the coupling fluctuations between modes, and improves the stability during continuous monitoring through quality weighting, positive definite regularization, and temporary order reduction mechanisms.

[0133] When an event is triggered, only the relevant rows and columns are updated algebraically to avoid fixed frame rate waveform reconstruction and full window batch calculation. The quality score enters the covariance construction process, so that the low quality modes are automatically downweighted. Positive definite regularization ensures that the covariance matrix is ​​stable in the SPD(5) manifold. Temporary downgrading comparison ensures that the system can continue to operate when some modes are disabled for a short time. Multi-scale expansion can simultaneously characterize short-term offset and long-term baseline drift.

[0134] Therefore, this step realizes the transformation from asynchronous event streams to comparable symmetric positive definite manifold joint observations, providing a mathematical foundation for subsequent scene posterior calculations, dynamic base point construction, and tangent space offset metrics.

[0135] S4, calculate the manifold distance between the current covariance matrix and the long-term baselines corresponding to different physiological scenarios, determine the current scenario and its posterior confidence, and adjust the event triggering threshold in step S2 in reverse according to the posterior confidence.

[0136] In this embodiment, the multimodal weighted covariance matrix obtained after step S3 The points have been constrained to be points on a symmetric positive definite manifold. Since the monitored objects will be in different physiological scenarios during continuous monitoring, and the normal physiological fluctuation range and modal coupling structure differ in different scenarios, the covariance matrix at all times cannot be compared with the same fixed baseline.

[0137] For example, in a postoperative follow-up monitoring scenario, the monitored subject may experience four physiological scenarios within a day: resting and awake, light activity, sleep, and rehabilitation exercise. During sleep, heart rate is lower, acceleration fluctuations are smaller, and respiratory rhythm is more stable; during light activity, acceleration, pulse wave, and respiratory impedance may all show moderate fluctuations; during rehabilitation exercise, heart rate, respiratory rate, and physical activity intensity are all significantly increased. If a uniform baseline is used without distinguishing between scenarios, normal scenario transitions can easily be misjudged as abnormal shifts, or truly necessary abnormal shifts may be missed within a specific scenario.

[0138] The input for this step includes: the current covariance matrix. The posterior probability vector of the scene at the previous moment Scene transition matrix , and the reverse re-judgment trigger flag returned by the abnormal offset classification step.

[0139] The output of this step includes: current scene label. posterior probability vector of the current scene And the trigger thresholds for each physiological modality event used to be sent to the event-driven front-end. .

[0140] The scene set can be set as follows: ;

[0141] Where 1 represents a resting and awake scenario, 2 represents a light activity scenario, 3 represents a sleep scenario, and 4 represents a rehabilitation exercise scenario. The posterior probability vector for each scenario is: ;

[0142] in, ,and: ,Right now Located in a four-dimensional probability simplex Inside.

[0143] Initialization of long-term baselines for various physiological scenarios in S4.0:

[0144] After the initial run or the failure of long-term baselines, the system performs cold start calibration. Cold start can be achieved by combining pre-set population prior baselines with individual stability windows: before obtaining sufficient individual samples, the initial scenario comparison is completed using the pre-set population prior baselines on the device; after obtaining stability windows that meet the quality score thresholds for each scenario (resting and awake, light activity, sleep, and rehabilitation exercise), long-term individual baselines are established based on the stability covariance matrix of the corresponding scenario.

[0145] ;

[0146] in, Indicates the first The initial long-term baseline of a physiological scenario; Indicates the use of initialization of the first The number of normal and stable windows for long-term baselines in each physiological scenario; Indicates the first The first physiological scenario The covariance matrix corresponding to each normal stable window; and and represent the matrix logarithm and matrix exponent, respectively. For example, for each physiological scenario, an initial long-term baseline for the individual can be established after obtaining no fewer than 20 stable windows; if the number is not reached, the population prior baseline continues to be used, and unstable samples and anomalous offset samples from the isolated record set are prohibited from participating in the initialization.

[0147] S4.1 Scene likelihood calculation based on SPD manifold distance:

[0148] In this embodiment, the system maintains a long-term baseline for each physiological scenario. ,in:

[0149] The long-term base point Indicates the first The long-term normal reference center of the multimodal covariance matrix under various physiological scenarios. For example, the long-term baseline of the sleep scenario may be characterized by low acceleration-related variance and relatively stable respiratory impedance-related fluctuations; the long-term baseline of the rehabilitation exercise scenario may be characterized by high acceleration, pulse wave, and respiratory impedance-related components.

[0150] Current covariance matrix With the Long-term baseline of each scenario The distance between them is calculated using affine-invariant Riemann distance: ;

[0151] in, The current covariance matrix and the first SPD manifold distance between long-term baselines in each scenario; Represents the matrix logarithm; Denotes the Frobenius norm; Representing the long-term baseline of the scenario The matrix is ​​the negative 1 / 2 power.

[0152] Based on this manifold distance, the current covariance matrix is ​​calculated at the [number]th [location]. Likelihood in a given scenario:

[0153] ;in, Indicates the first The observed covariance matrix in each scenario The likelihood; The temperature coefficient is used to classify scenes and adjust their sharpness. Normalization constant:

[0154] ;

[0155] This makes all scenarios seem normalized.

[0156] In one implementation, an instantaneous scene estimate can be obtained first:

[0157] in, This represents the instantaneous scene prediction result obtained based on the minimum manifold distance. This instantaneous scene prediction result is not directly used as the final scene label, but rather as the input for subsequent posterior state machine updates to avoid frequent jumps near scene boundaries.

[0158] S4.2 Posterior state machine update with scene transition prior:

[0159] Relying solely on instantaneous likelihood can easily introduce jitter at scene transition boundaries. For example, when a monitored object has just moved from a resting state to a state of light activity, the current covariance matrix may be close to both the resting baseline and the light activity baseline simultaneously; if instantaneous minimum distance classification is used directly, the scene label may switch frequently between the two scenes.

[0160] Therefore, this implementation combines the current likelihood with the scene transition prior and uses a Bayesian posterior state machine to update the scene posterior probability. Let... Let be the scene transition matrix, where Indicates the transition from the previous scene Transfer to the current scene The prior probability is then used to update the posterior probability of the scenario as follows:

[0161] ;

[0162] in, At the current moment, it is in the first position. The posterior probability of a scenario; For the previous moment in the first position The posterior probability of each scenario; the denominator is used for normalization so that the sum of the posterior probabilities of all scenarios is 1.

[0163] To reduce posterior jitter near scene boundaries, first-order low-pass smoothing is applied to the posterior confidence scores of each physiological scene:

[0164] ;

[0165] in, Indicates time Belongs to the Unsmoothed posterior confidence for each physiological scenario; and represents the smoothed posterior confidence at the current time and the previous time, respectively; Represents the posterior smoothing coefficient of the scene, and Subsequent scene determination, hysteresis judgment, and event trigger threshold feedback all use smoothed posterior scene confidence.

[0166] The final scene label can be determined according to the maximum a posteriori principle: ];

[0167] In a preferred embodiment, the scene transition matrix Initialization can be performed based on prior physiological activities. For example, the probability of transitioning directly from a sleep scenario to a rehabilitation exercise scenario can be set to a lower value; the probability of transitioning from resting and awake to a sleep scenario during the night can be higher than during the day; the probability of entering a rehabilitation exercise scenario can be adjusted according to a preset rehabilitation plan or monitoring cycle. During operation, the scenario transition matrix can also be individually fine-tuned based on historical stable samples, but this fine-tuning does not involve online training of the neural network and is only used to update the statistical prior of the scenario state machine.

[0168] Through the aforementioned posterior state machine, the system can combine which scenario the current observation covariance matrix resembles with whether the transition of that scenario is reasonable in the current time period and current state, thereby reducing instantaneous misjudgments.

[0169] Furthermore, step S4 also adjusts the event triggering threshold in step S2 with hysteresis and smoothing feedback based on the posterior confidence level, as follows:

[0170] S4.3 Inverse modulation of event-triggered thresholds with hysteresis and smoothing:

[0171] Obtain the current scene label and scene posterior probability Subsequently, this implementation method not only uses it for subsequent offset measurement, but also feeds it back to the differential modulation event front end in step S2 to adjust the trigger threshold of each modal event. The front end event trigger sensitivity can dynamically change with the current physiological scene and confidence level.

