Multimodal cross-calibrated continuous multi-metabolite monitoring method

Abstract

This invention relates to the field of metabolite monitoring technology and discloses a multimodal cross-calibrated method for continuous multi-metabolite monitoring. The method acquires multi-source metabolite signal sequences containing electrochemical, optical, and thermodynamic signals through an implanted biosensor array; it performs time-series segmentation of the multi-source metabolite signal sequences to generate metabolite concentration time-series segments; and it synchronously inputs these metabolite concentration time-series segments into a metabolite dynamic response model and a metabolite interaction feature model to generate a first metabolite concentration trajectory and a second metabolite concentration trajectory, respectively. This method, by leveraging multimodal signal fusion and cross-calibration, combined with time-series segmentation and multi-model analysis, can comprehensively capture the dynamic changes and interactions of metabolites, improving the completeness and accuracy of metabolite monitoring, and is suitable for continuous multi-metabolite monitoring scenarios.

Description

Technical Field

[0001] This invention relates to the field of metabolite monitoring technology, specifically a continuous multi-metabolite monitoring method with multimodal cross-calibration. Background Technology

[0002] In the field of modern health monitoring, continuous and precise monitoring of various metabolites in the human body is an important means of understanding physiological states and disease development. As intermediate or final products of chemical reactions in organisms, the concentration changes of metabolites are closely related to energy metabolism, organ function, and pathological states. For example, the dynamic changes of metabolites such as glucose, lactic acid, and ketone bodies can reflect various health problems such as diabetes and abnormal liver function.

[0003] Traditional methods for monitoring metabolites often rely on invasive, single-point sampling, such as blood tests. These methods not only fail to provide continuous monitoring but also cause patients frequent pain and increase the risk of infection. To overcome this limitation, various biosensors have emerged, among which implantable sensors, due to their direct contact with the bodily fluid environment, demonstrate advantages in continuous monitoring. However, single-type sensors are often limited by their own principles and are susceptible to interference from the biological environment, leading to insufficient stability of monitoring results. For example, electrochemical sensors are easily interfered with by other electroactive substances in bodily fluids, optical sensors experience signal drift due to changes in tissue scattering characteristics, and thermodynamic sensors are sensitive to local temperature fluctuations.

[0004] Metabolites in the human body do not exist in isolation; complex dynamic interactions exist between different metabolites, such as the regulatory feedback between glucose and insulin, and the metabolic linkage between lactate and pyruvate. Current monitoring technologies mostly analyze individual metabolites independently, neglecting the interactions between them and failing to comprehensively reflect the body's overall metabolic state. Furthermore, metabolite concentrations change dynamically over time, influenced by various factors such as diet and exercise. Extracting effective concentration trajectories from continuous signals remains a challenge for current technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal cross-calibrated continuous multi-metabolite monitoring method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for continuous multi-metabolite monitoring with multimodal cross-calibration, the method comprising:

[0007] Multi-source metabolite signal sequences are acquired using an implantable biosensor array, the multi-source metabolite signal sequences including electrochemical signals, optical signals and thermodynamic signals;

[0008] The multi-source metabolite signal sequence is time-series segmented to generate metabolite concentration time segments;

[0009] The time-series segments of metabolite concentrations are synchronously input into the metabolite dynamic response model and the metabolite interaction feature model to generate the first metabolite concentration trajectory and the second metabolite concentration trajectory, respectively.

[0010] Preferably, when constructing the metabolite dynamic response model, the metabolite migration path is discretized using biomembrane diffusion rate and enzyme reaction threshold as constraints;

[0011] Based on the metabolite migration path, adjacent path nodes are associated through a dynamic damping coefficient to generate a metabolite concentration gradient field.

[0012] The concentration inversion of the multi-source metabolite signal sequence is performed based on the metabolite concentration gradient field, and the concentration trajectory of the first metabolite is output.

[0013] Preferably, when constructing the metabolite interaction feature model, the phase aggregation and polarity effect features of the multi-source metabolite signal sequences are extracted;

[0014] The co-shift of the phase clustering and polarity effect features is quantified using a variation analysis model;

[0015] The competitive inhibition relationship of metabolites is dynamically weighted based on the cooperative offset, and the concentration trajectory of the second metabolite is output.

[0016] Preferably, cross-model calibration is performed on the first metabolite concentration trajectory and the second metabolite concentration trajectory, including:

[0017] Calculate the discrete difference sequence between the concentration trajectory of the first metabolite and the concentration trajectory of the second metabolite;

[0018] The discrete difference sequence is input into the trajectory convergence judgment unit to generate the trajectory offset weight factor;

[0019] When the trajectory offset weighting factor exceeds the preset convergence threshold, a metabolite signal resampling command is triggered.

[0020] Preferably, the method further includes:

[0021] In response to the metabolite signal resampling command, the multi-source metabolite signal sequence is subjected to frequency domain noise reduction processing to generate a noise-reduced metabolite signal segment.

[0022] The noise-reduced metabolite signal fragments are fed back to the metabolite dynamic response model and the metabolite interaction feature model to update the first metabolite concentration trajectory and the second metabolite concentration trajectory.

[0023] Preferably, the method further includes:

[0024] Based on the first metabolite concentration trajectory and the second metabolite concentration trajectory, a metabolite feature recombination matrix is ​​constructed;

[0025] Principal component decomposition was performed on the metabolite feature recombination matrix to extract the core metabolite modal feature vectors;

[0026] A fusion map of metabolite concentrations is generated based on the feature vectors of the core metabolites modalities.

[0027] Preferably, the method further includes:

[0028] A tolerance range for metabolite concentration fluctuations is set, and the fusion profile of the metabolite concentrations is dynamically screened.

[0029] When the concentration values ​​of consecutive time points in the metabolite concentration fusion map exceed the tolerance range of metabolite concentration fluctuations, an abnormal metabolite mode is marked.

[0030] The frequency domain divergence and time domain clustering features of the abnormal metabolite modes are extracted to generate a metabolite abnormal feature set.

[0031] Preferably, the method further includes:

[0032] The abnormal metabolite feature set is input into the metabolite mechanism matching model, which includes quantitative rules for substrate inhibition, product feedback inhibition, and covalent modification.

[0033] The matching degree between the metabolite abnormality feature set and each quantification rule is calculated using the least squares approximation algorithm, and the metabolite abnormality type identifier is output.

[0034] Preferably, the method further includes:

[0035] Based on the metabolite abnormality type identifier, retrieve the corresponding metabolite correction strategy library;

[0036] Based on the metabolite correction strategy library, metabolite concentration compensation coefficients are generated to dynamically correct the metabolite concentration fusion map.

[0037] Preferably, the stability of the dynamically corrected metabolite concentration fusion profile is verified, including:

[0038] Calculate the coefficient of variation and autocorrelation decay rate of the corrected concentration sequence;

[0039] When the coefficient of variation is lower than a preset stability threshold and the autocorrelation decay rate conforms to the exponential decay model, the calibrated continuous multi-metabolite monitoring results are output.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] By acquiring multi-source metabolite signal sequences containing electrochemical, optical, and thermodynamic signals through implantable biosensor arrays, it is possible to capture characteristic information of metabolites from multiple dimensions. Different types of signals have their own response characteristics when reflecting changes in metabolite concentration. The fusion of multi-source signals can cover information levels that a single signal cannot reach, reducing information loss caused by the simplification of signals.

[0042] Temporal segmentation of multi-source metabolite signal sequences to generate metabolite concentration time-series fragments helps decompose continuous dynamic signals into units with specific temporal characteristics, enabling subsequent analysis to more accurately focus on metabolite changes over different time periods. This temporal segmentation method can accommodate fluctuations in metabolite concentrations at different time scales, avoiding the loss of detail that may occur when processing the entire signal in a general way.

