Multi-dimensional health data fusion analysis system and method

Abstract

The present application relates to a kind of multi-dimensional health data fusion analysis system and method, first to the concurrent collection of intelligent bracelet multiple-source heterogeneous physiological movement data is space-time alignment and frequency domain denoising preprocessing, subsequently while extracting in multiple branch feature, based on the ratio of fundamental energy and full-band energy to pulse wave signal Real-time quality quantification, generate dynamic quality index.In the fusion stage, the quality index is used as the dynamic scaling mask factor in cross-modal attention mechanism, when the quality of core physiological signal decreases, automatically pull up the attention guide weight of high-quality auxiliary mode to damaged physiological characteristics, realize the feature self-healing repair in semantic level, instead of simply discarding or static weighting.Finally, through multi-task learning network, map out multi-dimensional health index, and combine individual dynamic baseline to carry out coupled perception joint abnormality judgment, so as to significantly improve the accuracy and continuity of health analysis in strong interference environment without increasing hardware.

Description

Technical Field

[0001] This invention belongs to the field of health data analysis technology, specifically relating to a multi-dimensional health data fusion analysis system and method. Background Technology

[0002] With the widespread adoption of smart wearable devices in personal health management, continuous and non-invasive monitoring of wearers' multi-dimensional physiological parameters using terminal devices such as smart bracelets has become a core technological direction in the digital health industry. Smart bracelets typically integrate multiple heterogeneous sensing units, including photoplethysmography (PPG) sensors, inertial motion sensors, and skin temperature sensors, enabling concurrent collection of multi-source physiological and movement data such as heart rate, blood oxygenation, exercise status, and body surface temperature. However, the data acquired by a single sensor has limited dimensions and is easily affected by environmental interference, making it difficult to comprehensively and accurately reflect the wearer's true health status. Therefore, constructing a multi-dimensional health data fusion and analysis scheme, deeply fusing and collaboratively analyzing multi-source heterogeneous sensor data at the feature level, has significant technical value and clinical implications for improving the accuracy and reliability of health assessments.

[0003] Existing multi-dimensional health data fusion and analysis solutions typically employ static weighted fusion or simple feature cascading and stitching strategies. These methods linearly combine or directly stitch together features extracted from various sensors with fixed weights before inputting them into downstream health assessment models. These solutions implicitly assume, in their underlying architecture, that the signal quality of data from each modality of the sensor remains consistently reliable over time. However, during complex dynamic scenarios such as running or ball sports, physical factors like large arm swings, periodic muscle deformation, and changes in wearing tightness can severely contaminate the photoplethysmography (PPG) signal with motion artifacts, causing a sharp drop in the signal-to-noise ratio (SNR) of that modality. In such cases, existing fusion solutions cannot quantify the signal quality confidence of each sensor data slice in real-time and dynamically. When low-quality PPG features containing significant noise are forcibly input into the static fusion model along with high-fidelity inertial motion features, the low-quality data directly pollutes the entire feature space, triggering cross-modal feature conflicts and leading to catastrophic distortion of output health indicators such as heart rate and blood oxygen saturation. The root cause of this defect is that the existing fusion engine lacks dynamic data quality perception capabilities and cross-modal semantic compensation mechanisms. When the quality of core physiological signals drops sharply, it cannot adaptively use high-quality auxiliary modal data to repair and fill in the damaged physiological features at the semantic level, which seriously restricts the accuracy and continuity of health analysis of smart bracelets in strong interference environments.

[0004] Therefore, an optimized multi-dimensional health data fusion and analysis scheme is desired. Summary of the Invention

[0005] This invention application provides a multi-dimensional health data fusion and analysis system and method. According to one aspect of this application, a multi-dimensional health data fusion and analysis method is provided, comprising: S1 uses multiple sensors built into the smart bracelet to concurrently acquire and spatiotemporally align preprocessing of multi-source physiological motion data streams to obtain pulse wave preprocessing sequences, inertial motion preprocessing sequences, and body surface temperature preprocessing sequences. S2, through a multi-branch feature extraction network, modal features are extracted from the pulse wave preprocessing sequence, inertial motion preprocessing sequence and body surface temperature preprocessing sequence to obtain the pulse wave feature matrix, motion rhythm feature matrix and body surface context feature vector; S3, Dynamic quantization of signal quality is performed on the preprocessed pulse wave sequence to obtain the pulse wave quality index; S4. Using the pulse wave quality index as a dynamic scaling factor, the pulse wave feature matrix, the motion rhythm feature matrix, and the body surface context feature vector are fused together with cross-modal semantic repair based on quality perception to obtain the cross-modal fused feature tensor. S5 uses a multi-task learning network to perform inverse mapping of health status indicators on cross-modal fusion feature tensors to obtain a multi-dimensional health indicator set. S6, based on an individual's dynamic physiological baseline, makes a deviation judgment on a multidimensional health indicator group to obtain a health assessment and early warning report.

[0006] According to another aspect of this application, a multi-dimensional health data fusion and analysis system is provided, comprising: The data acquisition and preprocessing module is used to concurrently acquire and spatiotemporally align and preprocess multi-source physiological motion data streams through multiple sensors built into the smart bracelet to obtain pulse wave preprocessing sequences, inertial motion preprocessing sequences, and body surface temperature preprocessing sequences. The modal feature extraction module is used to extract modal features from the pulse wave preprocessing sequence, inertial motion preprocessing sequence, and body surface temperature preprocessing sequence through a multi-branch feature extraction network to obtain the pulse wave feature matrix, motion rhythm feature matrix, and body surface context feature vector. The signal quality dynamic quantization module is used to perform dynamic signal quality quantization on the pulse wave preprocessing sequence to obtain the pulse wave quality index. The cross-modal semantic repair and fusion module is used to perform quality-aware cross-modal semantic repair and fusion on the pulse wave feature matrix, the motion rhythm feature matrix, and the body surface context feature vector using the pulse wave quality index as a dynamic scaling factor to obtain the cross-modal fused feature tensor. The Health Status Indicator Inverse Mapping Module is used to perform health status indicator inverse mapping on cross-modal fusion feature tensors through a multi-task learning network to obtain a multi-dimensional health indicator set. The deviation judgment module is used to make deviation judgments on a group of multidimensional health indicators based on an individual's dynamic physiological baseline in order to obtain a health assessment and early warning report.

[0007] Compared with existing technologies, this invention proposes a multi-dimensional health data fusion and analysis system and method. First, it performs spatiotemporal alignment and frequency domain denoising preprocessing on multi-source heterogeneous physiological motion data concurrently collected by a smart bracelet. Then, while extracting multi-branch features, it performs real-time quality quantization of the pulse wave signal based on the ratio of fundamental energy to full-band energy, generating a dynamic quality index. In the fusion stage, this quality index is used as a dynamic scaling mask factor in a cross-modal attention mechanism. When the quality of the core physiological signal decreases, the attentional guidance weight of the high-quality auxiliary modality on the damaged physiological features is automatically increased, achieving semantic-level feature self-healing repair, rather than simple discarding or static weighting. Finally, a multi-task learning network maps multi-dimensional health indicators, and combined with an individual dynamic baseline, performs coupled perception for joint anomaly judgment. This significantly improves the accuracy and continuity of health analysis under strong interference environments without increasing hardware requirements. Attached Figure Description

[0008] Figure 1 This is a flowchart of a multi-dimensional health data fusion and analysis method according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in the multi-dimensional health data fusion and analysis method according to an embodiment of this application; Figure 3 This is a flowchart of step S2 of the multi-dimensional health data fusion analysis method according to an embodiment of this application; Figure 4 This is a flowchart of step S5 of the multi-dimensional health data fusion analysis method according to an embodiment of this application; Figure 5 This is a flowchart of step S6 of the multi-dimensional health data fusion analysis method according to an embodiment of this application; Figure 6 This is a flowchart illustrating the multi-dimensional health data fusion analysis method according to embodiments of this application, which compares the absolute values ​​of multi-dimensional physiological deviation feature vectors dimension by dimension with preset medical safety boundary thresholds using an indicative function to obtain a set of abnormal alarm flags. Figure 7 This is a block diagram of a multi-dimensional health data fusion analysis system according to an embodiment of this application. Detailed Implementation

[0009] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0010] To address the technical problem that smart bracelets suffer from a sharp drop in the signal-to-noise ratio of core physiological signals due to motion artifacts in complex dynamic scenarios, and that existing static fusion solutions cannot dynamically perceive changes in data quality or use auxiliary modalities to semantically compensate for damaged features, ultimately leading to cross-modal feature conflicts and distortions in health assessments, this application proposes a multi-dimensional health data fusion and analysis method. The specific implementation process is as follows: First, using the high-frequency clock of the inertial motion signal as a reference, asynchronous timestamp alignment and zero-phase bandpass filtering are performed on the three heterogeneous signals of photoplethysmography pulse wave, six-axis inertial motion, and skin temperature. Second, the pulse wave feature matrix, motion rhythm feature matrix, and body surface context feature vector are extracted through a multi-branch network, and the fundamental energy ratio of the pulse wave signal in the current time window is calculated simultaneously. After normalization mapping, a real-time quality index is generated. Then, the inverted value of the quality index is used to construct an exponential dynamic mask. In the cross-modal attention calculation, the integral of the query and key is scaled and modulated element by element, so that the semantic guidance weight of the auxiliary modality is higher when the signal quality is lower, thereby achieving deep repair of damaged physiological features rather than simple discarding. The repaired fused features are mapped into continuous physiological indicators and discrete state labels through multi-task network branches and encapsulated into a multi-dimensional health indicator group. Finally, the individual dynamic baseline is updated based on the exponential weighted moving average, and global coupling anomaly judgment is achieved through cross-dimensional coupling weight matrix and quadratic joint scoring, outputting a health assessment early warning report.