[0172] If used directly Linear modulation of the threshold can lead to frequent threshold jumps due to small fluctuations in the posterior probability of the scene in the boundary region, further causing closed-loop oscillations in the "threshold-event rate-covariance-scene posterior" relationship. Therefore, this implementation uses a combination of dual-threshold hysteresis and first-order low-pass smoothing to generate the actual distribution threshold.

[0173] First, the original feedback threshold is obtained based on the posterior confidence level of the current dominant scenario:

[0174] ;

[0175] ;

[0176] in, Indicates the first Road physiological mode at time The posterior feedback factor; Indicates the first Feedback gain of physiological modes of the road; Indicates a delayed upper threshold; Indicates the current dominant scenario The smoothed posterior confidence level; This indicates that the maximum value among the values ​​within the parentheses is taken; Indicates the first The raw feedback threshold of the next moment in the physiological mode of the road; Indicates the first The basic event triggering threshold of the physiological modality of the road; Indicates the first Road physiological modalities in the current dominant scenarios The scene threshold factor.

[0177] Subsequently, a first-order low-pass smoothing was applied to the original feedback threshold:

[0178] ;

[0179] in, This indicates the final message sent to the event-driven frontend. Road physiological modal event trigger threshold; This indicates the threshold for triggering events that have already been issued; Indicates the threshold smoothing coefficient; and They represent the first The lower and upper limits of the event triggering thresholds allowed by the physiological modality of the road; This indicates the upper and lower limit truncation operation.

[0180] Simultaneously, a Schmitt double-threshold hysteresis is implemented at the state machine level. The system maintains internal threshold modulation states. When the posterior confidence of the current dominant scenario Above the upper threshold When the posterior confidence level is below the lower threshold, the system allows entry into sensitive mode; when the posterior confidence level is below the lower threshold, the system allows entry into sensitive mode. When this happens, the system exits sensitive mode and returns to normal mode, where:

[0181] ;

[0182] exist and Within the hysteresis interval, the system maintains the current mode to avoid threshold switching caused by fluctuations in the scenario confidence boundary.

[0183] In different scenarios, the threshold values ​​for each modality can be adjusted according to preset scenario threshold factors. For example, in a sleep scenario, the respiratory impedance channel threshold can be appropriately lowered to improve the sensitivity of capturing abnormal deviations in respiratory rhythm; at the same time, the acceleration channel threshold can be appropriately increased to suppress spurious events caused by body movements such as turning over. In a rehabilitation exercise scenario, the respiratory impedance channel threshold can be appropriately increased to avoid event storms caused by large changes in exercise-induced breathing; at the same time, the acceleration channel threshold can be lowered to enhance the ability to capture acute abnormal body movements.

[0184] Therefore, the output of the scene recognition module is no longer just the backend classification result, but directly controls the frontend event generation strategy, so that the event-driven frontend, covariance estimation and scene posterior update form a closed loop.

[0185] S4.4 Closed-loop stability constraints:

[0186] To avoid closed-loop oscillation caused by excessive reverse modulation, this implementation method can adjust the feedback gain. and low-pass smoothing coefficient Set stability constraints. Let the one-step composite mapping from the scene posterior to the front-end threshold, then to the covariance at the next time step, and finally to the scene posterior at the next time step be: Then it can be required that it meets the compression condition near the working point: ;

[0187] in, The Jacobian matrix represents the composite feedback mapping to the posterior probability of the scene; Represents the matrix norm; This represents the upper bound of the local sensitivity of the covariance matrix to changes in the event trigger threshold; It represents the upper bound of the local sensitivity of the scene's posterior to changes in the covariance matrix; This is the feedback gain coefficient; This is the low-pass smoothing coefficient.

[0188] when At point 1, the feedback mapping exhibits compressibility near the operating point, allowing threshold modulation and scene posterior updates to tend towards stability. In practical implementation, this can be achieved by limiting... upper limit, decrease Alternatively, the hysteresis bandwidth can be increased to satisfy this stability condition. This condition is an engineering constraint used to demonstrate that threshold inverse modulation will not infinitely amplify posterior fluctuations in the scene.

[0189] S4.5 Reversal of judgment based on anomalous evidence:

[0190] During continuous monitoring, the current scene state machine may lag behind abrupt changes in the actual scene. For example, when a monitored object suddenly enters a highly dynamic state from a resting state, the current covariance matrix may have a large distance relative to all known long-term baselines of the scene, and the modal contributions may exhibit a wide distribution. This situation may not be a true multimodal anomalous shift, but rather a scene mismatch caused by a lag in scene recognition.

[0191] To address this issue, this implementation introduces a reverse re-judgment branch for abnormal evidence. This branch is initiated by the reverse re-judgment trigger flag returned by the subsequent abnormal offset classification step. When the abnormal offset classification step determines that the current sample meets the scene mismatch condition, the reverse re-judgment trigger flag is set.

[0192] Upon triggering, this step resets the current scene's posterior probability vector to a uniform distribution:

[0193] Among them, arrows This indicates a forced assignment. This operation means temporarily abandoning prior historical scenarios, so that the scenario judgment in the next moment is mainly dominated by the manifold likelihood between the current covariance matrix and the long-term basis points of each scenario.

[0194] After the posterior reset, the system re-executes the manifold scene likelihood calculation in S4.1 and the posterior update in S4.2. Since the priors of each scene are temporarily balanced, the scene closest to the current observation will be selected more quickly. If a new scene label is obtained after reclassification, the threshold modulation state machine is reset synchronously, restoring the internal threshold modulation state to normal mode to prevent the sensitive modes left over from the old scene from continuing to affect the new scene.

[0195] Meanwhile, the subsequent dynamic base point construction steps will perform geodesic interpolation between the long-term base point of the old scene and the long-term base point of the new scene, so that the measurement reference system changes continuously during the scene switching process, and avoids the forced scene switching itself being mistaken for a new abnormal offset.

[0196] The reverse re-judgment of abnormal evidence is not a routine step executed every time, but is only initiated when the scenario mismatch trigger condition is met. This can handle sudden scenario switching and avoid frequent re-judgments that could cause scenario state oscillations.

[0197] S4.6 Output and technical effects of this step:

[0198] After the above processing, this step outputs the current scene label. , Scene posterior probability vector and the event trigger threshold for the next time step in each modality .

[0199] Among them, the current scene label The delivery process includes long-term baseline maintenance and dynamic baseline construction steps; scenario posterior probability vector. The delivery anomaly offset classification step is used to construct an adaptive threshold; event-triggered threshold. The data is sent back to the event-driven frontend to adjust the trigger sensitivity of each modal event in the next moment.

[0200] By calculating the distance between the current covariance matrix and the long-term baseline of each scene on the SPD manifold, scene recognition is based on a multimodal coupling structure rather than a single scalar threshold. Scene transition matrix and Bayesian posterior update reduce scene boundary jitter. Hysteresis and smoothing mechanisms prevent frequent jumps in event trigger thresholds. Scene posterior back-modulation of front-end thresholds enables the front-end event triggering strategy and back-end scene judgment to form a closed loop. Abnormal evidence reverse re-judgment mechanism enables the system to quickly correct scene labels when there are sudden scene changes or scene mismatches.

[0201] This provides a stable scenario reference basis for subsequent dynamic base point construction and abnormal offset judgment.

[0202] S5. Construct a dynamic base point for manifold mapping based on the posterior confidence level, and perform tangent space mapping on the current covariance matrix with the dynamic base point as a reference to obtain the offset magnitude and modal contribution of each physiological mode.

[0203] In this embodiment, the multimodal covariance matrix at the current time has been obtained through the aforementioned steps. posterior confidence level of the scene and the long-term baseline corresponding to different physiological scenarios. .in, This indicates the current combined fluctuation state of the five physiological signals. Indicates the first The long-term normal covariance center under physiological scenarios.