[0043] By simultaneously inputting time-series data of metabolite concentrations into both the metabolite dynamic response model and the metabolite interaction feature model, first and second metabolite concentration trajectories are generated, respectively, enabling multi-perspective analysis of metabolite dynamic changes. The dynamic response model focuses on capturing the direct trends of metabolite concentration changes over time, while the interaction feature model can uncover the correlations between different metabolites. The combination of the two models provides a more comprehensive view of the overall metabolic process of metabolites.

[0044] The cross-calibration mechanism for multimodal signals can correct for each other's deviations by utilizing the complementarity between different signals, reducing the impact of various interference factors in the biological environment on a single signal. This cross-validation method makes the final metabolite concentration trajectory closer to the actual metabolic state, reducing monitoring bias caused by interference from a single signal. Attached Figure Description

[0045] Figure 1 This is a time-series diagram of the multimodal cross-calibrated continuous multi-metabolite monitoring method described in this invention;

[0046] Figure 2 A flowchart for constructing a dynamic response model of metabolites and generating their trajectories;

[0047] Figure 3 The flowchart shows the cross-model calibration and resampling triggering process for dual trajectories.

[0048] Figure 4 This is a flowchart for metabolite concentration anomaly detection and feature extraction.

[0049] Figure 5 A flowchart for matching metabolite abnormality types. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figure 1 This invention provides a method for continuous multi-metabolite monitoring with multimodal cross-calibration, the method comprising:

[0052] High-precision acquisition and fusion of multi-source metabolite signals are achieved through an implantable biosensor array. The biosensor array comprises electrochemical sensing units, optical sensing units, and thermodynamic sensing units, each capturing metabolite signals with different physical properties. The electrochemical sensing unit detects metabolite concentration based on the current signal generated by redox reactions; the optical sensing unit reflects metabolite levels through changes in fluorescence intensity or absorption spectra; and the thermodynamic sensing unit measures metabolite activity based on the thermal effect of enzymatic reactions. After analog-to-digital conversion, the multi-source signals generate time-stamp-aligned electrochemical, optical, and thermodynamic signal sequences, constituting the multi-source metabolite signal sequence.

[0053] The time-series segmentation module employs a sliding window algorithm to dynamically divide multi-source metabolite signal sequences. The window length is 5 minutes with a 50% overlap rate, generating time-series segments of metabolite concentrations that include time-dimensional indices. Each time-series segment is simultaneously input into a metabolite dynamic response model and a metabolite interaction feature model. The metabolite dynamic response model generates a first metabolite concentration trajectory based on biophysical mechanisms, while the metabolite interaction feature model outputs a second metabolite concentration trajectory through statistical learning. The trajectory data output by both models carry the same timestamp, providing a time-series alignment basis for subsequent cross-calibration.

[0054] Example 1: See Figure 2 The core component of this invention, the metabolite dynamic response model, begins with biophysical mechanism modeling. Biomembrane diffusion kinetics are expressed through an unsteady-state mass transfer equation, with the diffusion rate parameter determined by Fick's second law and related to the molecular weight of the specific metabolite and membrane permeability. The enzyme reaction threshold is derived based on the steady-state solution of the Michaelis-Menten equation, and its value is dynamically adjusted with substrate affinity. A non-uniform mesh partitioning strategy is used to discretize metabolite migration paths, with 100 to 200 computational nodes set along the biomembrane thickness direction, and the node density is automatically increased in the enzyme activity reaction zone. The dynamic damping coefficient between adjacent nodes is proportional to the local concentration gradient, and the calculation process uses an implicit Euler method for iterative solution, forming a three-dimensional concentration gradient field with spatiotemporal continuity.

[0055] The multi-source signal inversion stage employs a regularized optimization framework: electrochemical signals are used as the primary observed variable for concentration inversion; the fluorescence intensity of optical signals is converted to equivalent concentration after Stokes shift calibration; and the concentration value of thermodynamic signals is derived based on the linear relationship between enthalpy change and temperature change. By constructing a loss function with gradient field constraints, and using the damping coefficient as a weight to fuse the three signal sources, the inversion algorithm converges to a preset residual threshold within 20 iterations using the conjugate gradient method. The generated first metabolite concentration trajectory is stored in minutes, and the trajectory data is stored as a sequence of timestamp-concentration value pairs.

[0056] The implementation of the metabolite interaction characteristic model begins with feature separation of multi-source signals. Time-series segments of electrochemical, optical, and thermodynamic signals are input into three independent data channels, each configured with a Butterworth filter bank with a specific frequency response. The filtered signal segments undergo zero-phase drift correction to eliminate phase distortion caused by time delay. A Hilbert transform is applied to the corrected signal to generate a two-dimensional matrix containing the instantaneous phase angle. Phase clustering is calculated using a sliding window mechanism, expanding two data points forward and backward from the current sampling point to form a quintuple. The phase change slope is fitted using the least squares method, and the reciprocal of this slope is stored as a phase clustering index in a circular buffer.

[0057] The polarity effect identification process is parallel to phase analysis. The raw signal, before phase processing, is standardized and then input into the Gaussian mixture model training module. Initialization uses a clustering seed selection method based on the signal amplitude histogram, prioritizing distribution extrema as initial cluster centers. An early stopping mechanism is incorporated into the expectation-maximization algorithm, terminating training when the likelihood function change rate falls below 0.01% for three consecutive iterations. The positive and negative polarity boundaries are dynamically determined by the intersection points of the Gaussian components, and the KL divergence value of the bimodal distribution is calculated simultaneously. The polarity distribution histogram is updated every five minutes, maintaining a moving window dataset containing the latest 200 data points.

[0058] The variation analysis module receives the processing results from the two aforementioned paths. The four-dimensional feature vector is synchronously generated via the timestamp alignment module, including: phase clustering difference (difference between the current value and the mean of the previous minute), polarity boundary displacement (percentage deviation of the current boundary position from the historical mean), amplitude kurtosis rate of change (ratio of the current window to the previous window), and skewness extreme value offset (absolute difference between the current skewness and the baseline). The covariance matrix is ​​updated using an exponentially weighted moving average method, assigning higher weight to recent data. The Mahalanobis distance calculation process includes condition number monitoring; when the condition number of the feature matrix exceeds 10^4, Tikhonov regularization is automatically triggered. The resulting covariance offset is transformed to the 0-1 interval using a hyperbolic tangent function, with the saturation parameter of the transformation function dynamically adjusted based on the metabolite molecular weight.

[0059] The competitive inhibition relationship modeling employs a hierarchical weighted structure. The base weight coefficients are negatively correlated with metabolite concentrations, achieved through a two-parameter logistic function mapping. The inflection point of the logistic function is set to the known competition constant kc value for the metabolite, with a fixed steepness coefficient of 0.05. A co-adjustment factor, acting as a dynamic adjustment factor, multiplicatively corrects the base weights. When the offset exceeds 0.7, an inhibition intensity multiplication mechanism is activated, and the system automatically creates a new weight calibration branch. This branch optimizes the logistic function parameter configuration by analyzing the competitive inhibition relationship patterns over the previous 24 hours. The weight calculation cycle is consistent with the original signal acquisition rate, generating a weight coefficient vector every minute, with the vector dimension equal to the number of metabolite species being measured.

[0060] The feature model output generation incorporates multiple protection mechanisms. When phase clustering analysis detects drastic fluctuations (standard deviation increase > 300%) for three consecutive cycles, it switches to emergency processing mode: freezing the current parameter configuration and recalculating using backup data from the previous stable state. A dynamic buffer is set up during polarity boundary determination; when boundary drift exceeds 120% of the historical maximum displacement, the system uses the previous cycle's valid boundary to replace the current calculation result. Range verification is added during the transmission of competition suppression weights; values ​​exceeding the reasonable biochemical range are automatically truncated to the boundary value. Each frame of output data is appended with an integrity marker, including: a four-dimensional feature vector hash value, a calculation interruption flag, a boundary threshold alarm status, and a dynamic weight checksum.