[0011] Figure 1 This is a flowchart of a multi-dimensional health data fusion analysis method according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the multi-dimensional health data fusion and analysis method according to an embodiment of this application. Figure 1 and Figure 2As shown, the multi-dimensional health data fusion analysis method according to an embodiment of this application includes the following steps: S1, using multiple sensors built into a smart bracelet to concurrently acquire and spatiotemporally align multi-source physiological motion data streams to obtain a pulse wave preprocessing sequence, an inertial motion preprocessing sequence, and a body surface temperature preprocessing sequence; S2, using a multi-branch feature extraction network to extract modal features from the pulse wave preprocessing sequence, the inertial motion preprocessing sequence, and the body surface temperature preprocessing sequence to obtain a pulse wave feature matrix, a motion rhythm feature matrix, and a body surface context feature vector; S3, processing the pulse wave data... S4. The pulse wave preprocessing sequence is dynamically quantized to obtain the pulse wave quality index; S5. The pulse wave quality index is used as a dynamic scaling factor to perform quality-aware cross-modal semantic repair and fusion on the pulse wave feature matrix, the motion rhythm feature matrix, and the body surface context feature vector to obtain the cross-modal fusion feature tensor; S6. Through a multi-task learning network, the cross-modal fusion feature tensor is inversely mapped to health status indicators to obtain a multidimensional health indicator group; S7. Based on the individual dynamic physiological baseline, the deviation of the multidimensional health indicator group is judged to obtain a health assessment and early warning report.

[0012] Specifically, in step S1, the multi-source physiological motion data stream is concurrently acquired and spatiotemporally aligned and preprocessed using multiple sensors built into the smart bracelet to obtain pulse wave preprocessing sequences, inertial motion preprocessing sequences, and skin surface temperature preprocessing sequences. It is worth noting that the multi-source physiological motion data stream includes photoplethysmography (PPG) signals, six-axis inertial motion signals, and skin temperature signals. It should be understood that because the PPG sensor, six-axis inertial motion sensor, and skin temperature sensor built into the smart bracelet each have independent hardware sampling clocks and different sampling frequencies, the three heterogeneous signals have inherent asynchronous offsets and inconsistent sampling intervals on the time axis. If the unaligned raw data is directly input into the subsequent feature extraction and fusion network, it will cause misalignment of features from different modalities in the time dimension, leading to the failure of cross-modal feature association and systematic deviations in health indicator mapping. Simultaneously, low-frequency baseline drift and high-frequency circuit thermal noise superimposed on the raw signals will further contaminate the frequency band of effective physiological information. Therefore, in the technical solution of this application, multiple sensors built into the smart bracelet are used to concurrently acquire and spatiotemporally align multi-source physiological motion data streams, including photoplethysmography pulse wave signals, six-axis inertial motion signals, and skin temperature signals, to obtain pulse wave preprocessing sequences, inertial motion preprocessing sequences, and skin surface temperature preprocessing sequences. This eliminates clock drift and sampling rate differences between multi-source heterogeneous signals and suppresses the interference of noise in each frequency band on effective physiological information. This ensures that the three input data received by the subsequent multi-branch feature extraction network are strictly aligned in the time dimension and that the signal-to-noise ratio meets the quality requirements of feature-level fusion, laying a reliable data foundation for semantic-level feature interaction in the cross-modal attention fusion stage.

[0013] More specifically, in the embodiments of this application, step S1 includes: using the high-frequency hardware clock timestamp of the six-axis inertial motion signal as the reference clock, performing asynchronous timestamp alignment and resampling mapping on the photoplethysmography pulse wave signal and the skin temperature signal to obtain a pulse wave aligned sequence, an inertial motion aligned sequence, and a body surface temperature aligned sequence; and using an infinite impulse response digital bandpass filter to perform forward and reverse zero-phase frequency domain noise suppression filtering on the pulse wave aligned sequence, the inertial motion aligned sequence, and the body surface temperature aligned sequence to obtain a pulse wave preprocessing sequence, an inertial motion preprocessing sequence, and a body surface temperature preprocessing sequence, respectively.

[0014] Specifically, in one example of this application, the smart bracelet's photoplethysmography (PPG) sensor acquires PPG signals at a sampling rate of 25 Hz, the six-axis inertial motion sensor acquires six-axis inertial motion signals at a sampling rate of 100 Hz, and the skin temperature sensor acquires skin temperature signals at a sampling rate of 1 Hz. These three signals operate concurrently at the hardware level and each has its own independent hardware timestamp. First, a unified target time grid is constructed using the high-frequency hardware clock timestamp of the six-axis inertial motion signal as the reference clock. Since the six-axis inertial motion signal has the highest sampling frequency among the three sensors, its timestamp sequence has the finest time resolution and can provide accurate interpolation anchors for other low-frequency signals. Therefore, the sampling time sequence of the six-axis inertial motion signal is directly used as the time reference for the inertial motion alignment sequence. Next, asynchronous timestamp alignment and resampling mapping are performed on the photoplethysmography (PPG) signal and skin temperature signal. Specifically, for each reference moment on the target time grid, two adjacent sampling points surrounding the reference moment are located in the original timestamp sequences of the PPG signal and skin temperature signal, respectively. The signal amplitude corresponding to the reference moment is calculated using a first-order linear interpolation algorithm, thereby resampling the PPG signal and skin temperature signal to the same 100 Hz sampling rate as the six-axis inertial motion signal, respectively, to obtain the pulse wave alignment sequence, the inertial motion alignment sequence, and the body surface temperature alignment sequence. The three alignment sequences strictly correspond to each sample on the time axis. Then, using an infinite impulse response digital bandpass filter, positive and negative zero-phase frequency domain noise suppression filtering is applied to the pulse wave alignment sequence, inertial motion alignment sequence, and body surface temperature alignment sequence, respectively. Specifically, the passband for the pulse wave alignment sequence is set to 0.5 Hz to 8 Hz to preserve the cardiac fundamental wave and its harmonic components while suppressing low-frequency baseline drift caused by respiration and high-frequency noise introduced by ambient light. The passband for the inertial motion alignment sequence is set to 0.1 Hz to 20 Hz to cover the main frequency range of daily human movement. The passband for the body surface temperature alignment sequence is set to 0 Hz to 0.1 Hz to preserve the slow temperature change trend and filter out high-frequency thermal noise from the sensor circuit. Furthermore, the implementation method of forward and reverse zero-phase filtering is to first pass the aligned sequence through an infinite impulse response filter along the forward time direction to complete a first filtering, and then pass the forward filtering result through the same filter along the reverse time direction to complete a second filtering. The phase delay introduced by the forward filtering and the phase advance introduced by the reverse filtering cancel each other out, thereby achieving frequency domain noise suppression while keeping the time domain phase of the signal waveform undistorted, and finally obtaining the pulse wave preprocessing sequence, inertial motion preprocessing sequence and body surface temperature preprocessing sequence.