[0204] because and Both lie on symmetric positive definite manifolds, and they cannot be directly subtracted using ordinary matrices. If calculated directly... The resulting matrix cannot guarantee that it is still within a symmetric positive definite manifold, nor can it represent an offset vector with a clear geometric meaning. Therefore, this step first constructs a dynamic base point (P(t)) based on the posterior confidence of the scene, and then uses this dynamic base point as the origin to map the current covariance matrix to the tangent space to obtain a tangent space offset matrix that can be used for Euclidean operations, and calculates the offset magnitude and modal contribution accordingly.

[0205] The input for this step includes: the current covariance matrix. posterior confidence of the current scene Long-term baseline for various physiological scenarios ;

[0206] S5.1 Determination of dynamic baselines for the stable and transitional phases of a scenario:

[0207] During the stable period of the scene, if the current scene label is If the posterior confidence level of the current scenario satisfies the stability condition, then the long-term baseline of the current scenario is directly adopted as the dynamic baseline: ;

[0208] in, This is the dynamic base point used for manifold mapping at the current moment. For the current scenario The corresponding long-term base point.

[0209] During the scene transition period, if the scene changes from the old scene... Towards new scenarios Switching will not directly change the mapping base point from Hard switch to Instead, a dynamic base point is constructed along the SPD manifold geodesic between the two. Specifically, the dynamic base point is determined by the following formula: ;

[0210] in, This indicates the long-term baseline of the scene before the switch. This indicates the long-term baseline of the scene after the switch. express The principal square root of the matrix, Denotes its inverse matrix. This represents the geodesic interpolation weight.

[0211] Interpolation weights It can be determined based on the posterior confidence level of the new scenario. In one implementation: ;

[0212] In a smoother implementation, normalization can be performed based on the relative confidence levels of the old and new scenarios: ;

[0213] in, To prevent small positive numbers with a denominator of zero.

[0214] when hour, Approaching the long-term baseline of the old scenario ;when hour, Approaching the long-term baseline of new scenarios Therefore, the dynamic reference point can slide continuously between the old and new scene reference points as the posterior confidence of the scene changes, avoiding artificial spikes in the offset amplitude caused by sudden changes in the reference frame during scene switching.

[0215] Matrix exponentiation can be achieved through eigenvalue decomposition. Let: ;

[0216] in, The eigenvector matrix, If is an eigenvalue diagonal matrix, then:

[0217] ,in, This means taking the values ​​of each eigenvalue separately. This ensures the dynamic base point. It remains a symmetric positive definite matrix.

[0218] S5.2 Affine-invariant logarithmic mapping with the dynamic base point as the origin:

[0219] Determining the dynamic baseline Then, the current covariance matrix Mapped to Let the tangent space be the origin. Specifically, first construct the whitening matrix: ;

[0220] in, This represents the whitened representation of the current covariance matrix relative to the dynamic base point.

[0221] Then, the logarithm of the matrix is ​​calculated to obtain the tangent space offset matrix: ;

[0222] in, , Represents a 5th-order symmetric matrix space; This can be understood as the tangent space offset representation of the current covariance matrix relative to the dynamic base point.

[0223] In another equivalent expression, the tangent space vector can also be written as:

[0224] ;

[0225] in, Indicated by The result of the logarithmic mapping of the manifold with base points. It is the tangent space offset matrix in whitening coordinates, which facilitates subsequent calculation of offset magnitude and modal contribution; It is a tangent space offset vector expressed in the original coordinates. Both originate from the same affine invariant logarithmic mapping.

[0226] The logarithm of a matrix can be calculated through eigenvalue decomposition. Let: ;

[0227] in, If is an eigenvalue diagonal matrix, then: ;

[0228] in, This indicates taking the natural logarithm of each eigenvalue.

[0229] This mapping possesses affine invariance. That is, when the current covariance matrix and the dynamic basis point are simultaneously subjected to the same invertible linear transformation... When in effect: ;

[0230] The offset magnitude obtained by the affine invariance metric remains unchanged. This property reduces the impact of electrode impedance variations, sensor gain drift, or overall amplitude scaling on offset judgment, allowing the system to focus more on the actual changes in the multimodal coupling structure rather than the hardware amplitude changes themselves.

[0231] S5.3 Calculate the offset magnitude based on the tangent space offset matrix:

[0232] Obtain the tangent space offset matrix Then, calculate the magnitude of the offset of the current covariance matrix relative to the dynamic base point: ;in, Indicates the offset magnitude. This represents the Frobenius norm. The larger the value, the more significant the deviation of the current multimodal covariance state from the normal baseline of the current scene; The smaller the value, the closer the current state is to the long-term normal covariance center in the current scenario.

[0233] because It is calculated in the tangent space with the dynamic base point as the origin, therefore It is not directly affected by hard jumps in the reference frame during scene transitions. During scene transitions, the dynamic base point moves smoothly along the geodesic, making... The posterior confidence level changes continuously with the scene, thereby reducing false alarms caused by abrupt changes in the reference point.

[0234] S5.4 Calculate modal contributions based on the tangent space offset matrix:

[0235] To further determine which modes or modal coupling directions the current offset mainly occurs in, this implementation method uses the tangent space offset matrix. Calculate the contribution of each mode.

[0236] For the The contribution ratio of the road mode can be determined by the following formula: ];

[0237] in, Indicates the first The percentage of the contribution of the road mode to the current offset. Represents the tangent space offset matrix The Line 1 Column elements, To prevent small positive numbers with a denominator of zero.

[0238] This calculation method will [the following] The ripple term of the first ECG mode itself and its coupling offset term with other modes are included in the contribution statistics. For example, the contribution of the first ECG mode includes not only... The reflected ECG fluctuation deviations also include , , , The offset of the coupling direction between the ECG and other modes can be calculated. This allows us to output the relative contribution of each mode to the current offset.

[0239] In the five-mode implementation, to This can be used to calculate the contribution percentages of ECG, PPG, blood pressure estimation, respiratory impedance, and acceleration modes, respectively. When the contributions of ECG and PPG are simultaneously elevated, an abnormal deviation in ECG-pulse coupling can be output as an indication parameter; when the contribution of respiratory impedance is elevated, an abnormal deviation in respiratory rhythm can be output as an indication parameter; when the contribution of acceleration is elevated and the quality scores of other modes decrease, a body motion disturbance-related indication parameter can be output. The above outputs are used for physiological state monitoring and risk indication, and are not directly used as disease diagnosis conclusions.

[0240] S5.5 Coupling Direction Decomposition Based on Feature Direction:

[0241] In a further embodiment, to obtain a finer-grained interpretation of the coupling direction, the tangent space offset matrix can be adjusted. Perform eigenvalue decomposition: ;

[0242] in, The characteristic direction matrix, This is an eigenvalue diagonal matrix. Each eigendirection represents a multimodal coupling offset direction, and the absolute value of each eigenvalue represents the offset intensity in that direction.

[0243] The system can select the characteristic direction corresponding to one or more eigenvalues ​​with the largest absolute value as the dominant offset direction, and output the modal coupling interpretation based on the load distribution of the characteristic direction across the five modes. For example, if the dominant direction has a high load on the ECG and PPG modes, it indicates that the current offset is mainly concentrated in the ECG-pulse coupling direction; if the dominant direction has a high load on the respiratory impedance and blood pressure estimation modes, it indicates that the current offset is mainly concentrated in the respiratory-blood pressure related coupling direction.

[0244] The directional decomposition result can be compared with the aforementioned modal contributions. They are used together for subsequent anomaly offset classification.

[0245] S5.6 Handling of multi-scale or order-reduction cases:

[0246] If the aforementioned covariance estimation step employed temporary modal order reduction comparison, then in this step... , , , and All after order reduction Computation within the submanifold. Wherein, This indicates the number of currently available modes. Modal contributions are normalized only within the set of currently available modes; missing modes do not participate in the current offset metric.

[0247] If the aforementioned covariance estimation step employs multi-scale block diagonal expansion, then dynamic basepoint construction and tangent space mapping can be performed on covariance sub-blocks of different scales respectively. For example, for short-term, medium-term, and long-term windows, the following results are obtained: ;

[0248] And calculate the corresponding offset magnitude respectively: ;

[0249] Thus, the system can distinguish between short-term abrupt shifts, medium-term state fluctuations, and long-term baseline drift, avoiding mutual masking of shifts at different time scales.