[0061] The model parameter self-optimization system runs continuously in the background. A feedback loop between feature parameters and weight coefficients is established: when the competitive suppression weight remains high (>0.85) for two consecutive hours, phase clustering analysis parameter retuning is triggered. The retuning process uses a particle swarm optimization algorithm, with the smoothness of the suppression weights as the optimization objective. Every day at midnight, the system initiates global consistency verification, comparing the deviation distribution between the current feature model output and the mass spectrometry calibration data. When the deviation exceeds a preset threshold, the iterative step size is adaptively reduced. Cross-calibration between feature channels employs a delay compensation design, aligning data streams with different processing delays through an interpolation algorithm, with interpolation errors controlled within 1% of the sampling interval.

[0062] Adaptive signal drift handling includes dedicated compensation channels. The optical signal channel employs a baseline recovery algorithm, identifying and eliminating slowly varying drift components through morphological filtering. The electrochemical signal channel implements voltammetric self-calibration, injecting a standard reference potential every eight hours to reset the zero point. The thermodynamic signal channel runs a temperature compensation program, constructing a quadratic regression model based on the original thermopile voltage and temperature sensor readings. Timing records of all compensation operations are embedded in the output data frame, forming a traceable parameter correction log.

[0063] Biologically specific compensation mechanisms are implemented differently based on metabolite categories. Heme-containing molecules are configured with light absorption compensation factors, and the optimal compensation coefficient is determined through multi-wavelength scanning. Charged molecules are loaded with electric field interference correction terms, and their intensity is used to establish a two-dimensional lookup table with respect to the electrochemical detection voltage. Molecules with allosteric effects trigger dynamic phase calibration, and the Hilbert transform window size is optimized using sliding correlation analysis. Compensation parameters are stored in a biological characteristic parameter library; the default configuration is loaded each time the system starts, and dynamic fine-tuning is performed during operation based on performance indicators, with the fine-tuning range limited to ±15% of the initial value.

[0064] The time-varying response module monitors changes in metabolic rate. The system continuously calculates the second derivative of concentration change per unit time. When acceleration exceeds 90% of historical peak values, it automatically shortens the feature statistics window to 60% of the original timescale. The signal reconstruction mechanism is activated upon detecting abnormal noise patterns, reconstructing damaged channel data using other sensor signals for the same metabolite. Model updates employ an incremental deployment strategy: new parameter configurations are first validated in shadow mode, and only after six consecutive hours of output data quality meeting standards are they switched to the production environment. All algorithm components are equipped with resource usage monitoring. When computational latency exceeds real-time requirements, a degraded operation mode is activated, temporarily shutting down secondary auxiliary computational processes that do not affect core functions.

[0065] Example 2: See Figure 3 The study focuses on cross-model calibration and dynamic feedback mechanisms, with its core being the establishment of a differential quantification system and adaptive correction process for two sets of metabolite concentration trajectories. The calculation of discrete difference sequences employs an asymmetric time window design: the forward window width is fixed at 15 minutes from the current time, while the backward window width is dynamically adjusted based on the metabolite half-life, with a minimum of 5 minutes. The difference operation introduces a weight decay factor, where data points further from the current time have lower weights, specifically implemented through an exponential decay function.

[0066]

[0067] Where: ΔC i T represents the cumulative difference at the i-th time point. 1,i+k and T 2,i+k w represents the concentration values ​​of the first and second metabolite concentration trajectories at time i+k, respectively. k Let m and n be the time decay weighting coefficients, and let w be the radii of the backward and forward windows, respectively. kThe configuration follows the principle of larger values ​​for closer samples and smaller values ​​for farther samples. The weight of the current time (k=0) is set to 1.0, and it decreases by 0.15 for each sampling interval further away, with a minimum cutoff value of 0.3. The values ​​of the window radius m and n are related to the type of metabolite. For metabolites with a half-life of less than 30 minutes, m takes 2 sampling points and n takes 4 sampling points; for metabolites with a longer half-life, a configuration scheme of m=4 and n=8 is used.

[0068] The trajectory convergence judgment unit adopts a three-level decision-making mechanism. The primary judgment is based on a simple threshold method, where three consecutive ΔC values ​​are considered convergent. i Values ​​exceeding five standard deviations of the baseline noise level are marked as suspicious offsets. Intermediate judgment initiates trend analysis, calculating the first and second derivatives of the difference sequence within the suspicious period. If both have the same sign and their absolute values ​​show an increasing trend, it is determined to be a potential divergence state. Advanced judgment invokes historical pattern matching, performing Dynamic Time Warping (DTW) distance calculation between the current difference sequence segment and typical anomaly patterns recorded in the database. If the distance value is less than a preset similarity threshold, trajectory offset is confirmed. The three-level judgment results generate the final decision through weighted voting, with historical pattern matching accounting for 60%, trend analysis for 30%, and simple thresholding for 10%.

[0069] After the metabolite signal resampling command is triggered, the system enters a multi-stage noise reduction process. The first stage of frequency domain noise reduction implements wavelet packet decomposition, selecting the sym4 wavelet basis function for a 7-level decomposition. Noise level estimation for each sub-band uses a robust statistical method based on the median. Thresholding of detail coefficients employs a non-uniform strategy: a hard thresholding function is used for high-frequency sub-bands (layers 5-7), and a soft thresholding function is used for low-frequency sub-bands (layers 1-4). Before reconstructing the signal, sub-band consistency verification is performed, comparing the zero-crossing rate differences of adjacent sub-band coefficients; abnormal sub-bands trigger a local reconstruction mechanism. The second stage implements spatial filtering, applying an adaptive Wiener filter to the denoised signal segments. The filter parameters are dynamically adjusted based on the local signal-to-noise ratio, and the window length is set to 1.5 times the metabolite characteristic period.

[0070] The feedback update mechanism employs a double-buffered design. The denoised signal segment is first stored in a preparatory buffer, where timestamp alignment and format conversion are performed. After 60 seconds of data in the preparatory buffer, a hot update of the model parameters is triggered: the metabolite dynamic response model, upon receiving a new signal, prioritizes recalculating the boundary conditions at the biomembrane interface and updates the first three rows of the diffusion coefficient matrix; the metabolite interaction feature model resets the phase analysis buffer and reinitializes the cluster centers of the Gaussian mixture model. Data updates in the formal buffer utilize atomic operations to ensure that partial updates to the model output do not cause logical inconsistencies. Each update operation records the version number and timestamp, forming a complete model evolution log.

[0071] The anomaly handling during resampling incorporates multiple safeguards. When wavelet decomposition detects impulse interference in the signal, it automatically switches to robust decomposition mode, employing a more redundant dual-tree complex wavelet transform. When Wiener filtering encounters local signal-to-noise ratios below 3dB, joint bilateral filtering is activated as a supplementary scheme, with the standard deviations of the spatial kernel and the range kernel adaptively configured based on noise characteristics. In the event of a model hot update failure, a rollback mechanism is implemented, restoring to the most recently validated version, while simultaneously marking the data for that period as a low-confidence interval. All abnormal events trigger the generation of a system diagnostic package, containing snapshots of the original signal segments, intermediate processing results, and environmental parameters.

[0072] The dynamic convergence threshold adjustment strategy is based on the dynamic range of metabolite concentrations. The system maintains a historical concentration extreme value table for each metabolite. When the current concentration is detected to be close to 90% of the historical highest or lowest value, the convergence threshold is automatically relaxed by 30%. The circadian rhythm compensation module adjusts the baseline noise level estimate according to the acquisition time; the noise estimate during the daytime (06:00-18:00) is 20% higher than at night. The drug interference detector continuously monitors characteristic patterns in the signal. When a known drug interference feature is identified, the trajectory calibration function for the relevant metabolite is temporarily disabled until 60 minutes after the interference pattern disappears.