[0015] Specifically, in step S2, a multi-branch feature extraction network is used to extract modal features from the pulse wave preprocessing sequence, inertial motion preprocessing sequence, and body surface temperature preprocessing sequence to obtain the pulse wave feature matrix, motion rhythm feature matrix, and body surface context feature vector. It should be understood that since the pulse wave preprocessing sequence, inertial motion preprocessing sequence, and body surface temperature preprocessing sequence respectively carry three fundamentally different physiological and physical information: cardiovascular volume changes, trunk motion state, and body surface thermal environment, and these three sequences differ significantly in signal morphology, time-frequency characteristics, and semantic level, using a single feature extraction architecture to process the three sequences uniformly will fail to fully exploit the specific semantic information contained in each modality of data. This results in a lack of sufficiently rich and discriminative feature representations in the subsequent cross-modal fusion stage, thus limiting the accuracy of health indicator mapping. Therefore, in the technical solution of this application, a multi-branch feature extraction network is further used to extract modal features from the pulse wave preprocessing sequence, inertial motion preprocessing sequence, and body surface temperature preprocessing sequence to obtain the pulse wave feature matrix, motion rhythm feature matrix, and body surface context feature vector. This allows for the design of adaptive feature extraction paths tailored to the time-frequency characteristics of each modal data, ensuring that the deep semantic information of each signal is fully and specifically represented. This provides high-quality query, key-value projection sources for the subsequent cross-modal attention fusion stage, ensuring that under the modulation of the signal quality-aware mask, high-quality auxiliary modal features possess sufficient semantic richness to effectively repair and compensate for damaged physiological features.

[0016] Figure 3 This is a flowchart of step S2 of the multi-dimensional health data fusion and analysis method according to an embodiment of this application. Figure 3 As shown, step S2 includes: S21, performing sliding window convolution feature mapping on the pulse wave preprocessing sequence based on a one-dimensional temporal convolutional network to obtain the pulse wave feature matrix; S22, performing multi-dimensional feature reduction and aggregation on the inertial motion preprocessing sequence to obtain the motion rhythm feature matrix; S23, performing environmental context deep feature extraction on the body surface temperature preprocessing sequence to obtain the body surface context feature vector.

[0017] More specifically, in a concrete example of this application, the multi-branch feature extraction network comprises three parallel, independent processing branches, corresponding to the feature extraction tasks of the pulse wave preprocessing sequence, the inertial motion preprocessing sequence, and the body surface temperature preprocessing sequence, respectively. First, the pulse wave preprocessing sequence is input into a one-dimensional temporal convolutional network branch, which consists of multiple cascaded one-dimensional convolutional layers. Each convolutional layer is configured with several one-dimensional convolutional kernels of fixed receptive field spans. The convolutional kernels move gradually along the time axis in a sliding window manner on the pulse wave preprocessing sequence, performing a dot product operation with the local signal segment at each time position and superimposing a bias term to extract the local waveform morphology features at that time position. After multiple convolutional kernels are processed in parallel, they are stacked in the channel dimension. Through the cascaded mapping of layer-by-layer convolution and nonlinear activation, the final output is a pulse wave feature matrix containing multi-channel cardiovascular volume change semantics. The row dimension of this matrix corresponds to the number of feature channels, and the column dimension corresponds to the number of time steps. Next, the inertial motion preprocessing sequence is input into the motion feature extraction branch. This branch performs temporal encoding and multi-dimensional feature dimensionality reduction aggregation on the three-axis acceleration and three-axis angular velocity components in the six-axis inertial motion data. Specifically, it performs step-by-step downsampling along the time axis through a multi-layer convolutional pooling structure to compress the high-frequency sampled six-axis motion data into a low-dimensional motion rhythm representation. It retains macroscopic motion pattern information such as the wearer's step frequency, arm swing period, and posture transition, while filtering out high-frequency jitter noise. Finally, it outputs a motion rhythm feature matrix. The row dimension of this matrix corresponds to the number of motion feature channels, and the column dimension corresponds to the number of time steps after downsampling. Then, the preprocessed body surface temperature sequence is input into the environmental context feature extraction branch. Since the frequency of skin temperature signal changes is much lower than that of pulse wave and inertial motion signals, the information it carries mainly reflects the slow trend changes of the wearer's thermal environment. Therefore, this branch uses a fully connected linear mapping layer to perform deep environmental context feature extraction on the preprocessed body surface temperature sequence. After multiplying the temperature time series data by a learnable weight matrix and superimposing a bias vector, redundant negative components are filtered out through a nonlinear activation function, and finally, a body surface context feature vector is output. This vector encodes the global state information of the wearer's body surface thermal environment within the current time window in a compact one-dimensional form. Furthermore, the network parameters of the three branches are independently optimized during the training phase, and the weights are not shared between the branches. This ensures that the pulse wave feature matrix focuses on encoding the temporal morphological features of cardiovascular volume pulsation, the motion rhythm feature matrix focuses on encoding the periodic rhythmic features of trunk dynamic behavior, and the body surface context feature vector focuses on encoding the global state features of the external thermal environment. The three form a complementary and non-redundant multimodal feature representation system in semantic space.

[0018] Specifically, in step S3, the pulse wave preprocessing sequence is dynamically quantized to obtain a pulse wave quality index. It should be understood that smart bracelet wearers inevitably experience complex dynamic scenarios such as running, arm swinging, and muscle contraction during daily activities. These movements cause severe motion artifact interference to the photoplethysmography (PPG) sensor, resulting in drastic fluctuations in the signal quality of the pulse wave preprocessing sequence across different time windows. The subsequent cross-modal attention fusion stage needs to dynamically adjust the semantic guidance intensity of the auxiliary modality based on the real-time reliability of the pulse wave signal. Without real-time quantitative evaluation of the pulse wave signal quality, the fusion network cannot distinguish whether the pulse wave features at the current moment are reliable physiological information or artifact components heavily contaminated by noise, thus failing to effectively trigger the cross-modal semantic repair mechanism. Therefore, in the technical solution of this application, the pulse wave preprocessing sequence is further dynamically quantified to obtain a pulse wave quality index. This index assigns a continuous quality confidence scalar between zero and one to the pulse wave signal within each time window, serving as the basis for calculating the dynamic scaling factor in the subsequent fusion stage. In this way, the cross-modal attention mechanism can obtain real-time data quality perception capabilities, maintain the autonomous expression weight of physiological features when the pulse wave signal quality is high, and automatically increase the repair guidance intensity of the auxiliary mode when the signal quality drops sharply, fundamentally solving the feature conflict problem caused by the lack of quality perception in existing static fusion schemes.

[0019] More specifically, in a concrete example of this application, the core idea of ​​dynamic signal quality quantification is to determine the degree of motion artifact contamination in a given pulse wave preprocessing sequence based on the proportion of effective cardiac fundamental energy to the total energy across the entire frequency band within the current time window. First, the pulse wave preprocessing sequence is segmented according to the same time windows as the feature extraction step. Spectral analysis is performed on the pulse wave segments within the current time window to calculate the total fundamental energy within the effective physiological heartbeat frequency band, which corresponds to the frequency range mapped from the resting to the exercise state of a normal human body. Next, the total energy across the entire frequency band of the pulse wave sequence within this time window is calculated. This total energy includes the fundamental energy and the energy of non-physiological frequency bands introduced by motion artifacts, ambient light interference, and circuit noise. Then, the ratio of the total fundamental energy to the total energy across the entire frequency band is calculated. This ratio reflects the proportion of energy occupied by the effective physiological components in the pulse wave signal within the current time window. A higher ratio indicates less artifact contamination, while a lower ratio indicates that motion artifact energy occupies the main frequency band of the signal. Furthermore, the energy ratio is subtracted by the fundamental energy ratio threshold bias term, multiplied by the steepness hyperparameter, and the result is negative. This negative value is then input into the Sigmoid activation function for nonlinear normalization mapping, ultimately yielding the pulse wave quality index. The calculation formula for this dynamic quantization process of signal quality is as follows: in, The pulse wave quality index, Time windows in pulse wave preprocessing sequences Inside The sum of the fundamental wave signal energy, For time windows Total energy across the entire frequency band of the internal pulse wave sequence. This is the fundamental frequency energy ratio threshold bias term. The steepness hyperparameter controls the nonlinear mapping gradient of the Sigmoid function near the boundary point. In the above formula, the numerator part... With the denominator ratio This directly reflects the proportion of the effective cardiac fundamental component relative to the total signal energy in the pulse wave preprocessing sequence within the current time window. When the wearer is at rest or in a state of mild activity, the motion artifact energy is low, and the proportion of fundamental energy is close to one, resulting in a relatively large ratio. When the wearer engages in strenuous exercise, motion artifacts inject a large amount of noise energy across a wide frequency band, causing the proportion of fundamental energy to drop sharply, and this ratio approaches zero. Bias Term The introduction of this feature allows the center decision point of the Sigmoid function to be flexibly adjusted according to different sensor hardware characteristics and wearing scenarios, while the steepness hyperparameter... It controls the sharpness of the transition between high-quality and low-quality signals in the quality index, resulting in a larger [quality index]. The value causes the quality index to exhibit an approximately step-like binarization characteristic; a smaller value... The value maintains a smooth, continuous transition; in this scheme, an appropriate value is selected. The value is designed to balance the sensitivity and robustness of quality judgment. The final output pulse wave quality index serves as the direct input to the dynamic scaling factor in the subsequent cross-modal attention fusion step. The closer the value is to one, the more reliable the current pulse wave features are, and the autonomous weights of the physiological modality should be maintained during fusion. The closer the value is to zero, the more contaminated the current pulse wave features are, and the stronger the semantic repair guidance of the auxiliary modality should be during fusion.