[0250] S6. Based on the offset amplitude, modal contribution, posterior confidence and quality score, the current physiological state is classified as abnormal. Normal samples that meet the stability conditions are used to update the long-term baseline of the corresponding physiological scenario. Abnormal offset samples are written into the isolation record set and are prohibited from participating in the update of the long-term baseline.

[0251] In this embodiment, the aforementioned steps have yielded the current covariance matrix's offset magnitude, modal contribution, scene posterior confidence, and quality score relative to the dynamic baseline. The offset magnitude characterizes the overall deviation of the current multimodal physiological state from the current physiological scene's long-term baseline; the modal contribution characterizes which physiological modalities or modal coupling directions the offset is primarily concentrated in; the scene posterior confidence characterizes whether the current scene recognition is stable; and the quality score determines whether the current sample is suitable for backflow to update the long-term baseline.

[0252] The input for this step includes: offset magnitude. Modal contribution vector posterior confidence level of the scene Current scene tags Local bandwidth of the scene Quality score vector Current covariance matrix and optional affine invariant families , .

[0253] The modal contribution vector can be expressed as: ;

[0254] in, Indicates the first The first mode or the first The contribution percentage of each dominant coupling direction to the current offset, and satisfying: ;

[0255] The quality rating vector is: ;

[0256] in, Indicates the first Road mode at time Signal quality score.

[0257] The output of this step includes: current offset type label. Long-term baseline of the scenario after normal sample backflow update Abnormal sample isolation record set And the scene re-judgment trigger signal that is fed back into the scene recognition process in step S4 when the scene does not match.

[0258] S6.1 Constructing a scene-adaptive threshold family:

[0259] To avoid false alarms or false negatives caused by fixed thresholds under different physiological scenarios, this step constructs an adaptive threshold family based on the local bandwidth of the current scenario, the posterior confidence of the scenario, and the quality score. The local bandwidth of the scenario... This represents the acceptable range of normal covariance fluctuations in the current scenario; the posterior confidence level of the scenario. The quality score indicates the stability of the current scene recognition. Used to reflect whether the current sample is reliable.

[0260] In one implementation, the global offset magnitude threshold can be constructed using the following formula:

[0261] ;in, This represents the global offset magnitude threshold in the current scene; Represents the global offset threshold scaling constant; Indicates the local bandwidth of the current scene; This represents the posterior confidence level of the current dominant scenario; This represents the quality score adjustment factor; This represents the lowest quality score among the currently participating monitoring modes.

[0262] The formula indicates that the greater the background fluctuation of the current scene, the more lenient the threshold; the more uncertain the scene recognition, the more conservative the threshold; and the worse the current signal quality, the more relaxed the threshold should be to avoid misjudgment caused by low-quality signals.

[0263] The concentration threshold for a single indicator can be set using the following formula: ;

[0264] in, Indicates the concentration threshold for a single indicator; This represents the scaling constant for the concentration of a single index; To prevent small positive numbers with a denominator of zero, this threshold is used to determine whether the current offset is mainly concentrated in a single mode or a single coupling direction.

[0265] The trend offset threshold can be set using the following formula: ;

[0266] in, Indicates the trend offset threshold; This represents the trend threshold scaling constant; This indicates the length of the trend observation time window. This threshold is used to determine whether the offset magnitude continues to change unidirectionally over a longer time window.

[0267] The multimodal joint offset threshold can be set using the following formula: ;

[0268] in, Indicates the multimodal joint offset threshold; This represents the joint offset threshold scaling constant.

[0269] Using the aforementioned threshold family, the system can dynamically adjust the offset classification criteria based on different physiological scenarios and current data quality. For example, in sleep scenarios, the background fluctuation is small, and the threshold is relatively low, which is beneficial for capturing small abnormal deviations in respiratory rhythm; in rehabilitation exercise scenarios, the background fluctuation is large, and the threshold is relatively high, which can avoid misjudging physiological fluctuations caused by normal exercise as abnormal deviations.

[0270] S6.2 Determine whether a sample meets the stability window criterion:

[0271] Before classifying abnormal offsets, this step first determines whether the current sample meets the basic conditions for backflow to update the long-term reference point. The stability window criterion includes at least the quality score condition, the scene stability condition, and the offset magnitude condition.

[0272] The quality scoring criteria can be expressed as: ;

[0273] in, This sets a minimum quality score threshold. When the quality score of any key modality falls below this threshold, the current sample is not allowed to be backflowed to update the long-term baseline.

[0274] The stability condition of the scene can be expressed as: ;

[0275] in, This represents the scene confidence threshold. When the posterior confidence of the current scene is insufficient, it indicates that the system may be in a scene switching or scene mismatch state, and the current sample is not suitable for updating the long-term baseline.

[0276] The offset magnitude condition can be expressed as: ;

[0277] in, This is the current offset magnitude. This is the adaptive offset amplitude threshold for the current scene.

[0278] Only when all of the above conditions are met can the current sample be classified as a normal stable sample; if any condition is not met, the current sample cannot enter the long-term baseline update channel.

[0279] S6.3 Multidimensional Offset Type Classification:

[0280] This step classifies the current sample based on offset magnitude, modal contribution, trend change, affine invariants, and scene posterior confidence. The classification results include at least normal state, single-index offset, trend offset, multimodal joint offset, and scene mismatch offset.

[0281] When the current offset magnitude exceeds the global offset threshold, and the modal contribution is highly concentrated in a single mode or a single coupling direction, the current sample is classified as a single-index offset: ,and ,in, This indicates the percentage of the largest modal contribution.

[0282] Single-index offset typically indicates a significant abnormal shift in a particular mode or coupling direction. If, in conjunction with a quality score, the quality of that mode is found to be low, then equipment status check parameters or mode quality anomaly indicators can be output, rather than directly providing a disease diagnosis. For example, an ECG quality score approaching 0 with an excessively high ECG-related contribution may indicate an abnormal electrode contact status; a decreased PPG quality score with an excessively high PPG contribution may indicate an abnormal photoelectric contact status.

[0283] If the current sample does not meet the single-index offset criteria, but the offset magnitude continues to rise or fall within the trend observation window, it is classified as a trend offset. Trend determination can be performed using the sliding window linear regression slope or the Mann-Kendall trend test.

[0284] When using sliding window linear regression, the regression slope of each offset amplitude relative to time within the trend observation time window is used as the trend quantity.

[0285] When the absolute value of the regression slope exceeds the trend determination scale corresponding to the current scenario, and the regression slope direction is consistent across multiple consecutive update times, the current sample is classified as a trend-biased sample.

[0286] Trend offset is used to indicate a persistent drift of the current multimodal physiological state relative to the scene baseline. For example, when blood pressure estimation-related components and respiratory impedance-related components deviate continuously over a long observation window, chronic change trend indicators can be output for subsequent follow-up analysis.

[0287] If the current offset magnitude exceeds the joint offset determination scale, and the number of physiological modalities that meet the contribution significance threshold is not less than two, then the current sample will be classified as a multimodal joint offset.

[0288] The number of physiological modes that meet the contribution significance threshold is determined by comparing the modal contribution of each physiological mode with the contribution significance threshold one by one; only physiological modes whose modal contribution reaches or exceeds the contribution significance threshold are included in the count.

[0289] Multimodal joint migration indicates that multiple modalities or multiple modal coupling directions shift simultaneously. For example, when the contributions related to ECG, PPG, and blood pressure estimation increase simultaneously, abnormal migration parameters for ECG, pulse, and blood pressure coupling can be output; when the contributions related to respiratory impedance and blood pressure estimation increase simultaneously, abnormal migration parameters for respiratory-blood pressure coupling can be output. The above results are used for physiological state monitoring and risk indication, and are not directly used as disease diagnosis conclusions.

[0290] When dividing the data into multimodal joint migrations, a feature visualization trigger signal can be sent to the aforementioned migration metric step to output feature radar map distribution data. This radar map is only used to display the contribution distribution of each mode and does not perform original waveform inversion.