[0073] A signal quality assessment system is implemented throughout the entire processing chain. Upon the raw signal entering the system, six quality metrics are immediately calculated: signal-to-noise ratio (SNR), baseline stability, peak-to-valley symmetry, sampling integrity, noise dispersion, and interference pulse count. The denoised signal undergoes a morphological consistency check, comparing its matching degree with the original signal in terms of zero-crossing distribution and extreme point intervals. The trajectory data output by the model includes a confidence score, which integrates information from three dimensions: input signal quality, model residuals, and historical prediction accuracy. All quality assessment results are appended to the output data stream as metadata for subsequent analysis modules to access.

[0074] Version control and data traceability mechanisms ensure the auditability of the processing. Each signal resampling generates a unique processing batch ID, which persists throughout the entire feedback loop. Model parameter change records include three basic pieces of information: the modified content, the effective period, and the operator's identifier. The data lineage tracking system records the complete transformation path from the original signal to the final concentration value, including the version numbers and configuration parameters of all intermediate processing steps. Audit logs are stored using a copy-on-write strategy to ensure that historical records are not overwritten by subsequent operations.

[0075] Example 3: See Figure 4This study focuses on metabolite data fusion and anomalous modality identification. The construction of the metabolite feature reconstruction matrix employs a high-order tensor structure. The concentration trajectories of the first and second metabolites are concatenated along the time dimension to form a two-dimensional subspace, with the metabolite species dimension serving as the third axis for expansion. This three-dimensional tensor is reduced in dimensionality through Tucker decomposition, compressing the core tensor size to less than 30% of the original data. The extraction of the core metabolite modal feature vectors utilizes an incremental orthogonal iterative algorithm, with each iteration proceeding as follows:

[0076]

[0077] Where: V t+1 Let U represent the eigenvector matrix updated at time t+1, R be the Schmitt orthogonalization operator, and U be the eigenvector matrix updated at time t+1. t It is the temporal projection matrix of the previous time step, X. t W is the feature recombinant tensor of the current time window. t This represents a diagonal matrix of metabolite weights. The physical meanings of the parameters are as follows: the row vectors of V correspond to the core feature modes, the column vectors of U represent time-varying characteristics, and the diagonal elements of W reflect the fusion contribution of different metabolites. The iteration termination condition is set to the angle between two adjacent feature vectors changing by less than 0.1 radians. The output includes an orthogonalized feature matrix and a set of time-varying projection coefficients.

[0078] The metabolite concentration fusion map is achieved through bilinear mapping. The feature vector matrix and the real-time concentration trajectory undergo tensor shrinking to generate a low-dimensional embedding representation with temporal continuity. This representation is then reorganized in Hilbert space, transforming discrete projection points into a continuous surface using radial basis function kernels. The fusion map comprises three basic data layers: a principal component loading layer recording feature vector elements, a dynamic projection layer storing time-dependent coefficient changes, and a confidence layer labeling the standard deviation estimates for each region. The map resolution is set to 10 frames per second, with each frame independently encoded as a variable-length byte stream.

[0079] The dynamic calculation of the tolerance range for metabolite concentration fluctuations is based on quantile regression. The system maintains concentration distribution statistics within a sliding historical window, with the window length adaptively varying between 30 minutes and 6 hours, depending on the metabolite turnover rate. The upper boundary is calculated using an asymmetric strategy: the range from the first to the third quartile is multiplied by a scaling factor of 1.7 as the base range, and a periodic adjustment component is superimposed when a circadian rhythm pattern is detected. The boundary update frequency matches the dynamic characteristics of the metabolites: rapidly changing metabolites are updated every 3 minutes, while stable metabolites are updated hourly.

[0080] The anomaly detection mechanism implements a multi-level screening process. Primary screening, based on static threshold rules, marks discrete data points exceeding tolerance boundaries. Secondary analysis applies a temporal continuity test, performing forward and backward searches on initial anomalies: a clustering index is calculated within a 5-minute time window; when the clustering of discrete anomalies exceeds 0.85, a modality construction procedure is triggered. The definition of anomalous metabolite modalities requires simultaneous spatial and temporal constraints: spatially, it must include synergistic anomalies of two or more metabolites; temporally, its duration must exceed three times the average fluctuation period of the metabolite.

[0081] The frequency domain divergence feature is calculated using complex analytic wavelet transform. The Morette wavelet basis function decomposes the anomalous mode signal across six scales, extracting the information entropy of the wavelet coefficients as the frequency domain dispersion index. Temporal clustering analysis employs phase space reconstruction, with the delay time determined by the first minimum of the mutual information function, and the embedding dimension optimized using a pseudo-nearest neighbor algorithm. The feature fusion unit combines the frequency domain entropy value with the Lyapunov exponent into a 24-dimensional hybrid feature vector, where 14 dimensions describe the frequency domain characteristics and 10 dimensions characterize the temporal chaotic behavior. All features are standardized to eliminate the influence of different physical dimensions.

[0082] The encapsulation of the abnormal feature set adopts a hierarchical data structure. The metadata layer records basic attributes such as the start time, duration, and types of metabolites involved in the abnormal mode. The feature layer stores three data types: original waveform segments, frequency domain divergence parameters, and time domain clustering indices. The diagnostic layer retains intermediate results during the detection process, including wavelet decomposition tree diagrams, phase space reconstruction projections, and feature selection records. Data blocks are encoded into binary streams according to a preset template, with a version identifier and integrity check code added to the header, and a data lineage marker appended to the tail to indicate the acquisition channel and processing history of the original signal.

[0083] Data buffering and transmission mechanisms implement flow control. The generation of the feature reconstruction matrix adopts a dual-pipeline architecture, automatically switching to a simplified calculation mode when the main channel computation latency exceeds 100 milliseconds. The anomaly detection module sets up a priority queue to handle anomalies whose duration exceeds a critical threshold (usually 5 minutes). The result output port implements bandwidth management rules, allowing fused graph data to be transmitted first, while anomaly feature sets are sent in batches during network idle periods. All data packets are locally buffered in a circular manner with a retention period of 72 hours.

[0084] The system's robustness design incorporates multi-level fault-tolerance strategies. Metastability detection is implemented during eigenvector iterative calculation; if the algorithm fails to converge after 30 iterations, an auxiliary convergence engine is activated, reducing the optimization dimensionality through random projection of the subspace. Boundary condition verification is added to the abnormal mode construction process to eliminate false alarms caused by sensor drift. When a hardware clock anomaly is detected, the spectrum generation process is immediately frozen, and an event-time-based signal reconstruction mechanism is initiated. All fault-tolerance events trigger error logging, with log entries containing four key elements: timestamp, fault level, recovery measures, and checksum.

[0085] The data processing pipeline implements an asynchronous execution framework. Three core modules—matrix construction, graph generation, and anomaly detection—run in parallel, exchanging data via shared memory. The message queue uses a publish / subscribe model, keeping inter-module communication latency within 20 milliseconds. The task scheduler implements dynamic load balancing, migrating computationally intensive tasks across CPU cores. When the system load exceeds 85% of the threshold, the feature vector dimension is automatically reduced to maintain critical functionality. The resource monitor continuously tracks memory usage, CPU utilization, and disk I / O, triggering a three-tiered alarm strategy in case of abnormal conditions.