[0020] Specifically, in step S4, the pulse wave quality index is used as a dynamic scaling factor to perform quality-aware cross-modal semantic repair and fusion on the pulse wave feature matrix, the motion rhythm feature matrix, and the body surface context feature vector to obtain a cross-modal fusion feature tensor. It should be understood that, because the pulse wave feature matrix is ​​contaminated by motion artifacts during vigorous exercise, its feature representation deviates from the true cardiovascular physiological semantics. Meanwhile, the trunk motion state and thermal environment information carried by the motion rhythm feature matrix and the body surface context feature vector maintain high fidelity at the same time. If the traditional static weighted splicing or fixed-weight linear fusion strategy is used to merge the three features, the contaminated pulse wave features will be injected into the fusion space with equal weight, directly contaminating the entire feature representation and causing cross-modal feature conflicts. The pulse wave quality index generated in the previous step provides a scalar basis for real-time quantification of the reliability of the pulse wave signal and can be used as a control signal for dynamically adjusting the interaction weights of each modality during the fusion process. Therefore, in the technical solution of this application, the pulse wave quality index is further used as a dynamic scaling factor to perform quality-aware cross-modal semantic repair fusion of the pulse wave feature matrix, the motion rhythm feature matrix, and the body surface context feature vector to obtain a cross-modal fused feature tensor. This automatically enhances the semantic guidance and repair strength of the auxiliary modality for physiological features when the pulse wave signal quality deteriorates, and maintains the autonomous expression weight of the physiological modality when the signal quality is normal. In this way, a leap from traditional physical-level denoising to semantic-level cross-modal feature self-healing and compensation can be achieved, so that the fused feature tensor can maintain an accurate representation of the real physiological state under any intensity of motion interference, fundamentally eliminating the problem of health indicator distortion caused by static fusion schemes that cannot perceive changes in data quality.

[0021] More specifically, in the embodiments of this application, step S4 includes: concatenating the motion rhythm feature matrix and the body surface context feature vector along the feature channel dimension to obtain a concatenation result, projecting the pulse wave feature matrix into a query matrix, and projecting the concatenation result into a key matrix and a value matrix respectively; and performing quality-aware dynamic mask modulation and semantic-level self-healing fusion on the query matrix, key matrix, and value matrix based on a dynamic mask factor constructed from the inverted value of the pulse wave quality index to obtain a cross-modal fusion feature tensor.

[0022] Specifically, in a concrete example of this application, firstly, the kinetic rhythm feature matrix and the surface context feature vector are concatenated along the feature channel dimension to obtain the concatenated result. This concatenation operation horizontally cascades the multi-channel temporal features representing trunk kinetic rhythms with the global state features representing the surface thermal environment along the channel dimension, forming a unified auxiliary modality joint feature space. The surface context feature vector is aligned to the same number of time steps as the kinetic rhythm feature matrix through dimensional expansion before concatenation. Next, the pulse wave feature matrix is ​​projected into a query matrix using three independent sets of learnable linear projection matrices, and the concatenated result is projected into a key matrix and a value matrix, respectively. This projection operation maps the physiological modality and the auxiliary modality to a unified-dimensional attention hidden layer space. The query matrix carries the request information for semantic retrieval from the pulse wave features to the auxiliary modality, the key matrix carries the semantic index information that can be matched in the auxiliary modality, and the value matrix carries the actual semantic content that can be transmitted in the auxiliary modality. Then, a dynamic masking factor is constructed based on the inverted value of the pulse wave quality index. Specifically, the pulse wave quality index is subtracted from the numerical value, multiplied by a learnable control coefficient, and then exponentially calculated to generate an exponential dynamic masking factor. This masking factor approaches one when the pulse wave quality index is close to one (i.e., the signal quality is excellent) without producing an additional amplification effect. When the pulse wave quality index is close to zero (i.e., the signal is severely damaged), it grows exponentially, thus significantly amplifying the attention weights of the auxiliary modality. Further, the attention score of the dot product of the query matrix and the transpose of the key matrix is ​​calculated and divided by the arithmetic square root of the key dimension for basic scaling. The dynamic masking factor is applied to this attention score matrix element-wise. After Softmax probability normalization, it is multiplied by the value matrix to complete weighted aggregation. Finally, nonlinear dimensionality reduction fusion is performed through a multilayer perceptron to obtain the cross-modal fusion feature tensor. The calculation formula for this fusion process is as follows: in, The output is the cross-modal fusion feature tensor; This is a nonlinear dimensionality reduction operation for a multilayer perceptron. This is a normalization function that transforms the attention score into a probability distribution along the last dimension; This is the query matrix obtained by linear projection of the pulse wave feature matrix; This is the transpose of the key matrix obtained by linear projection of the splicing result; is a scalar dimension of the attention hidden layer key space, used to prevent the gradient from vanishing due to excessively large dot product values; For element-wise multiplication, the Hadamah product operator is used. For exponential operations with the natural constant as the base; Learnable control coefficients for controlling the nonlinear amplification of the quality mask; It is an all-one matrix used to calculate the inversion value of the pulse wave quality index; This is the pulse wave quality index output from the previous step; This is the value matrix obtained by linear projection of the spliced ​​result. In this formula, the fractional part... The raw semantic relevance score between pulse wave physiological features and auxiliary modal features was calculated, reflecting the degree of semantic matching between pulse wave features and various locations of the body surface environment at the current time step. Dynamic masking factor. middle The pulse wave quality index is inverted as an indicator of contamination level. When the wearer is at rest, the pulse wave quality index is close to one, and the inversion value is close to zero. The index calculation result approaches one, and the masking factor does not amplify the attention score. The fusion process maintains the autonomous expression of pulse wave characteristics. When the wearer is engaged in strenuous exercise, the pulse wave quality index is close to zero, and the inversion value is close to one. (The last sentence appears to be incomplete and requires further context.) The amplified result of the exponential operation is significantly greater than one, and the masking factor has an exponential amplification effect on the attention score. This forces the attention probability after Softmax normalization to concentrate on the highly relevant positions of the auxiliary modality, thereby injecting the high-quality motion rhythm and environmental context semantic information carried in the value matrix into the fusion result with greater weight, achieving semantic-level self-healing repair of damaged pulse wave physiological features. Finally, after nonlinear dimensionality reduction and integration by a multilayer perceptron, the cross-modal fusion feature tensor can maintain an accurate representation of the wearer's true cardiovascular physiological state under any intensity of motion disturbance.

[0023] Specifically, in step S5, a multi-task learning network is used to perform inverse mapping of health status indicators on the cross-modal fusion feature tensor to obtain a multi-dimensional health indicator group. It should be understood that, since the cross-modal fusion feature tensor, after quality-aware semantic repair and fusion, already contains a deep joint representation of pulse wave physiological semantics, trunk movement rhythm, and body surface thermal environment context, this tensor is still in a high-dimensional hidden feature space and has not yet been transformed into health status indicators with clear physical meaning and clinical reference value. Smart bracelets, for health assessment of end users, need to simultaneously output continuous numerical physiological parameters (such as real-time heart rate and blood oxygen saturation estimates) and discrete categorical state determinations (such as sleep stages and arrhythmia classifications). These two types of indicators mathematically belong to regression and classification tasks, respectively. If a complete feature extraction and prediction network is independently constructed for each task, it will lead to redundant model parameters and an inability to utilize the shared underlying physiological feature correlations between different tasks. Therefore, in the technical solution of this application, a multi-task learning network is further used to inversely map health status indicators onto the cross-modal fusion feature tensor to obtain a multi-dimensional health indicator group. This allows for joint prediction of continuous indicator regression and discrete state classification simultaneously through parallel task-specific branches, based on shared underlying feature representations. This enables efficient decoding of multi-dimensional health status information from the fusion feature tensor in a single forward inference, while leveraging the implicit regularization effect between multiple tasks to improve the generalization ability and robustness of various indicator predictions, providing complete and structured health indicator inputs for subsequent individual baseline deviation judgment.