[0291] If the posterior confidence of the current scene is lower than a preset threshold, or if the current covariance matrix has a large distance relative to multiple scene reference points, but the modal contribution does not meet the characteristics of single-index offset, trend offset, or multimodal joint offset, then the current state is classified as scene mismatch offset. ;in, This is the scene confidence threshold.

[0292] Scene mismatch offset typically indicates that the system is in a scene transition, unrecognized activity, or state machine lag state. In this case, the system does not output an abnormal alarm conclusion, but instead feeds back the scene re-judgment trigger signal to the scene recognition step, resetting the scene posterior probability and recalculating the current scene label, thereby avoiding misjudging normal scene transitions as abnormal offsets.

[0293] If none of the above-mentioned single-index offset, trend offset, multimodal joint offset, and scene mismatch offset are triggered, and the current sample meets the quality scoring condition, scene stability condition, and offset magnitude condition, the current sample will be classified as a normal stable sample. ;in, Indicates the current offset type label.

[0294] S6.4 Normal stable sample reflux to update long-term baseline:

[0295] When the current sample is classified as a normal, stable sample, the system allows the sample to be re-entered to update the long-term baseline corresponding to the current scenario. The long-term baseline update uses a logarithmic Euclidean mean form:

[0296] ;

[0297] in, This represents the updated long-term baseline for the current scenario; This represents the long-term baseline of the current scenario before the update. Indicates the dynamic learning rate; Represents the matrix logarithm; Represents the matrix index.

[0298] Dynamic learning rate The posterior confidence and quality score can be set based on the current scenario, for example:

[0299] ;

[0300] in, As the base learning rate, when the scene confidence is high and the quality scores of each modality are good, the base point update step size is larger; when the scene confidence is low or the signal quality is poor, the base point update step size is automatically reduced. Through this update method, the long-term base point can slowly evolve with the normal recovery process of the monitored object, individual differences and long-term physiological state changes, but will not be quickly pulled off course by abnormal offset samples.

[0301] S6.5 Abnormal offset samples are written to the isolated record set:

[0302] When the current sample is classified as a single-indicator offset, trend offset, multimodal joint offset, or scenario mismatch offset, the system does not allow the sample to participate in long-term baseline updates, but instead writes it to the isolated record set:

[0303] ;

[0304] in, Represents a set of isolated records; Represents a timestamp; Represents the current covariance matrix; Indicates the offset type label; Indicates the current scene label; Represents the modal contribution vector; Indicates the offset magnitude; Represents the quality score vector; and It represents an affine invariant family.

[0305] The purpose of the isolation record set is to preserve anomalous offset samples and their context information while isolating them from the normal long-term baseline update path. In other words, samples written into the isolation record set will not enter the long-term baseline update formula, thereby preventing anomalous samples from contaminating the normal baseline in the data path.

[0306] S6.6 Lifecycle Management of Isolation Records

[0307] Due to limited end-side storage resources, the lifecycle management of isolated record sets can be achieved using anomaly sparse scoring. For newly written anomalous offset samples, their minimum manifold distance relative to existing isolated records can be calculated:

[0308] ;

[0309] in, Indicates the score for exceptional rarity; This represents the manifold distance between the current anomalous offset sample and the historical isolated sample; This represents the average distance scale of similar offset samples; To prevent small positive numbers with a denominator of zero.

[0310] When the edge-side isolated record set reaches its capacity limit, the system prioritizes retaining records with higher rarity scores and gradually removes highly repetitive low-rarity records. This allows for the preservation of more valuable anomalous offset samples within limited storage space. When the device is connected to the network or charging, high-rarity isolated records can be uploaded to a remote monitoring platform for subsequent review and offline calibration of model parameters.

[0311] S6.7 Scene Re-judgment Trigger and Visual Trigger:

[0312] When the current sample is classified as a scene mismatch offset, this step generates a scene re-judgment trigger signal and feeds it back into the scene recognition process in step S4. Upon receiving this trigger signal, the scene recognition step resets the scene posterior probability to a uniform distribution and recalculates the scene label based on the manifold distance between the current covariance matrix and each scene reference point. This mechanism handles sudden scene transitions or state machine lag issues, preventing false alarms of scene jumps as abnormal offsets.

[0313] When the current sample is classified into a multimodal joint migration, this step generates a feature visualization trigger signal, which is then fed back into the migration metric step to output radar chart distribution data of feature loads for each modality. This radar chart is used to display the multimodal contribution distribution, making it easy to see which modes or coupling directions the migration mainly involves. It should be noted that this visualization result is auxiliary information for physiological state monitoring and does not directly constitute a disease diagnosis conclusion.

[0314] Parameter initialization and implementation settings:

[0315] In one feasible parameter initialization method, the baseline event triggering threshold for each physiological modality is determined by the quantile of the signal change amplitude within the normal stable window of the corresponding physiological modality. The baseline event triggering threshold for the ECG modality can be taken as the lower quantile of the normal R-wave ascending amplitude, and the ECG refractory period can be set to 200 ms for example; the baseline event triggering threshold for the PPG modality can be taken as a preset proportion of the typical systolic ascending amplitude; the baseline event triggering thresholds for continuous blood pressure estimation, thoracic respiratory impedance, and wrist acceleration modalities are determined based on the changes in blood pressure during the cardiac cycle, the changes in inspiratory-expiratory transition impedance, and the acceleration changes at the boundary between resting and mild activity, respectively.

[0316] For example, modal disabling threshold The modal recovery threshold can be set to 0.60, the scene confidence threshold can be set to 0.75, and the hysteresis upper threshold can be set to 0.70. It can be set to 0.80, the lower hysteresis threshold. It can be set to 0.60, the threshold smoothing coefficient. It can be set between 0.10 and 0.30. The above values ​​are used to illustrate one implementation method, and the actual values ​​can be determined based on sensor noise, wearing position, and calibration data of the monitored object.

[0317] In positive definite constraints, the basic canonical strength ε0 is determined according to the mean of the trace of the mass-weighted covariance matrix. to Multiplier setting, maximum number of conditions It can be set to to Between; the rolling time window can be set to short 30 s, medium 60 s and long 300 s, and the event update weight of each time window is calculated based on the current average event interval and the corresponding time window length.

[0318] To avoid directly outputting disease diagnosis conclusions, the abnormal offset classification only outputs monitoring auxiliary parameters such as offset type, offset amplitude, modal contribution, scenario posterior confidence, and quality score; the specific disease name, treatment plan, or health status conclusion is determined by independent manual interpretation or external medical decision-making process.

[0319] To verify the effectiveness of the event-driven and Riemannian manifold-based multimodal dynamic physiological monitoring method described in this invention under different physiological scenarios, different levels of motion artifacts, and different modal quality conditions, this embodiment constructs a repeatable constructed simulation environment and follows the... Figure 1 The method flow shown executes the following steps in sequence: multimodal signal construction, event processing, event-driven covariance incremental update, scene posterior recognition, dynamic base point construction, tangent space offset measurement, and abnormal offset classification.

[0320] This embodiment constructs 200 virtual monitoring subjects, each corresponding to a 30-day continuous wearing scenario, generating approximately 144,000 hours of constructed multimodal physiological signals. Each virtual monitoring subject is configured with individualized baseline heart rate, pulse wave amplitude, pulse wave conduction time, baseline blood pressure, respiratory rate, respiratory impedance amplitude, and daily activity level, ensuring normal individual differences among different virtual monitoring subjects.

[0321] Each virtual monitoring subject includes at least the following four physiological scenarios within a day: resting and awake scenario; light activity scenario; sleep scenario; and rehabilitation exercise scenario.