[0086] Data storage formats are standardized. Metabolite concentration fusion maps are organized using an HDF5 file structure, comprising three basic datasets: a time-series dataset recording sampling times with millisecond precision, a feature matrix dataset storing compressed feature vectors, and a metadata dataset recording processing parameters. Anomaly feature sets are saved in a custom indexed binary format, with the index header containing feature vector length, time range, and anomaly type identifier. All files are digitally signed, and the verification algorithm employs elliptic curve cryptography; the private key is stored in a hardware security module. The backup mechanism implements an incremental backup strategy, generating a differential snapshot every 8 hours, with a retention period of 30 days.

[0087] Example 4: See Figure 5 This study focuses on the matching and dynamic correction of abnormal metabolite mechanisms, using a glucose metabolism abnormality case to demonstrate the complete processing flow. When the system detects abnormal blood glucose concentration fluctuations lasting 15 minutes, the abnormal metabolite feature set is encapsulated into a structured data packet and transmitted to the matching engine. This data packet contains a 32-dimensional feature vector, including 12 dimensions of time-domain features, 14 dimensions of frequency-domain features, and 6 dimensions of cross-features. Standardization preprocessing is performed before feature matching to eliminate differences in dimensions between different sensors. The mechanism matching model includes three types of rule bases, as shown in Table 1.

[0088] Table 1: The mechanism matching model includes three types of rule bases.

[0089] Rule Type Feature Dimension Matching weight Dynamic parameters Response threshold Substrate inhibition 8 0.45 kc value 0.72 Product Feedback 10 0.33 Hill coefficient 0.65 Covalent modification 14 0.22 Phosphorylation rate 0.58

[0090] The matching process employs a hierarchical decision-making mechanism. Primary matching calculates the cosine similarity between the feature vector and each rule template, filtering candidate rules with a similarity exceeding 0.6. Intermediate matching performs constraint optimization, solving for a weight allocation scheme within the candidate rule space, requiring three constraints: the sum of all rule weights is 1, rules of the same type are mutually exclusive, and dynamic parameters are within a biochemically reasonable range. Advanced matching performs cross-validation, using matching results of similar anomalies from historical data as a reference; the current matching scheme must maintain at least 80% consistency with historical patterns. The final output includes a dominant anomaly type identifier and auxiliary type labels, such as "substrate inhibition dominant (75%), accompanied by biofeedback (25%)".

[0091] The correction strategy library employs a tree-like index structure. The first level categorizes by metabolite type, the second by anomaly type, and the third stores specific correction algorithms. For substrate inhibition anomalies in glucose metabolism, the strategy library includes three compensation schemes: dynamic adjustment of the Michaelis constant, competitive inhibition weight compensation, and membrane permeability correction algorithm. The system automatically selects a combination of schemes based on the matching confidence level. When the confidence level of the dominant type exceeds 80%, a single scheme is used; otherwise, a mixed compensation mode is activated. After receiving the matching results, the compensation coefficient generation module first queries the current metabolic environment parameters, including pH, body temperature, and blood oxygen saturation, using these environmental factors as adjustment parameters for the compensation algorithm.

[0092] Before dynamic correction can be implemented, a safety verification process is required. After the concentration compensation coefficient is generated, it enters a four-stage verification process: biochemical rationality check to confirm that the coefficient value is within the known physiological range; mass conservation verification to ensure that the total change in metabolite amount before and after correction does not exceed 5%; temporal consistency check to prevent abrupt changes in compensation amount at adjacent time points; and cross-metabolite impact assessment to avoid causing secondary anomalies. After all verifications pass, the compensation coefficient is encoded into an operation instruction set, which includes information such as correction magnitude, action time window, and metabolic pathway identifier. The instruction set is transmitted to the fusion map update module through a secure channel.

[0093] Taking the detection of abnormal liver glycogenolysis as an example, the complete system execution flow is as follows: The abnormal feature set shows an abnormal increase in energy in the frequency domain feature in the 0.05-0.1Hz band, and the time domain feature exhibits pulse-like fluctuations. The mechanism matching model calculates that the similarity between this feature and the substrate inhibition rule is 0.81, and the similarity with the product feedback rule is 0.43. Dynamic parameter estimation shows that the kc value deviates from the baseline level by 28%, and the Hill coefficient does not change significantly. The matching engine determines it as "abnormal glycogen phosphorylase substrate inhibition" with a confidence level of 83%. The correction strategy library calls the liver glycogen metabolism-specific compensation channel, and combined with the current body temperature of 36.7℃ and blood pH of 7.35, generates a compensation coefficient of 0.15±0.03. During the safety verification phase, it is found that this correction may affect lactate metabolism, and the system automatically adds a lactate buffer compensation subroutine. The final correction instruction includes a main compensation amount of 0.12 and a buffer compensation amount of 0.04, with an effect duration of 20 minutes.

[0094] A version rollback mechanism ensures the safety of corrections. Before each concentration correction, the system automatically saves the current fusion graph state to the rollback buffer. After the correction is implemented, three key indicators are continuously monitored: the smoothness of the compensated concentration curve, the covariance of related metabolites, and changes in system energy consumption. When any indicator exceeds the expected range, an automatic rollback process is triggered: first, the current compensation algorithm is stopped; then, the original data is restored from the buffer; and finally, the abnormal characteristics are reassessed. A diagnostic report is generated for the rollback event, recording key information such as abnormal behavior, environmental parameters, and compensation schemes for subsequent algorithm optimization.

[0095] Metabolic network balance monitoring serves as a supplementary safeguard. During dynamic correction, the system tracks the concentration trends of 12 relevant metabolites in real time and establishes a sub-network dynamics model. This model calculates the metabolic flow balance index every 5 minutes, issuing a regulation signal when the index deviates from the baseline value by more than 15%. The regulation signal operates through two pathways: directly feeding back to the compensation coefficient generation module to adjust the correction intensity; and indirectly triggering an increase in the monitoring frequency of relevant metabolites. Network balance data is visualized in the form of heatmaps to help operators understand the overall metabolic state.

[0096] The case database continuously accumulates processing experience. Each completed anomaly handling case is abstracted into a standardized template and stored in the database, containing five components: the original anomaly feature vector, the matching process record, the compensation strategy adopted, the tracking of the correction effect, and a snapshot of environmental parameters. The database implements version control, creating a new branch to store case data each time the algorithm is updated. The query engine supports multi-condition combined searches, for example, it can retrieve "substrate inhibition anomaly cases under body temperature >37℃ in the past three months".

[0097] The system maintenance module implements preventative management. Four core maintenance tasks are executed daily at midnight: rule base integrity verification via hash value comparison; strategy base index reconstruction to optimize query efficiency; case database compression and archiving to free up storage space; and compensation algorithm performance analysis to generate optimization suggestions. Maintenance operations employ a gradual execution strategy; any task taking more than 30 minutes is automatically paused and moved to the background for lower priority execution. Detailed logs are recorded for all maintenance activities, including start time, duration, resource utilization, and any abnormal events.

[0098] The user interface design emphasizes context awareness. In the glucose anomaly handling case, the interface dynamically displays three related views: the left view shows a feature comparison between the original abnormal waveform and the matching rules; the central view presents the decision tree and selection path of the compensation strategy; and the right view tracks the changes in the metabolic network after the correction is implemented. Operation controls are displayed context-sensitively; for example, the manual intervention option is automatically hidden when a high-confidence match is detected. Alarm information is presented in three levels: primary anomalies are displayed as blue alert boxes; if no action is taken for 30 minutes, it escalates to an orange warning; and if no action is taken for 60 minutes, it turns into a red emergency alert. All user operation records are audited, including operation content, timestamps, operator identity, and system status snapshots.