[0024] Figure 4 This is a flowchart of step S5 of the multi-dimensional health data fusion and analysis method according to an embodiment of this application. Figure 4 As shown, step S5 includes: S51, performing shared low-level feature extraction and multi-task branch prediction on the cross-modal fusion feature tensor to obtain continuous physiological indicator prediction vector and discrete state probability vector; S52, performing maximum a posteriori probability index extraction on the discrete state probability vector to obtain discrete state classification label; S53, performing structured encapsulation of continuous physiological indicator prediction vector and discrete state classification label according to timestamp to obtain multidimensional health indicator group.

[0025] Accordingly, in step S51, shared low-level feature extraction and multi-task branch prediction are performed on the cross-modal fusion feature tensor to obtain continuous physiological indicator prediction vectors and discrete state probability vectors. It should be understood that although the cross-modal fusion feature tensor, after quality-aware semantic repair fusion, possesses a high robustness representation ability of the wearer's true physiological state, it still exists in the network's internal feature space as a high-dimensional hidden layer tensor and has not yet been decoded into a health state value with clear physical dimensions and clinical reference significance. The health assessment task of the smart bracelet essentially includes two types of prediction objectives with drastically different mathematical properties: the need to output continuous numerical physiological parameter estimations and discrete categorical state determinations. Although these two tasks differ in output form, they share a high degree of commonality in their underlying feature dependencies. Constructing independent complete networks for each would lead to parameter redundancy and prevent the utilization of complementary regularization effects between tasks. Therefore, in the technical solution of this application, continuous physiological indicator prediction vectors and discrete state probability vectors are obtained by performing shared low-level feature extraction and multi-task branch prediction on the cross-modal fusion feature tensor. This allows for the extraction of task-independent general physiological pattern representations through a shared low-level network, and then the decoding of regression and classification prediction targets is completed through parallel task-specific branches. In this way, continuous physiological parameters and discrete state probabilities can be output simultaneously in a single forward inference, leveraging the implicit regularization effect of the shared layer to improve the generalization ability of each branch prediction, while reducing the computational overhead and model storage requirements of edge inference.

[0026] More specifically, in a concrete example of this application, the network architecture for shared low-level feature extraction and multi-task branch prediction consists of a shared encoding module and two parallel task-specific decoding branches. First, the cross-modal fusion feature tensor is globally averaged along the time dimension, compressing it into a fixed-length one-dimensional feature vector to eliminate the variability of the time step dimension and adapt to the fixed input dimension requirements of subsequent fully connected layers. Next, this one-dimensional feature vector is input to a shared fully connected representation layer, which consists of two cascaded fully connected networks. Each layer performs weight matrix multiplication and bias stacking, followed by a nonlinear mapping using the ReLU activation function. Layer by layer, the feature dimension is compressed from the number of channels in the fusion tensor to the preset shared hidden layer dimension, outputting a compact global task-independent feature representation. This representation simultaneously encodes low-level cardiovascular dynamic pattern information that contributes to both heart rate regression and state classification. Then, the global task-independent feature representations are distributed to the regression branch, which consists of a single fully connected output layer. The number of output neurons equals the number of continuous physiological parameters to be predicted. A linear mapping is performed on the shared features to directly output a continuous physiological indicator prediction vector. Each dimension of this vector corresponds to the real-time heart rate estimate and blood oxygen saturation estimate, respectively, without additional activation function constraints to maintain the continuous numerical domain of the regression output. Further, the global task-independent feature representations are simultaneously distributed to the classification branch, which also consists of a single fully connected output layer. The number of output neurons equals the total number of discrete state categories. A linear mapping is performed on the shared features to obtain the unnormalized logarithmic score for each category. Subsequently, the logarithmic score is normalized along the category dimension using the Softmax function, transforming it into a posterior probability distribution for each category, outputting a discrete state probability vector. Each dimension of this vector corresponds to the posterior probability value of the wearer's current cardiac rhythm (normal sinus rhythm, atrial premature beats, ventricular premature beats, atrial fibrillation, etc.) or sleep stage (awake, light sleep, deep sleep, REM sleep, etc.).

[0027] Accordingly, in step S52, the discrete state probability vector is extracted using maximum a posteriori probability index to obtain discrete state classification labels. It should be understood that since the discrete state probability vector represents the posterior confidence of the wearer's current state category in the form of a probability distribution, with the sum of the probability values ​​of each dimension being one and each dimension being a continuous real number, this soft probability distribution, while preserving classification uncertainty information, cannot be directly used as a deterministic label input for the discrete state field in the subsequent multidimensional health indicator group. Subsequent individual baseline deviation judgments and early warning report generation require a clear single-category judgment result rather than a fuzzy probability distribution, in order to perform a deterministic comparison with the preset medical safety boundary. Therefore, in the technical solution of this application, the discrete state probability vector is further extracted using maximum a posteriori probability index to obtain discrete state classification labels, thereby hardening the continuous probability distribution into a uniquely determined category judgment result, completing the transformation from a soft decision space to a hard decision space. In this way, clear discrete state determination labels can be provided for subsequent structured encapsulation steps, so that the multidimensional health indicator group can simultaneously contain two complementary types of health state description information: continuous numerical parameters and deterministic categorical labels.

[0028] More specifically, in a concrete example of this application, firstly, the discrete state probability vector output by the classification branch is received. This vector contains the posterior probability values ​​of various discrete state categories, such as normal sinus rhythm, premature atrial contractions, premature ventricular contractions, and atrial fibrillation, at the wearer's current moment. Next, all elements of the discrete state probability vector are traversed along the category dimension, and the category index corresponding to the probability element with the largest value is located. This operation is mathematically equivalent to performing an Argmax operation on the discrete state probability vector, selecting the category index that maximizes the posterior probability as the final judgment result. Then, the extracted category index is mapped to a pre-established state semantic label dictionary, converting the integer index into a discrete state classification label with clear clinical meaning. This label represents the wearer's discrete health status judgment result at the current moment in the form of a single deterministic category. Furthermore, the maximum a posteriori probability index extraction operation follows the Bayesian optimal decision criterion, that is, under the zero-one loss function, selecting the category with the highest posterior probability as the final judgment can make the classification error rate reach the theoretical minimum, thereby ensuring that the discrete state classification label has the highest judgment accuracy in a statistical sense, and providing reliable discrete state field input for the subsequent structured encapsulation of multidimensional health indicator groups.

[0029] Accordingly, in step S53, the continuous physiological indicator prediction vector and the discrete state classification label are structurally encapsulated by timestamp to obtain a multidimensional health indicator group. It should be understood that since the continuous physiological indicator prediction vector and the discrete state classification label originate from two independent parallel branches in the multi-task learning network, they differ fundamentally in data type, dimensional structure, and semantic level. The continuous physiological indicator prediction vector is a floating-point vector containing values ​​such as real-time heart rate and blood oxygen saturation, while the discrete state classification label is a single-category label representing heart rhythm status or sleep stage. If the two are not standardized and integrated under a unified time index, the subsequent individual baseline deviation judgment step will be unable to access both continuous and discrete health indicators simultaneously using a standardized data interface, resulting in a lack of structural consistency in the data input for deviation calculation and threshold comparison. Therefore, in the technical solution of this application, the continuous physiological indicator prediction vector and the discrete state classification label are further structurally encapsulated by timestamp to obtain a multidimensional health indicator group, thereby integrating heterogeneous health indicators into a standardized composite data structure under a unified time anchor. This provides a unified, time-aligned, and semantically complete multidimensional health status snapshot input for subsequent individual dynamic physiological baseline deviation judgment, ensuring that deviation measurement and abnormal early warning processes can simultaneously obtain continuous parameters and discrete state information through standardized field access interfaces.

[0030] More specifically, in a concrete example of this application, firstly, the absolute timestamp corresponding to the current prediction performed by the system is obtained. This timestamp is consistent with the reference clock used in the spatiotemporal alignment preprocessing in the preceding steps, ensuring that the time index in the multidimensional health indicator group strictly corresponds to the time axis of the original collected data. Next, using this absolute timestamp as the primary key, the continuous physiological indicator prediction vector is written as a continuous field into the structured data dictionary. This field contains continuous numerical physiological parameters such as the current real-time heart rate estimate and blood oxygen saturation estimate. Then, the discrete state classification label is written as a discrete field into the same structured data dictionary. This field records the deterministic discrete health state category determination result of the wearer at the current moment. Furthermore, the timestamp primary key, continuous fields, and discrete fields are encapsulated into a complete structured data dictionary instance to obtain a multidimensional health indicator group. This indicator group integrates continuous physiological parameters and discrete state labels into a self-contained multidimensional health status record unit using a single time index as the anchor point. Subsequent deviation judgment steps can directly access the continuous physiological indicators within the group through the field names to calculate the difference with the individual's dynamic physiological baseline vector, while also accessing the discrete state labels for contextual auxiliary judgment of state transitions.