[0322] In one exemplary timetable, 10 PM to 6 AM the next day is set as a sleep scenario, waking up in the morning is set as a resting and awakening scenario, getting up, washing, and daily indoor activities are set as a light activity scenario, and a preset time period in the morning is set as a rehabilitation exercise scenario. Transition segments of varying durations can be added to the boundaries of each scenario to simulate real-world transitions such as changing from a lying to a sitting position, from a sitting to walking, and returning to rest after exercise. Five physiological signals are constructed: channel 1 is an ECG signal, channel 2 is a PPG signal, channel 3 is a continuous blood pressure estimation signal, channel 4 is a thoracic respiratory impedance signal, and channel 5 is the combined amplitude of wrist three-axis acceleration. To form a continuous reference signal, the sampling rate of the ECG channel is set to 250Hz, the PPG channel to 125Hz, the continuous blood pressure estimation channel to 125Hz, the thoracic respiratory impedance channel to 25Hz, and the acceleration channel to 50Hz. The aforementioned sampling rate is only used to generate simulation reference waveforms. After entering the method of this invention, each signal is converted into an asynchronous event stream through differential modulation and is no longer transmitted to subsequent modules according to the aforementioned fixed frame rate. The PPG signal is constructed from red light and infrared dual-wavelength photoelectric signals, and their ratio can be used to form an auxiliary SpO2 estimate. Since the covariance matrix of this embodiment adopts five modes, the SpO2 estimate is not used as an independent sixth mode, but as an auxiliary state quantity or quality evaluation quantity of the PPG channel.

[0323] The continuous blood pressure estimation signal is constructed based on the pulse wave conduction time (PTT) between the R wave moment in ECG and the characteristic point of the pulse wave in PPG, and mapped to the blood pressure estimate through preset individual calibration parameters. During the simulated intermittent cuff NIBP calibration, the constructed cuff reference value is used to correct the calibration deviation between PTT and blood pressure.

[0324] To test the response of this invention to different types of abnormal offsets, this embodiment injects the following structurally abnormal segments into the normal baseline signals of four scenarios: ECG-pulse coupling abnormal offset segment; respiratory rhythm abnormal offset segment; and body position change-related blood pressure abnormal offset segment. The ECG-pulse coupling abnormal offset segment is formed by simultaneously changing the heartbeat triggering timing, QRS-related waveform characteristics, PPG pulse amplitude, and ECG-PPG time coupling relationship in the ECG. Figure 4 This is uniformly referred to as ECG-related structural abnormality offset, and the method actually outputs ECG-pulse coupling abnormality offset prompt parameters.

[0325] The respiratory rhythm deviation segment reduces or pauses the periodic changes in thoracic respiratory impedance within a preset time period, and simultaneously alters the structure of respiratory-related blood pressure fluctuations. The postural change-related blood pressure deviation segment causes a short-term decrease in continuous blood pressure estimates after the acceleration signal exhibits postural change characteristics, and simultaneously constructs compensatory changes in heart rate and PPG amplitude.

[0326] This embodiment injected a total of 8762 ECG-pulse coupling abnormality segments, 1208 respiratory rhythm abnormality segments, and 396 positional change-related blood pressure abnormality segments. These numbers are from a simulation configuration used to establish stable statistical comparisons and do not represent the actual number of patient events.

[0327] To verify robustness under motion artifact conditions, constructed motion artifacts of varying intensities were injected into the ECG, PPG, and acceleration channels. The motion artifact signal-to-noise ratio was set as follows:

[0328] In the simulation, a fixed random seed of 42 was used to generate noise parameters and event locations, and a five-fold partitioning method with independent virtual objects was used for verification. Each set of parameters was run 5 times and the average value was taken to reduce the fluctuations caused by a single random construction.

[0329] This embodiment sets three covariance time scales: short-time, medium-time, and long-time.

[0330]

[0331] Among them, short window Primarily used to capture abnormal offsets on the order of seconds, in the medium time window. Used to characterize fluctuations within a scene, long-time window Used to characterize slower baseline changes.

[0332] An exponential forgetting factor is set for each time scale based on the corresponding window length:

[0333] ;

[0334] in, The average event interval of the current event stream. .

[0335] The event-driven frontend performs differential modulation based on the basic thresholds for each modality, and then inversely modulates the thresholds using the posterior confidence of the scene. Scene switching employs a hysteresis interval consisting of an upper and lower threshold, and first-order low-pass smoothing is applied to the final threshold. The square root, negative half-power, logarithm, and exponent of the matrix are all calculated through symmetric matrix eigenvalue decomposition.

[0336] When the quality score is lower than the modal disabling threshold, the same-dimensional principal submatrix is ​​extracted from the fifth-order covariance matrix and the long-term basis points of the scene using the selection matrix, and the calculation continues in the corresponding low-dimensional SPD submanifold.

[0337] To verify the synergistic effect of the various technical aspects of this invention, this embodiment conducts simulation evaluations of the complete scheme and comparative scheme of this invention under constructed multimodal physiological signals, preset abnormal offset segments, and different motion artifact signal-to-noise ratios. The following results are all engineering verification data obtained under uniform simulation conditions and do not represent actual clinical trial results.

[0338] The sensitivity for detecting abnormal offsets is calculated using the following formula: ;

[0339] in, The number of structurally anomalous fragments correctly detected. This represents the number of undetected structurally anomalous fragments.

[0340] The false positive rate is represented by the number of abnormal offset alerts in the daily erroneous output: ;

[0341] in, (D) represents the number of abnormal offset prompts output during non-abnormal periods, and (D) represents the number of simulation monitoring days.

[0342] The equivalent terminal power consumption is estimated by the following formula: ;

[0343] in, , and These represent the number of analog-to-digital conversions, the number of operations, and the number of memory accesses within the statistical period; , and These are the equivalent energy consumption for a single operation, respectively. For the statistical period.

[0344] The end-to-end delay is the time between the first entry of an anomaly-related event into the shared timestamp bus and the output of the anomaly offset classification result.

[0345] Figure 5 The detection rate, false alarm suppression rate, power efficiency, and delay efficiency are all normalized to 0 to 1 and calculated according to the following formulas:

[0346] ;

[0347] ;

[0348] ;

[0349] ;

[0350] in, This represents the upper limit of the preset false positive rate used for normalization; The unit is mW; the unit of Delay is ms; the values ​​25 and 500 represent the upper limit of power consumption normalization and the upper limit of delay normalization in this embodiment, respectively.

[0351] Overall performance score: .

[0352] like Figure 1 As shown, the five simulated physiological signals are sequentially processed through preconditioning, quality assessment, and differential modulation eventization to form asynchronous event tuples with a unified timestamp reference. The event tuples directly drive the incremental update of the multimodal covariance matrix without performing fixed frame rate waveform reconstruction.

[0353] The covariance matrix after positive definite regularization is used as the observation point on the SPD manifold. The manifold distance is compared with the long-term reference points of four scenarios: resting and awake, light activity, sleep, and rehabilitation exercise, to obtain the posterior confidence of the scenario. This confidence is used to determine the current scenario and, after hysteresis and smoothing, to inversely adjust the event triggering thresholds of each modality.

[0354] like Figure 2 As shown, when the scene is stable, the current covariance matrix fluctuates around the corresponding long-term base point; when the scene changes, the dynamic base point moves continuously along the geodesic between the old scene base point and the new scene base point to reduce artificial offset spikes caused by hard switching of base points.

[0355] like Figure 3As shown, the event-driven front-end, scene recognition, and anomaly classification form an interlocking closed loop. Normal samples that meet the stability window criterion are used to update the long-term baseline of the corresponding scene; abnormal offset samples and low-quality samples are written into the isolated record set and do not participate in the long-term baseline update.

[0356] Figure 4 Comparison curves of the sensitivity of ECG-pulse coupling abnormal offset detection under different motion artifact signal-to-noise ratios.

[0357] At an SNR of 5 dB, the detection sensitivity of the complete scheme of this invention is 94.6%, and that of the MM-CNN-Transformer scheme is 84.1%; at an SNR of −5 dB, the sensitivity is 87.2% and 72.5%, respectively. This indicates that the detection sensitivity of the complete scheme of this invention decreases only slightly as motion artifacts increase.

[0358] The detection sensitivity of the non-reverse modulation scheme of this invention is 87.9% at an SNR of 5dB, which is lower than that of the complete scheme, indicating that the scene posterior confidence inverse adjustment of the event threshold helps to maintain effective event capture capability in complex scenes.

[0359] Figure 5 The normalized performance of each scheme in terms of detection rate, false alarm suppression rate, power efficiency, and delay efficiency is shown.