[0099] Example 5: Focusing on the stability verification and result output of dynamically corrected metabolite concentration data. Before the verification process begins, the system loads the corrected metabolite concentration fusion map and extracts the concentration sequence data at consecutive time points. The coefficient of variation is calculated using a dynamic window mechanism, with each three consecutive time points as the basic calculation unit, and the window sliding step along the time axis is ten seconds. Within each window, the ratio of the standard deviation to the mean is calculated, and this ratio is compared with a preset stability threshold in real time. When the coefficient of variation of a window exceeds the set threshold, the window is marked as pending inspection; when five consecutive windows meet the threshold, it is considered a locally stable region, and a region locking mechanism is triggered to keep the calculation parameters of that segment unchanged.

[0100] The analysis of autocorrelation decay rate was performed independently and in parallel. The system constructed a time-delay sequence with an initial delay step size set to twice the sampling interval, and a maximum delay depth covering a 120-second time period. For each metabolite channel, the autocorrelation coefficient at different delay times was calculated, forming a decay curve. The curve fitting process adopted a piecewise modeling strategy: a high-order autoregressive model was established in the initial delay (first 30 seconds); a moving average model was converted in the middle (30-90 seconds); and a hybrid modeling was implemented in the later stage (after 90 seconds). The white noise test of the model residuals used the rank-sum verification method to calculate the round distribution probability of the residual sequence. When the residuals passed the 95% confidence test and the decay curve showed a monotonically decreasing trend, the autocorrelation characteristics of the channel were deemed to meet the requirements.

[0101] The dynamic optimization system for validation parameters operates continuously. The window width for calculating the coefficient of variation automatically adjusts based on the concentration change rate: when the concentration difference between adjacent time points continuously exceeds three times the historical average, the window shrinks to a single-point calculation mode; when the concentration change is gradual, the window expands to ten time points. The delay step size for autocorrelation analysis is also adaptive, automatically matching the main periodic component when periodic fluctuations are detected. The validation threshold itself also has a floating range; normal physiological fluctuations caused by circadian rhythm patterns can trigger a temporary 30% increase in the threshold, with the specific increase amount referring to the current-day characteristic curve of the metabolite.

[0102] Comprehensive measures have been implemented to ensure the integrity of the concentration sequence. When data gaps are detected during validation, a bidirectional interpolation compensation mechanism is activated: small gaps (within three seconds) are filled using cubic spline interpolation; large gaps are filled by switching to a data imputation mode based on feature vector reconstruction. Boundary handling follows specific rules: missing start points are filled using a backward imputation strategy; missing end points are filled using a forward imputation process. All imputation operations are logged in detail, including the imputation location, imputation method, and source of the replacement values.

[0103] The output format for stable results is clearly defined. Qualified data is organized into a four-tuple structure: a timestamp accurate to milliseconds, a normalized metabolite identifier, a floating-point concentration value, and a confidence percentage value. The timestamp includes time zone information and conforms to the ISO 8601 standard; the metabolite identifier uses a combination of IUPAC nomenclature and KEGG encoding; the concentration value retains five significant digits; the confidence level ranges from 50% to 100%, discretely ranging in 0.5% steps. Data is grouped and packaged hourly, with each hourly data packet including a header checksum and packet sequence number.

[0104] Data transmission employs a layered security mechanism. The wireless transmission module uses a time-slice strategy, breaking down each data packet into multiple 128-byte segments, each independently encrypted and signed. The encryption algorithm uses AES-256-GCM mode, with the initialization vector changed hourly; the signature mechanism is based on the ECDSA-secp384r1 standard. During transmission gaps, it automatically switches to local buffer mode, with buffered data using a cyclic overwrite strategy and a retention period of 48 hours. When the network recovers, incremental synchronization is performed, retransmitting only unacknowledged data segments.

[0105] The local storage system employs a triple redundancy architecture. The primary storage area is an SQLite relational database containing four core tables: a time-series table recording all timestamps; a metabolite metadata table storing identifier resolution information; a concentration value table stored by time partition; and an audit log table tracking data change records. The secondary storage area implements a columnar storage structure, with hourly concentration data converted to Parquet format for compressed archiving. The cache area retains a memory mirror of the most recent 72 hours of data, using an LRU eviction algorithm to manage memory space. All storage areas implement data consistency checks, and an automatic reconciliation process is performed every 6 hours.

[0106] The operation audit module records all events throughout the entire process. The audit content covers six dimensions: system startup and shutdown events, significant configuration changes, verification process events, data completion operations, anomaly handling records, and result output records. Each audit entry includes four basic pieces of information: Greenwich Mean Time, operation type code, affected data range, and operator identifier. The audit logs are write-protected and stored using blockchain technology to ensure the integrity of the evidence chain. The hash value is anchored to a public timestamp service every hour.

[0107] A tiered fault failover mechanism is implemented. If the main processor load consistently exceeds 90% for five minutes, a backup computing core is automatically activated to share the verification task. When the storage system detects bad sectors on the disk, it immediately isolates the damaged sectors and redirects the data to reserved space. If transmission interruption exceeds two hours, an emergency protocol is triggered: a data simplification mode is activated, retaining only critical metabolite data; non-essential background services are stopped; and the sampling frequency is reduced to a baseline level. All emergency failover operations are accompanied by audio-visual warnings, and the degradation status is continuously displayed on the user interface.

[0108] The energy consumption monitoring system implements refined management. The processor is configured with a seven-level dynamic frequency adjustment strategy, automatically reducing the frequency to the base frequency when the computing load is below 40%. The wireless transmission module implements an intelligent sleep mechanism, switching to low-power standby after more than 30 seconds of network inactivity. The screen backlight automatically adjusts according to ambient light; when the system detects no user operation on the device, it automatically reduces the backlight brightness after five minutes of inactivity. All energy consumption data is recorded every ten minutes, forming a complete power consumption curve that is stored in the system diagnostic database.

[0109] Example 6: A non-invasive or minimally invasive continuous metabolite monitoring device is designed for children with inherited metabolic disorders. It employs a flexible skin patch-type biosensor array for signal acquisition. This array integrates a miniature electrochemical sensing unit, a near-infrared optical sensing unit, and a miniature thermocouple thermodynamic sensing unit, with an overall thickness of less than 0.5 mm and rounded edges to conform to the delicate skin of children. The electrochemical sensing unit is fabricated using printed electrode technology, with platinum as the working electrode and silver / silver chloride as the reference electrode, capturing the redox current signals of free amino acids (such as phenylalanine and tyrosine) in the blood. The near-infrared optical sensing unit uses a 780-1050 nm wavelength light source, generating optical signals by detecting changes in the characteristic absorption peaks of different metabolites (such as organic acids and sugars). The thermodynamic sensing unit acquires thermodynamic signals by monitoring minute temperature fluctuations (accuracy up to ±0.01℃) in local tissue caused by metabolic reactions. All three signals are converted into digital signals by a low-power signal processor integrated within the patch and transmitted in real-time to a terminal device via Bluetooth, forming a multi-source metabolite signal sequence including timestamps.

[0110] For the multi-source metabolite signal sequences transmitted to the terminal, time-series segmentation was first performed. Considering that the metabolic state of children is easily affected by feeding and activity and fluctuates rapidly, an adaptive sliding window algorithm was adopted. The window length was dynamically adjusted according to the intensity of signal fluctuations in the previous 5 minutes. When fluctuations were severe, the window was shortened to 2 minutes, and when they were stable, it was extended to 5 minutes. The overlap rate was maintained at 50% to ensure temporal continuity. The metabolite concentration time-series segments generated after segmentation simultaneously carried electrochemical, optical, and thermodynamic signal characteristics and corresponding time indices, providing a basis for subsequent model analysis.