[0031] Specifically, in step S6, based on the individual's dynamic physiological baseline, a deviation judgment is made on the multidimensional health indicator group to obtain a health assessment early warning report. It should be understood that since the real-time heart rate and blood oxygen saturation values ​​included in the continuous physiological indicator prediction vector of the multidimensional health indicator group only reflect the wearer's instantaneous physiological state at the current moment, and there are significant physiological differences between different individuals, the normal physiological parameter range of the same individual will also slowly drift during different age stages, seasonal changes, and long-term health status evolution. If a fixed group statistical threshold is used to uniformly determine anomalies for all users, it will be impossible to distinguish between normal physiological differences between individuals and true pathological deviations, leading to continuous false alarms for athletes with high baseline heart rates or missed alarms for elderly users with low baseline heart rates. Therefore, in the technical solution of this application, a deviation judgment is further made on the multidimensional health indicator group based on the individual's dynamic physiological baseline to obtain a health assessment early warning report. This establishes a personalized reference benchmark that adaptively updates with the long-term evolution of the individual's physiological state, and based on this benchmark, standardized deviation measurement and anomaly threshold judgment are performed on the health indicators at the current moment. In this way, the judgment criteria for abnormal warnings can be accurately adapted to the individual physiological characteristics and long-term trends of each wearer, effectively suppressing false alarms caused by individual differences while maintaining high sensitivity, and ultimately outputting a health assessment and warning report that combines personalized accuracy with clinical reference value.

[0032] Figure 5 This is a flowchart of step S6 of the multi-dimensional health data fusion and analysis method according to an embodiment of this application. Figure 5 As shown, step S6 includes: S61, iteratively updating historical physiological data to obtain an individual dynamic physiological baseline vector; S62, dynamically measuring the deviation between the continuous physiological indicators in the multidimensional health indicator group and the individual dynamic physiological baseline vector to obtain a multidimensional physiological deviation feature vector; S63, comparing the multidimensional physiological deviation feature vector dimension by dimension with a preset medical safety boundary threshold using an indicator function to obtain an abnormal alarm flag set; S64, structurally encapsulating the multidimensional health indicator group and the abnormal alarm flag set to obtain a health assessment early warning report.

[0033] Accordingly, in step S61, the historical physiological data is iteratively updated to obtain an individual dynamic physiological baseline vector. It should be understood that the normal physiological parameters of a smart bracelet wearer are not fixed static constants, but rather drift slowly and continuously with factors such as age, adaptation to physical training, seasonal climate changes, adjustments to sleep patterns, and long-term health evolution. For example, the resting heart rate of a wearer who consistently engages in aerobic exercise will decrease month by month as they adapt to training. In winter, the wearer's baseline blood oxygen level may fluctuate slightly due to changes in the respiratory mucosa. If a fixed initial baseline value is used as the reference benchmark for subsequent deviation judgment, this fixed baseline will gradually deviate from the wearer's current true normal physiological level over time, leading to a systematic shift in deviation measurement, which in turn causes a decrease in the sensitivity of abnormal judgment or an increase in the false alarm rate. Therefore, in the technical solution of this application, an individual dynamic physiological baseline vector is obtained by iteratively updating historical physiological data, thereby constructing a dynamic reference benchmark that can adaptively track the wearer's long-term physiological trend changes, ensuring that deviation judgment always uses the wearer's recent true normal level as the comparison anchor. This ensures that no matter what long-term, slow changes the wearer's physiological state undergoes, the judgment benchmark of the abnormality warning system will always remain synchronized with its current normal physiological level, fundamentally eliminating systematic judgment bias caused by outdated baselines.

[0034] More specifically, in a concrete example of this application, the iterative update of the individual dynamic physiological baseline vector is implemented using an exponentially weighted moving average algorithm. First, the individual dynamic physiological baseline vector saved at the previous time step is read from the smart bracelet's local storage. Each dimension of this vector corresponds to the historical cumulative estimates of continuous physiological parameters such as the baseline heart rate and baseline blood oxygen saturation. Upon initial system startup, this vector is initialized with the mean of resting physiological data collected during the wearer's initial wearing period. Next, a continuous physiological indicator prediction vector is extracted from the current multidimensional health indicator set as the new observation input for this iteration. This vector contains the current real-time heart rate and blood oxygen saturation estimates obtained through cross-modal fusion and multi-task network mapping. Then, the current continuous physiological indicator prediction vector and the individual dynamic physiological baseline vector from the previous time step are weighted and mixed according to a preset smoothing factor. The smoothing factor is set to a small positive number close to zero to ensure that the baseline is sensitive only to long-term trends and remains inert for short-term fluctuations. New observations contribute to the baseline update with the smoothing factor as the weight, while the historical baseline is maintained by subtracting the smoothing factor from the weight. The weighted sum of the two yields the updated individual dynamic physiological baseline vector at the current time. Further, the updated individual dynamic physiological baseline vector is written back to the smart bracelet's local storage, overwriting the old value from the previous time step, completing one full iteration. This updated individual dynamic physiological baseline vector serves as the reference input for subsequent dynamic deviation measurement steps. The calculation formula for this iterative update is as follows: in, This is the individual dynamic physiological baseline vector output after the update at the current time tt; This is the current-moment continuous physiological indicator prediction vector extracted from the multidimensional health indicator set; This is the individual dynamic physiological baseline vector stored locally at the previous time step; This is the smoothing factor in the exponentially weighted moving average algorithm, and its value is a positive real number between zero and one. Looking at the structure of the formula... This partially characterizes the contribution of newly observed physiological indicators to the instantaneous update of the baseline at the current moment, and its weight is determined by a smoothing factor. Control, in this example Setting it to 0.01 means that only one percent of new observation information is injected into the baseline in each iteration, ensuring that a single outlier measurement will not significantly disturb the baseline. The inertial retention of the historical cumulative baseline is partially represented by a weight of 0.99, ensuring that the main information of the baseline comes from the cumulative average effect of a large number of past time steps, thus maintaining high stability against short-term noise fluctuations and occasional outliers. The weighted mixing mechanism of the two makes the individual dynamic physiological baseline vector mathematically equivalent to the weighted mean of all historical observations after applying exponentially decaying weights. The closer the observation is to the current time, the greater its contribution, while the contribution of observations further away from the current time decays exponentially to approach zero. This characteristic allows the baseline to maintain robustness to short-term noise while naturally tracking the slow drift trend of physiological parameters caused by long-term training adaptation or seasonal changes in the wearer.

[0035] Accordingly, in step S62, the continuous physiological indicators in the multidimensional health indicator group are dynamically deviated from the individual dynamic physiological baseline vector to obtain a multidimensional physiological deviation feature vector. It should be understood that since the real-time heart rate and blood oxygen saturation values ​​included in the prediction vector of the continuous physiological indicators in the multidimensional health indicator group are instantaneous absolute measurements at the current moment, and there are significant differences in dimensions and numerical scales between different physiological dimensions—the normal fluctuation range of heart rate is several to more than ten beats per minute, and the normal fluctuation range of blood oxygen saturation is only one to two percent change—if the absolute difference is directly used as a measure of deviation, it will be impossible to fairly compare the severity of deviations in different dimensions under a unified numerical scale. This leads to the larger-dimensional dimensions dominating anomaly judgment in the subsequent threshold decision stage, while smaller-dimensional dimensions are systematically ignored. Therefore, in the technical solution of this application, the continuous physiological indicators in the multidimensional health indicator group are further dynamically deviated from the individual's dynamic physiological baseline vector to obtain a multidimensional physiological deviation feature vector. This transforms the absolute offset of each dimension into a standardized relative deviation scaled by the wearer's own historical fluctuations, eliminating dimensional differences between different physiological dimensions. This allows subsequent threshold decisions to perform fair and consistent anomaly comparisons of all physiological dimensions within a unified dimensionless standardized space, ensuring that abnormal deviations in heart rate and blood oxygen saturation can be detected with the same sensitivity.