[0360] The four indicators of the complete solution of this invention are as follows: ;

[0361] Its overall performance score is: The equivalent end-side power consumption is 3.8mW, and the end-to-end latency is 42ms. The equivalent power consumption of the MM-CNN-Transformer scheme is 21.5mW, and the end-to-end latency is 165ms. This shows that the present invention, under the unified estimation conditions of this embodiment, combines high detection performance, low equivalent power consumption, and short processing latency.

[0362] like Figure 6 As shown, the complete solution of this invention has an FPR range of 0.21 to 0.52 times / day in four scenarios, with an average of 0.35 times / day across scenarios; the MM-CNN-Transformer solution has an FPR range of 0.86 to 3.42 times / day, with an average of 1.92 times / day.

[0363] Compared to the MM-CNN-Transformer approach, the complete approach of this invention reduces the average FPR across scenes by approximately 81.8%. This result indicates that scene-independent long-term baselines and dynamic baselines during scene transitions can reduce false alarms caused by normal physiological fluctuations in different scenes.

[0364] Figure 7The characteristic load distribution of the five physiological modalities is shown. In the preset ECG-pulse coupling abnormal offset segment, the normalized loads of the complete scheme of this invention on the ECG and PPG axes are 0.92 and 0.88, respectively, and the loads on the respiratory impedance and acceleration axes are 0.05 and 0.02, respectively.

[0365] Decoupling purity is defined as: Based on the above load data: The results show that tangent space orientation decomposition can concentrate the offset loads onto modes associated with structural anomalies and reduce their diffusion into irrelevant modes. Figure 7 Used to illustrate the interpretable distribution of characteristic loads, not to represent a direct output of disease diagnosis conclusions.

[0366] like Figure 8 As shown, the average FPR across scenarios for the basic fixed threshold and Euclidean scalar aggregation scheme is 4.15 times / day.

[0367] After adding SPD covariance representation and affine invariant manifold measure, FPR decreased to 1.23 times / day; after further adding scenario-independent long-term basis points and dynamic basis points, FPR decreased to 0.76 times / day; after adding scenario posterior confidence inverse modulation closed loop, FPR further decreased to 0.35 times / day.

[0368] The above ablation results indicate that SPD manifold characterization, scenario-based base points, dynamic base point geodesic transition, and event threshold inverse modulation all contribute to reducing false alarms across scenarios, and that these technical aspects form a synergistic relationship.

[0369] Under the constructive simulation conditions of this embodiment, the detection sensitivities of the complete scheme of the present invention are 94.6% and 87.2% when the SNR is 5dB and −5dB, respectively, the average FPR across scenarios is 0.35 times / day, the equivalent end-side power consumption is 3.8mW, the end-to-end delay is 42ms, and the modal characteristic load decoupling purity is 1.73.

[0370] The above results demonstrate that this invention can reduce redundant data processing through an event-driven front-end, preserve the multimodal coupling structure through event-driven covariance updates, reduce false alarms during scene switching through long-term and dynamic baselines, and prevent long-term baselines from being contaminated by anomalous data through normal sample backflow and abnormal sample isolation. The listed values ​​are all constructive simulation results, and their values ​​will vary depending on the signal model, noise model, parameter settings, and energy consumption estimation model.

[0371] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal dynamic physiological monitoring method based on event-driven and Riemannian manifolds, characterized in that, Includes the following steps: S1: Collect multiple physiological signals from the monitored object, pre-condition and quality assessment of each physiological signal, and obtain the conditioning signal and the corresponding quality score; S2 performs differential modulation event processing on each conditioning signal and uniformly stamps the trigger events based on the shared clock. When the difference obtained from the differential modulation event processing reaches the event trigger threshold of the corresponding physiological mode, an asynchronous event tuple is generated. S3, based on the asynchronous event tuple and quality score, perform event-driven incremental update and positive definite constraint on the covariance matrix of the multi-path physiological signals to obtain the current covariance matrix on the symmetric positive definite manifold; S4, calculate the manifold distance between the current covariance matrix and the long-term reference points corresponding to different physiological scenarios, the long-term reference points Indicates the first The long-term normal reference center of the multimodal covariance matrix under each physiological scenario is determined, the current scenario and its posterior confidence are determined, and the event triggering threshold in step S2 is adjusted in reverse according to the posterior confidence. S5. Construct a dynamic base point for manifold mapping based on the posterior confidence level, and perform tangent space mapping on the current covariance matrix with the dynamic base point as a reference to obtain the offset magnitude and modal contribution of each physiological mode. S6. Based on the offset amplitude, modal contribution, posterior confidence and quality score, the current physiological state is classified as abnormal. Normal samples that meet the stability conditions are used to update the long-term baseline of the corresponding physiological scenario. Abnormal offset samples are written into the isolation record set and are prohibited from participating in the update of the long-term baseline.

2. The multimodal dynamic physiological monitoring method according to claim 1, characterized in that, Step S1 includes: The multiple physiological signals include electrocardiogram signals, photoplethysmography pulse wave signals, continuous blood pressure estimation signals, thoracic respiratory impedance signals, and wrist acceleration signals. The photoplethysmography (PPG) signal is acquired by the red light photoelectric channel and the infrared photoelectric channel, and the blood oxygen saturation estimate is obtained based on the ratio of red light intensity to infrared light intensity. The blood oxygen saturation estimate is used as an auxiliary feature or quality assessment metric of the PPG mode. The continuous blood pressure estimation signal is obtained based on the pulse wave conduction time between the R wave in the electrocardiogram signal and the pulse wave characteristic point in the photoplethysmography signal, and is periodically calibrated based on the results of intermittent cuff blood pressure measurement. Each physiological signal is subjected to filtering processing adapted to the corresponding signal frequency band, and a corresponding quality score is generated based on the short-time signal-to-noise ratio, signal saturation ratio and sensor contact state. The quality score is used for covariance weighting and modal disability judgment in step S3, and for normal and stable sample determination in step S6.

3. The multimodal dynamic physiological monitoring method according to claim 2, characterized in that, Step S2 includes: For each conditioning signal, maintain the corresponding reference level and continuously compare the difference between the current conditioning signal and the reference level after the previous asynchronous event was triggered; When the difference reaches the event trigger threshold of the corresponding physiological mode, and the refractory period of the corresponding physiological mode is exceeded from the time of the previous asynchronous event trigger, an asynchronous event tuple containing the channel index, event trigger timestamp, event polarity and quality score is generated, and the corresponding reference level is updated stepwise according to the event polarity and event trigger threshold. The event trigger pulses of each physiological modality are connected to the same shared reference clock bus. The arbitration unit stamps the events in the order of arrival and compensates the event trigger timestamps of each asynchronous event tuple according to the pre-stored channel delay calibration value. Within the rolling time window, new asynchronous events are added according to the event trigger timestamp, and asynchronous events exceeding the rolling time window are cleared to form an asynchronous event tuple set. The asynchronous event tuple set and quality score are directly delivered to step S3, and the asynchronous event tuple is not reconstructed into a fixed frame rate time domain waveform between steps S2 and S3.

4. The multimodal dynamic physiological monitoring method according to claim 3, characterized in that, Step S3 includes: A reference level vector is constructed based on the current reference level of each physiological mode; when the k-th physiological mode... When an asynchronous event arrives, the product of the event polarity and the event trigger threshold is determined as the reference level step, and the covariance increment matrix is ​​calculated based on the reference level step. ； The covariance matrix is ​​updated incrementally using an event-driven method as follows: ； in, Indicates the first The first physiological modality The covariance increment matrix formed by asynchronous events; This represents the reference level step caused by the asynchronous event; Indicates the first 1 standard basis vector; This represents the reference level vector before the arrival of this asynchronous event; This represents the weighted average vector before the arrival of this asynchronous event; This represents the covariance matrix before the arrival of this asynchronous event; This represents the covariance matrix after event-driven incremental updates have been completed but before quality weighting and positive definite constraints have been applied. This indicates the event update weight corresponding to the length of the scrolling time window; The covariance increment matrix is ​​only in the first... line, number List and No. The diagonal elements generate an increment, so that the covariance sub-blocks corresponding to the remaining physiological modes remain unchanged when each asynchronous event arrives; Based on the quality scores of each physiological modality, the covariance matrix that has completed the event-driven incremental update is weighted accordingly, and positive definite regularization terms are adaptively added according to the condition number of the quality-weighted covariance matrix to obtain the current strictly symmetric positive definite covariance matrix. When the quality score of any physiological mode is lower than the mode disabling threshold, extract the same-dimensional principal submatrix corresponding to the physiological mode whose quality score is not lower than the mode disabling threshold from the current covariance matrix and the long-term basis points corresponding to each physiological scenario, and continue to execute steps S4 to S6 on the low-dimensional symmetric positive definite manifold formed by the same-dimensional principal submatrix.