[0111] The application of the metabolite dynamic response model in this monitoring instrument requires parameter adjustments based on the physiological characteristics of the stratum corneum of children's skin. When constructing the model, constraints are placed on the stratum corneum diffusion rate (corrected based on children's skin thickness and moisture content) and the reaction threshold of specific enzymes (such as phenylalanine hydroxylase). The migration path of metabolites from the dermis to the sensor surface is discretized into 50-80 nodes, with the node density increased closer to the sensor. Adjacent nodes are associated using a dynamic damping coefficient, which adjusts with changes in children's skin temperature (feedback from thermodynamic signals in real time). The damping coefficient decreases as temperature increases to reflect accelerated diffusion, thereby generating a metabolite concentration gradient field containing spatial distribution information. Based on this gradient field, concentration inversion is performed on multi-source signal sequences. For example, changes in the current value of electrochemical signals and the absorbance of optical signals are correlated through gradient field constraints, ultimately outputting the first metabolite concentration trajectory. This trajectory visually presents the changing trends of key metabolites such as phenylalanine over time.

[0112] The metabolite interaction feature model processes the same time segments synchronously, focusing on capturing the correlations between different metabolites. The model first extracts the phase clustering of signals, analyzing the phase synchronicity of electrochemical and optical signals along the time axis to determine the metabolic synergy between phenylalanine and tyrosine. Simultaneously, it identifies polarity effect characteristics, i.e., the consistency of the amplitude change direction of different metabolite signals (e.g., whether the carbohydrate signal changes inversely when the organic acid concentration increases). A variation analysis model quantifies the synergistic shift of these features; for example, when the phenylalanine signal leads the tyrosine signal beyond the normal range, or when their polarity changes from positive to negative, the specific shift value is calculated. Based on this shift, the competitive inhibition relationship between metabolites is dynamically weighted; for example, the inhibitory effect of phenylalanine on tyrosine hydroxylase increases in weight with increasing shift. The final output is a second metabolite concentration trajectory that reflects metabolite interactions.

[0113] To ensure monitoring accuracy, the monitor performs cross-model calibration on the two trajectories. The discrete differences between the first and second metabolite concentration trajectories at the same time point are calculated, forming a difference sequence. This sequence is input into the trajectory convergence judgment unit to generate a trajectory offset weighting factor, which comprehensively considers the amplitude, duration, and trend of the difference. When the factor exceeds a preset child-specific metabolite convergence threshold (preset according to different types of inherited metabolic diseases, such as a lower threshold for children with phenylketonuria than for children with organic acidemia), the terminal automatically triggers a metabolite signal resampling command. Upon response, the system performs frequency domain noise reduction on the original multi-source signal sequence, using wavelet thresholding to filter out motion noise and ambient light interference generated by the child's activities, generating a denoised signal segment, which is then fed back to the two models to update the concentration trajectories, ensuring that the trajectories more closely reflect the child's actual metabolic state.

[0114] Based on the updated two trajectories, the system constructs a metabolite feature recombination matrix. Rows represent time points, and columns contain the concentration values ​​and interaction feature parameters of different metabolites. Principal component decomposition is performed on the matrix to extract core metabolite modal feature vectors, which focus on key metabolites associated with inherited metabolic diseases (such as phenylalanine and its metabolites in phenylketonuria). A metabolite concentration fusion map is generated based on the feature vectors. The map uses time as the horizontal axis and metabolite concentration as the vertical axis, with different colors distinguishing the fusion weights of electrochemical, optical, and thermodynamic signal inversion results.

[0115] The monitor incorporates a tolerance range for metabolite concentration fluctuations specific to children. This range is dynamically adjusted based on metabolic data from healthy children of the same age and the individual patient's historical data. During dynamic screening of the fusion spectrum, if the concentration values ​​at three consecutive time points exceed the tolerance range, the system automatically identifies the abnormal metabolite mode. The frequency domain divergence features (e.g., energy distribution of the abnormal signal in a specific frequency band) and temporal clustering features (e.g., duration of abnormality, concentration peak interval) of this mode are extracted to generate a metabolite abnormality feature set. This feature set is input into a metabolite mechanism matching model, which includes quantitative rules applicable to children for substrate inhibition (e.g., phenylalanine inhibition of enzymes), product feedback inhibition (e.g., inhibition of upstream reactions by metabolite accumulation), and covalent modification. The matching degree between the feature set and each rule is calculated using a least-squares approximation algorithm, outputting an abnormality type identifier, such as "phenylalanine substrate inhibition abnormality" or "organic acid product feedback inhibition abnormality."

[0116] Based on the abnormality type identifier, the system retrieves the corresponding pediatric metabolite correction strategy library. This library contains specific compensation schemes for different inherited metabolic diseases; for example, the correction strategy for phenylketonuria considers the impact of dietary phenylalanine intake on metabolism. Metabolite concentration compensation coefficients are generated based on the strategy library. The coefficients are dynamically adjusted according to the child's weight, age, and current metabolic status to dynamically correct the fusion profile. After correction, the system calculates the coefficient of variation and autocorrelation decay rate of the concentration sequence. When the coefficient of variation is below the pediatric metabolic stability threshold and the autocorrelation decay rate conforms to the exponential decay model (indicating that metabolic fluctuations tend to stabilize), the terminal outputs the calibrated continuous multi-metabolite monitoring results. The results are presented in graphical form, including concentration trend graphs, abnormal event markers, and mechanism analysis, providing continuous and comprehensive metabolite information for the clinical monitoring of children with inherited metabolic diseases.

[0117] Example 7: A personalized recipe generation app for obese and diabetic children can achieve precise adaptation by integrating multimodal cross-calibrated continuous multi-metabolite monitoring technology. This app works in conjunction with an implantable biosensor array. After the sensor array is implanted subcutaneously, it continuously collects multi-source metabolite signal sequences, including electrochemical, optical, and thermodynamic signals. Electrochemical signals primarily reflect changes in the concentration of electroactive metabolites such as glucose and lactic acid in the child's blood. Optical signals capture the levels of metabolites such as insulin and ketone bodies by monitoring changes in absorption spectra at specific wavelengths. Thermodynamic signals help determine the activity levels of processes such as fat breakdown and glycogen synthesis based on changes in heat during metabolic reactions. These multi-source signals are transmitted in real-time to the app's signal processing module, forming a timestamp-aligned continuous signal sequence.

[0118] The app's time-series segmentation module uses a sliding window algorithm to divide the received multi-source metabolite signal sequences. The window length is set to 30 minutes per segment, with 50% overlap between adjacent windows, generating time-series segments of metabolite concentration with time indices. Each time-series segment is synchronously input into the app's built-in metabolite dynamic response model and metabolite interaction feature model. In the metabolite dynamic response model, the system discretizes the migration path of glucose and other metabolites in the body into multiple computational nodes, using the diffusion rate of glucose to biological membranes (such as cell membranes) and the response threshold of related enzymes (such as hexokinase) as constraints. Adjacent nodes are then linked through a dynamic damping coefficient to construct a metabolite concentration gradient field. Based on this gradient field, the model performs concentration inversion on the multi-source signals, generating a first metabolite concentration trajectory. This trajectory clearly shows the dynamic changes in glucose and other metabolites over time, such as the rise in postprandial blood glucose and the rate of decrease during fasting.

[0119] Metabolite interaction modeling focuses on analyzing the interactions between different metabolites. The model first extracts phase clustering and polarity effect features from multi-source signals, such as the phase synchronicity of glucose and insulin signal changes and the polar influence of lipid metabolites on glucose signals. A variation analysis model quantifies the co-shifts of these features, and then dynamically weights the competitive inhibitory relationships between metabolites (such as the inhibition of glucose uptake by free fatty acids) based on these shifts to generate a second metabolite concentration trajectory. This trajectory reflects the interactive effects between metabolites, such as the slowing effect of increased fatty acid concentration on glucose clearance rate after high-fat intake.