[0036] More specifically, in a concrete example of this application, the dynamic deviation measurement process employs the Z-score standardization method. First, a continuous physiological indicator prediction vector is extracted from the current multidimensional health indicator set. This vector contains continuous numerical physiological parameters such as the current real-time heart rate estimate and blood oxygen saturation estimate. Next, the continuous physiological indicator prediction vector is compared dimension-by-dimensionally with the individual dynamic physiological baseline vector updated in the previous step to obtain the absolute offset vector for each physiological dimension. A positive value indicates that the current indicator is higher than the individual baseline, and a negative value indicates that the current indicator is lower than the individual baseline. Then, the continuous physiological indicator standard deviation vector, obtained based on historical sliding window statistics, is retrieved from the smart bracelet's local storage. Each dimension of this standard deviation vector records the historical fluctuation amplitude statistics of parameters such as heart rate and blood oxygen saturation within a preset time window for the wearer. Further, the absolute offset vector is divided dimension-by-dimensionally by the corresponding standard deviation vector to complete the Z-score standardization operation, resulting in a multidimensional physiological deviation feature vector. The values ​​of each dimension of this vector represent the distance by which the current physiological indicator deviates from the individual baseline by how many historical standard deviations. The formula for calculating this standardized deviation metric is as follows: in, The output is a multidimensional physiological deviation feature vector, which represents the standardized deviation of each physiological dimension from the individual baseline at the current moment. It is a dimensionless real number vector. This is the current-moment continuous physiological indicator prediction vector extracted from the multidimensional health indicator set; This is the individual dynamic physiological baseline vector updated in the previous step; This is the standard deviation vector of continuous physiological indicators calculated in real time based on a sliding window of historical data. (Molecular part) The absolute offset of the current continuous physiological indicator prediction vector relative to the individual's dynamic physiological baseline vector was calculated. This offset retains the deviation direction information; a positive value corresponds to an increase in the indicator deviation, and a negative value corresponds to a decrease in the indicator deviation. In the health monitoring scenario of a smart bracelet, a positive deviation in the heart rate dimension may indicate tachycardia or sympathetic nerve excitation, while a negative deviation may indicate bradycardia or increased vagal tone. A negative deviation in the blood oxygen dimension may indicate tissue hypoxia or abnormal respiratory function. (The denominator is missing from the original text.) The absolute offsets of each dimension are normalized to a relative scale based on the wearer's historical fluctuation range. For dimensions with large historical fluctuations (such as a large heart rate standard deviation for active users), the standardized deviation value of the same absolute offset is smaller, indicating that the offset is within the user's normal fluctuation range. For dimensions with small historical fluctuations (such as extremely small blood oxygen standard deviation in healthy adults), even if the absolute offset is small, the standardized deviation value will be significantly amplified, ensuring that minor but clinically significant decreases in blood oxygen can be sensitively detected. Through this standardization operation, each dimension of the multidimensional physiological deviation feature vector is uniformly mapped to a dimensionless standard deviation multiple space, enabling the subsequent threshold judgment step to perform fair and consistent anomaly comparisons across all physiological dimensions using a unified standardized threshold.

[0037] Accordingly, in step S63, the multidimensional physiological deviation feature vector is compared with a preset medical safety boundary threshold dimension by dimension using an indicator function to obtain an abnormal alarm flag set. It should be understood that, since the multidimensional physiological deviation feature vector, after Z-score standardization, has uniformly mapped the deviation of each physiological dimension to a dimensionless standard deviation multiple space, this vector still exists in the form of continuous real numbers and has not yet been transformed into a binary judgment result that can directly trigger terminal warning prompts. The health warning function of the smart bracelet for the wearer needs to provide a clear normal or abnormal judgment for each physiological dimension so that vibration or interface reminders can be promptly issued to the wearer when an abnormality is detected. Therefore, in the technical solution of this application, the multidimensional physiological deviation feature vector is further compared with a preset medical safety boundary threshold dimension by dimension using an indicator function to obtain an abnormal alarm flag set. This transforms the continuous standardized deviation values ​​into discrete binary alarm flags, providing a deterministic judgment result of whether each physiological dimension is abnormal. In this way, the mapping from the continuous deviation space to the discrete alarm space can be completed with minimal computational overhead, providing clear warning status field inputs for the structured encapsulation of subsequent health assessment and early warning reports.

[0038] More specifically, in a specific example of this application, the anomaly threshold judgment adopts a decision strategy of independent threshold comparison for each dimension. First, a corresponding medical safety boundary threshold is preset for each physiological dimension of the multidimensional physiological deviation feature vector. This threshold is determined based on the normal fluctuation range of various physiological parameters in clinical medical literature. For example, the threshold for the heart rate dimension is set to 3 standard deviations, and the threshold for the blood oxygen saturation dimension is set to 2 standard deviations. The thresholds for different dimensions can be independently configured according to their differences in clinical sensitivity. Next, the absolute value of each dimension of the multidimensional physiological deviation feature vector is taken to eliminate the deviation direction information and retain only the deviation magnitude, so that whether the current indicator is positively higher or negatively lower than the baseline, it is included in the anomaly judgment range. Then, the absolute value of the deviation of each dimension is independently compared with its corresponding preset medical safety boundary threshold. If and only if the absolute value of the deviation of a certain dimension exceeds the threshold corresponding to that dimension, the indicator function outputs one to indicate that the dimension triggers an anomaly alarm; otherwise, it outputs zero to indicate that the dimension is within the safe range. Thus, alarm Boolean values ​​are generated dimension by dimension and aggregated into a set of anomaly alarm flags. Further, the calculation formula for this independent decision for each dimension is as follows: in, This is the alarm Boolean value for the i-th physiological dimension in the abnormal alarm flag set, and its value is either zero or one. This is an indicator function; it takes the value 1 when the logical condition inside the parentheses is true, and the value 0 otherwise. This is for absolute value operations; , where is the standardized deviation component of the i-th dimension in the multidimensional physiological deviation feature vector; This is the medical safety boundary threshold preset by the system for the i-th physiological dimension.

[0039] Specifically, in the abnormal threshold judgment mechanism of the first embodiment described above, the system adopts a judgment strategy of independent threshold comparison for each dimension of the multidimensional physiological deviation feature vector. That is, each physiological dimension, such as heart rate deviation and blood oxygen deviation, is independently compared with its own preset fixed medical safety boundary threshold. The alarm flag of a dimension is triggered only when the absolute value of the deviation of a certain dimension exceeds the threshold corresponding to that dimension. This mechanism implicitly assumes a key assumption in its underlying mathematical modeling: the physiological dimensions are independent of each other and do not affect each other, and the abnormal judgment of each dimension can be completed independently under the condition of isolation.

[0040] However, in the real human physiological system, there are significant cross-dimensional physiological coupling relationships between various physiological indicators. Taking heart rate and blood oxygen saturation monitored by a smart bracelet as an example, when the wearer is in an early state of cardiopulmonary dysfunction, blood oxygen saturation will slightly decrease, and the heart will activate a compensatory accelerated pumping mechanism to maintain tissue oxygen supply, resulting in a synchronous increase in heart rate. In this typical compensatory coupling scenario, the absolute values ​​of heart rate deviation and blood oxygen deviation may not exceed their respective independently set safety thresholds, but the coordinated shift pattern of one rising and one falling is precisely an early pathological signal of cardiopulmonary dysfunction. Because the first embodiment makes isolated judgments for each dimension, it cannot perceive this cross-dimensional coordinated shift pattern, causing the system to miss reports when facing such subthreshold coordinated anomalies. At the same time, the first embodiment also faces the problem of contextual false alarms. When the wearer is in a high-temperature environment, the increase in body surface temperature will trigger a compensatory increase in heart rate. At this time, although the heart rate deviation exceeds the independent threshold, combined with the contextual information of the temperature deviation, it can be judged that the increase in heart rate is a normal thermoregulatory physiological response and not a pathological abnormality. However, the original independent decision-making mechanism for each dimension cannot incorporate the deviation information of the temperature dimension into the decision-making process of the heart rate dimension, thus generating false alarms for this normal physiological compensation scenario.

[0041] To address the aforementioned technical deficiencies, a second embodiment is proposed. More specifically, in another specific example of this application, Figure 6 This is a flowchart illustrating the multi-dimensional health data fusion analysis method according to embodiments of this application, which compares the absolute values ​​of multi-dimensional physiological deviation feature vectors dimension by dimension with preset medical safety boundary thresholds using an indicative function to obtain a set of abnormal alarm flags. Figure 6As shown, step S63 includes: S631, encoding the cross-dimensional coupling strength of the multidimensional physiological deviation feature vector to obtain the physiological coupling weight matrix; S632, performing coupling weighted energy aggregation on the multidimensional physiological deviation feature vector based on the physiological coupling weight matrix to obtain the joint coupling anomaly score scalar; S633, comparing the joint coupling anomaly score scalar with the preset joint coupling anomaly threshold through an indicator function to obtain the anomaly alarm flag bit.