5. The multimodal dynamic physiological monitoring method according to claim 4, characterized in that, Step S4 includes: Calculate the affine invariant Riemann distance between the current covariance matrix and the long-term baseline corresponding to each physiological scene, and determine the scene likelihood of the current covariance matrix corresponding to each physiological scene based on each affine invariant Riemann distance; The scene likelihood of each physiological scene, the posterior confidence of the scene at the previous moment, and the scene transition matrix are fused by Bayesian method to obtain the posterior confidence of the scene corresponding to each physiological scene at the current moment, and the posterior confidence of the scene is smoothed by low-pass method. The physiological scene with the highest posterior confidence after smoothing is determined as the current scene. Based on the current scene, the posterior confidence of the smoothed scene corresponding to the current scene, the basic threshold of each physiological modality, and the corresponding scene threshold factor, the feedback threshold of each physiological modality at the next moment is determined. The feedback threshold is subjected to low-pass smoothing and upper and lower limit constraints, and the constrained feedback threshold is used as the event trigger threshold for the next moment of the corresponding physiological mode in step S2. When the posterior confidence of the smoothed scene corresponding to the current scene exceeds the upper hysteresis threshold, the corresponding physiological mode is allowed to enter the preset threshold modulation state; when the posterior confidence of the smoothed scene is lower than the lower hysteresis threshold, the corresponding physiological mode is made to exit the preset threshold modulation state; when the posterior confidence of the smoothed scene is between the upper hysteresis threshold and the lower hysteresis threshold, the current threshold modulation state remains unchanged.

6. The multimodal dynamic physiological monitoring method according to claim 5, characterized in that: The physiological scenarios include at least a resting and waking scenario, a light activity scenario, a sleep scenario, and a rehabilitation exercise scenario, with each physiological scenario corresponding to an independently maintained long-term baseline. The scene transition matrix is ​​initialized based on the reachability order, time phase, and preset activity plan between each physiological scene, so that the prior probability of scene transition that does not conform to the preset activity order is lower than the prior probability of scene transition between adjacent physiological scenes. When the posterior confidence level of the sleep scenario reaches a high confidence level, the event trigger threshold of the thoracic respiratory impedance mode is reduced and the event trigger threshold of the wrist acceleration mode is increased; when the posterior confidence level of the rehabilitation exercise scenario reaches a high confidence level, the event trigger threshold of the thoracic respiratory impedance mode is increased and the event trigger threshold of the wrist acceleration mode is reduced. When the abnormal offset classification result output in step S6 indicates that the current covariance matrix does not match the long-term baseline of all known physiological scenarios, the posterior confidence of each physiological scenario is reset and the scenario classification is re-executed. At the same time, the event trigger threshold modulation state of each physiological modality is restored to the preset initial state.

7. The multimodal dynamic physiological monitoring method according to claim 6, characterized in that, The construction of dynamic base points in step S5 includes: When the current physiological scenario is in a stable state, the long-term baseline corresponding to the current physiological scenario is determined as the dynamic baseline. When the monitored object transitions from the pre-switch physiological scene to the post-switch physiological scene, the geodesic interpolation weight is determined based on the posterior confidence of the pre-switch physiological scene and the posterior confidence of the post-switch physiological scene. Based on the geodesic interpolation weight, the dynamic base point is continuously moved along the affine invariant geodesic between the long-term base point of the pre-switch physiological scene and the long-term base point of the post-switch physiological scene. As the posterior confidence of the physiological scene after the switch increases, the dynamic reference point gradually approaches the long-term reference point of the physiological scene after the switch; as the posterior confidence of the physiological scene after the switch decreases, the dynamic reference point remains close to the long-term reference point of the physiological scene before the switch, so as to avoid the mapping reference point from jumping directly and forming a non-physiological offset peak when the physiological scene is switched.

8. The multimodal dynamic physiological monitoring method according to claim 7, characterized in that, Step S5, which involves tangent space mapping of the current covariance matrix, includes: The current covariance matrix is ​​whitened with reference to the dynamic base point, and the current covariance matrix after whitening is subjected to matrix logarithm operation to obtain the tangent space offset matrix with the dynamic base point as the origin. The offset magnitude of the current multimodal physiological state relative to the dynamic base point is determined based on the Frobenius norm of the tangent space offset matrix; Based on the diagonal and off-diagonal elements corresponding to each physiological mode in the tangent space offset matrix, the self-fluctuation term of each physiological mode and the coupling fluctuation term between each physiological mode and other physiological modes are counted to obtain the modal contribution of each physiological mode. The tangent space offset matrix is ​​decomposed into eigendirections, and the dominant mode coupling direction involved in the current offset is determined based on the loads of the dominant eigendirections on five physiological modes: electrocardiogram, photoplethysmography pulse wave, continuous blood pressure estimation, thoracic respiratory impedance, and wrist acceleration. The offset magnitude, modal contribution, and dominant modal coupling direction are all obtained from the tangent space offset matrix, and the original time-domain waveform is not inverted based on the current covariance matrix.

9. The multimodal dynamic physiological monitoring method according to claim 8, characterized in that, The abnormal offset classification in step S6 includes: The offset judgment scale corresponding to the current scene is determined based on the local fluctuation bandwidth of the current scene, the posterior confidence of the current scene, and the lowest quality score among the physiological signals. When the offset magnitude exceeds the offset determination scale, and the modal contribution is concentrated in a single physiological modality or a single modal coupling direction, the current sample is classified as a single index offset. When the offset amplitude changes unidirectionally continuously within the preset trend observation time window and exceeds the trend judgment scale corresponding to the current scene, the current sample is classified as a trend offset. When the offset magnitude exceeds the joint offset determination scale corresponding to the current scene, and the modal contribution of at least two physiological modalities exceeds the contribution significance threshold, the current sample is classified as a multimodal joint offset. When the posterior confidence of the current scene is lower than the scene confidence threshold, and the current sample does not meet the judgment conditions of single index offset, trend offset or multimodal joint offset, the current sample is classified as scene mismatch offset, and a scene re-judgment trigger signal is output to step S4. When the criteria for single-index offset, trend offset, multimodal joint offset, and scene mismatch offset are not met, and the scene posterior confidence and the quality scores of each physiological signal corresponding to the current scene reach the corresponding stability threshold, the current sample is classified as a normal stable sample.

10. The multimodal dynamic physiological monitoring method according to claim 9, characterized in that, Step S6 also includes: For the normal stable sample, the current covariance matrix corresponding to the normal stable sample is fused to the long-term base point corresponding to the current scene using the logarithmic Euclidean mean method. The base point update step size is determined according to the scene posterior confidence and the quality score of each physiological signal corresponding to the current scene, so that the base point update step size decreases as the scene posterior confidence or quality score decreases. For the single-index offset, trend offset, multimodal joint offset, and scenario mismatch offset, the corresponding current covariance matrix, timestamp, current scenario, offset type, offset magnitude, modal contribution, and quality score are written into the isolated record set, and abnormal offset samples written into the isolated record set are excluded during the long-term baseline update process. When the isolated record set reaches the preset capacity, the rarity of the newly written abnormal offset sample is determined according to the manifold distance between the newly written abnormal offset sample and the historical abnormal offset sample. The abnormal offset samples with higher rarity are retained first, and the abnormal offset samples with high repetition with the retained abnormal offset samples are deleted. When the quality score of the disabled physiological modality recovers to above the modality recovery threshold and continues to reach the preset recovery time, the current covariance matrix corresponding to the complete physiological modality and the long-term base points of each physiological scene are reactivated, and scene recognition, dynamic base point construction and abnormal offset classification are continued.

Images