[0120] To ensure trajectory accuracy, the app performs cross-model calibration on the first and second metabolite concentration trajectories. The system calculates the discrete difference sequence of the two trajectories and inputs it into the trajectory convergence judgment unit to generate a trajectory offset weight factor. When this factor exceeds a preset convergence threshold, the app triggers a metabolite signal resampling command to perform frequency domain noise reduction on the original multi-source signal, eliminating noise caused by interference factors such as motion and body temperature fluctuations, generating a denoised signal segment, and feeding it back to the two models to update the concentration trajectory.

[0121] Based on the updated two trajectories, the app constructs a metabolite feature recombination matrix, extracts core metabolite modal feature vectors through principal component decomposition, and then generates a metabolite concentration fusion map. The map clearly displays the concentration change curves of key metabolites such as glucose, insulin, and free fatty acids throughout the day. The app presets tolerable ranges for metabolite concentration fluctuations for obese and diabetic children. For example, it sets a stricter range for the normal fluctuation range of fasting blood glucose. When the metabolite concentration at multiple consecutive time points in the fusion map exceeds this range, the system marks it as an abnormal metabolite modality. For example, if a child's glucose concentration 2 hours after a meal exceeds the tolerable range for three consecutive time points, accompanied by a slow increase in insulin concentration, the system will mark this as an abnormal modality and extract the frequency domain divergence features (such as the frequency distribution of blood glucose fluctuations) and temporal clustering features (such as the duration and amplitude of abnormality) of this modality to form a set of abnormal metabolite features.

[0122] The app inputs anomaly feature sets into a metabolite mechanism matching model, which includes quantitative rules for substrate inhibition, product feedback inhibition, and covalent modification. The app calculates the matching degree between the feature set and each rule using a least-squares approximation algorithm. If the matching results indicate that the anomaly primarily stems from substrate inhibition (e.g., high glucose intake leading to suppressed glucokinase activity), the system outputs a corresponding anomaly type identifier. Based on this identifier, the app retrieves the corresponding metabolite correction strategy from the correction strategy library, generates a concentration compensation coefficient, and dynamically corrects the fused spectrum, such as correcting blood glucose prediction biases caused by substrate inhibition.

[0123] The corrected fusion map undergoes stability verification. The system calculates the coefficient of variation and autocorrelation decay rate of the corrected concentration sequence. When both meet stability criteria, the final metabolite monitoring results are used to generate personalized diets. For example, if monitoring reveals that a child's glucose concentration trajectory significantly exceeds the tolerance range after consuming refined rice and flour, and the insulin interaction characteristics show a delayed response, the app will, based on the child's age, weight, activity level, and other information, reduce the proportion of refined rice and flour in the diet and increase whole grains rich in dietary fiber. If an abnormally high concentration of free fatty acids is detected after a high-fat diet, inhibiting glucose metabolism, the app will adjust the fat sources in the diet, replacing some saturated fatty acids with unsaturated fatty acids (such as olive oil and deep-sea fish oil) and controlling the total daily fat intake. The app will adjust the diet recommendations in real time based on dynamic changes in metabolite concentrations. For instance, if a child's activity level increases on a particular day, leading to faster glucose consumption and low blood sugar, the diet will appropriately increase the proportion of carbohydrates to maintain metabolic balance. In this way, the app can provide personalized recipes tailored to the metabolic characteristics of obese and diabetic children based on accurate metabolite monitoring data, helping them control their blood sugar and manage their weight.

[0124] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various 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 method of continuous multi-metabolite monitoring of multi-modal cross- calibration, characterized in that, The method comprises the following steps: Collecting multi-source metabolite signal sequences by implanting a biosensor array, the multi-source metabolite signal sequences including electrochemical signals, optical signals and thermodynamic signals; Temporal segmentation of the multi-source metabolite signal sequences to generate metabolite concentration time series fragments; Synchronous input of the metabolite concentration time series fragments into a metabolite dynamic response model and a metabolite interaction feature model to generate a first metabolite concentration trajectory and a second metabolite concentration trajectory, respectively; When constructing the metabolite dynamic response model, discretizing the metabolite migration path with the biofilm diffusion rate and enzyme reaction threshold as constraint conditions; Based on the metabolite migration path, generating a metabolite concentration gradient field by associating adjacent path nodes through a dynamic damping coefficient; Concentration inversion of the multi-source metabolite signal sequences according to the metabolite concentration gradient field to output the first metabolite concentration trajectory; When constructing the metabolite interaction feature model, extracting phase aggregation and polarity effect features of the multi-source metabolite signal sequences; Quantifying the synergistic offset of the phase aggregation and polarity effect features through a variation analysis model; Based on the synergistic offset, dynamically weighting the metabolite competition inhibition relationship to output the second metabolite concentration trajectory; Cross-model calibration of the first metabolite concentration trajectory and the second metabolite concentration trajectory.

2. The method of multi-modal cross-calibration continuous multi-metabolite monitoring of claim 1, wherein, The cross-model calibration of the first metabolite concentration trajectory and the second metabolite concentration trajectory comprises: Calculating the discrete difference sequence of the first metabolite concentration trajectory and the second metabolite concentration trajectory; Inputting the discrete difference sequence into a trajectory convergence judgment unit to generate a trajectory offset weight factor; When the trajectory offset weight factor exceeds a preset convergence threshold, triggering a metabolite signal resampling instruction.

3. The method of continuous multi-metabolite monitoring of multi-modal cross- calibration according to claim 2, wherein, Further comprising: In response to the metabolite signal resampling instruction, performing frequency domain noise reduction processing on the multi-source metabolite signal sequences to generate a noise-reduced metabolite signal fragment; Feeding the noise-reduced metabolite signal fragment back to the metabolite dynamic response model and the metabolite interaction feature model to update the first metabolite concentration trajectory and the second metabolite concentration trajectory.

4. The method of multi-modal cross-calibration continuous multi-metabolite monitoring of claim 1, wherein, Further comprising: Based on the first metabolite concentration trajectory and the second metabolite concentration trajectory, constructing a metabolite feature reorganization matrix; Principal component decomposition of the metabolite feature reorganization matrix to extract a core metabolite modal feature vector; Generating a metabolite concentration fusion map according to the core metabolite modal feature vector.

5. The method of multi-modal cross-calibration continuous multi-metabolite monitoring of claim 4, wherein, Further comprising: Setting a metabolite concentration fluctuation tolerance interval to dynamically screen the metabolite concentration fusion map; When the concentration values of consecutive time points in the metabolite concentration fusion map exceed the metabolite concentration fluctuation tolerance interval, marking an abnormal metabolite mode; Extracting the frequency domain divergence features and time domain aggregation features of the abnormal metabolite mode to generate a metabolite abnormal feature set.

6. The method of multi-modal cross-calibration continuous multi-metabolite monitoring of claim 5, wherein, Further comprising: Inputting the metabolite abnormal feature set into a metabolite mechanism matching model, the metabolite mechanism matching model including quantification rules of substrate inhibition, product feedback inhibition and covalent modification; The matching degree of the metabolite abnormality feature set and each quantification rule is calculated by a least square approximation algorithm, and an metabolite abnormality type identifier is output.

7. The method of multi-modal cross-calibration continuous multi-metabolite monitoring of claim 6, wherein, Also comprising: According to the metabolite abnormality type identifier, a corresponding metabolite correction strategy library is called; Based on the metabolite correction strategy library, a metabolite concentration compensation coefficient is generated to dynamically correct the metabolite concentration fusion spectrum.

8. The method of multi-modal cross-calibration continuous multi-metabolite monitoring of claim 7, wherein, The dynamically corrected metabolite concentration fusion spectrum is subjected to stability verification, including: The coefficient of variation and the autocorrelation decay rate of the corrected concentration sequence are calculated; When the coefficient of variation is lower than a preset stability threshold and the autocorrelation decay rate conforms to an exponential decay model, a calibrated continuous multi-metabolite monitoring result is output.

Images