[0042] In step S631, the multidimensional physiological deviation feature vector is encoded with cross-dimensional coupling strength to obtain a physiological coupling weight matrix. It should be understood that because the aforementioned decision strategy of independent threshold comparison for each dimension treats each physiological dimension as an independent and isolated channel, it cannot perceive the objectively existing cross-dimensional coupling relationship between different dimensions such as heart rate deviation and blood oxygen deviation in the human physiological system. This leads to false negatives when multiple dimensions are in sub-threshold states but jointly present a pathological coupling pattern in compensatory coupling scenarios, and false positives when a single dimension exceeds the threshold in normal physiological compensation scenarios but belongs to normal responses such as thermoregulation. The fundamental reason is the lack of an explicit quantification encoding mechanism for the magnitude of deviation coordination between dimensions during the decision process. Therefore, in the technical solution of this application, the multidimensional physiological deviation feature vector is further encoded with cross-dimensional coupling strength to obtain a physiological coupling weight matrix, thereby explicitly quantifying and encoding the implicit physiological coupling relationship between dimensions, compensating for the structural defects of isolated decision-making for each dimension. In this way, the system can perceive the strength of physiological coupling between multiple dimensions, providing a quantitative basis for cross-dimensional correlation for subsequent joint decisions, thereby realizing a paradigm shift from independent decision-making in each dimension to globally coupled joint decision-making at the algorithm level.

[0043] In a specific example of this application, after completing the standardized calculation of the multidimensional physiological deviation feature vector, it is necessary to explicitly quantize and encode the implicit physiological coupling relationship between each dimension to compensate for the structural defect of isolated decision-making between each dimension in the first embodiment. First, the multidimensional physiological deviation feature vector output from the previous step is received. This vector is a column vector with dimension D multiplied by 1, where D is the number of physiological monitoring dimensions, and each dimension corresponds to standardized deviation components such as heart rate deviation and blood oxygen deviation. Next, the multidimensional physiological deviation feature vector is subjected to its own outer product operation, that is, the column vector is multiplied by its own transpose to construct a symmetric square matrix with dimension D multiplied by D. Each element in this square matrix is ​​mathematically equal to the product of the deviation components of the corresponding two dimensions, thereby encoding the deviation coordination amplitude information between any two physiological dimensions at the current time in a two-dimensional matrix space. Then, the symmetric matrix is ​​normalized using the Frobenius norm to eliminate the dimensional influence of the absolute amplitude. Specifically, the square root of the sum of the squares of all elements in the matrix is ​​used as the normalization denominator, and each element in the matrix is ​​divided by the norm value, ultimately yielding the physiological coupling weight matrix. The formula for constructing the physiological coupling weight matrix is ​​as follows: in, The output physiological coupling weight matrix is ​​a symmetric square matrix of dimension D×D, where D is the number of physiological monitoring dimensions. The input multidimensional physiological deviation feature vector is a D×1 column vector; The transpose of the multidimensional physiological deviation feature vector is a 1×D row vector; The Frobenius norm operation, i.e., the arithmetic square root of the sum of the squares of all elements of the matrix, is used for normalization to eliminate dimensional differences. It should be noted that the off-diagonal elements in this symmetric matrix precisely characterize the cooperative magnitude of the deviations of the corresponding two physiological dimensions at the current moment. When both the heart rate and blood oxygenation dimensions experience significant shifts simultaneously, the corresponding matrix element values ​​exhibit a secondary amplification effect, thus explicitly encoding the compensatory coupling pattern of heart rate increases accompanied by blood oxygen decreases, which was originally ignored by independent threshold decisions, into a computable value. The diagonal elements, on the other hand, retain the independent deviation energy of each dimension. Through this outer product dimensionality-upgrading operation, the one-dimensional deviation vector is elevated to a two-dimensional coupling matrix, providing a quantitative basis for cross-dimensional correlations in subsequent joint decisions, enabling the system to perceive the strength of physiological coupling between multiple dimensions. Furthermore, from the physical meaning of this formula, the outer product operation... The essence is to expand the pairwise interactions between the components of a one-dimensional deviation vector into a two-dimensional matrix. Taking the heart rate deviation component and the blood oxygen deviation component as examples, when the wearer is in an early state of cardiopulmonary dysfunction, the heart rate deviation is positive (compensatory increase) and the blood oxygen deviation is negative (insufficient oxygen supply decrease), and their product is negative. The absolute magnitude of this negative value directly reflects the strength of this compensatory coupling mode. However, when the wearer is in a high-temperature environment, the heart rate and body temperature rise in the same direction, and their product is positive. This coupling mode differs fundamentally from the pathological compensatory mode in its sign characteristics. (Flobenius norm normalization operation) By scaling the absolute magnitude of all elements in the matrix to a unit energy scale, the differences in matrix numerical scale caused by the different absolute magnitudes of the wearer's current deviation are eliminated. This ensures that the physiological coupling weight matrix only encodes the distribution pattern of the relative coupling strength between each dimension, rather than the absolute deviation magnitude information. As a result, the subsequent quadratic joint scoring calculation can run stably in the normalized weight space, unaffected by fluctuations in the deviation magnitude.

[0044] In step S632, the multidimensional physiological deviation feature vectors are coupled and weighted to aggregate energy based on the physiological coupling weight matrix to obtain a joint coupled anomaly score scalar. It should be understood that although the physiological coupling weight matrix encodes the cross-dimensional coupling strength relationship between each dimension into a two-dimensional matrix, this matrix itself is still a D-by-D square matrix structure. It has not yet undergone joint aggregation with the actual deviation magnitude of each dimension and cannot be directly used for threshold comparison in anomaly judgment. It is necessary to jointly aggregate the cross-dimensional coupling information with the deviation information of each dimension to generate a unified score scalar that comprehensively reflects the global anomaly degree. This scalar contains both the independent deviation energy of each dimension and the cross-coupling energy between any two dimensions, thus providing a single comprehensive criterion input for subsequent joint threshold judgment. Therefore, in the technical solution of this application, the multidimensional physiological deviation feature vectors are further coupled and weighted to aggregate energy based on the physiological coupling weight matrix to obtain a joint coupled anomaly score scalar. This compresses the multidimensional deviation information and the cross-dimensional coupling weight information into a single global anomaly degree measurement scalar through quadratic form operations. In this way, the independent bias energy of each dimension and the cross-dimensional cross-coupling energy can be encoded simultaneously in a single scalar value. This allows the subsequent decision-making process to complete the joint determination of global coupling anomalies by simply comparing the threshold of a single scalar, fundamentally overcoming the structural defect that independent decision-making in each dimension cannot perceive the collaborative offset mode.

[0045] In a specific example of this application, after obtaining the physiological coupling weight matrix, it is necessary to jointly aggregate the cross-dimensional coupling information with the deviation information of each dimension to generate a unified score that comprehensively reflects the global degree of abnormality. First, the physiological coupling weight matrix and the multi-dimensional physiological deviation feature vector output from the previous step are received. The physiological coupling weight matrix is ​​a normalized symmetric square matrix of D multiplied by D, and the multi-dimensional physiological deviation feature vector is a column vector of D multiplied by 1. Next, a quadratic operation is performed between the multi-dimensional physiological deviation feature vector and the physiological coupling weight matrix, i.e., the deviation vector transpose is multiplied by the coupling weight matrix and then multiplied by the deviation vector itself. Mathematically, this operation compresses the product of vector, matrix, and vector into a single scalar value. Then, this scalar value is output as the joint coupling anomaly score scalar to the subsequent joint threshold decision stage. The calculation formula for the joint coupling anomaly score is as follows: in, The output joint coupling anomaly score scalar is a non-negative real number. The larger the value, the more severe the cross-dimensional joint anomaly. This is the transpose row vector of the multidimensional physiological deviation feature vector; This is the physiological coupling weight matrix output from the previous step; This is a column vector of multidimensional physiological deviation feature vectors. It should be noted that the quadratic form operation not only accumulates the squared energy of each dimension's deviation (diagonal term contribution), but more importantly, it superimposes the cross-coupling energy between any two dimensions (off-diagonal term contribution). In typical scenarios of early cardiopulmonary dysfunction, heart rate deviation is positive (compensatory increase) and blood oxygen deviation is negative (insufficient oxygen supply decrease). While neither may exceed an independent threshold, the cross-coupling term in the quadratic form transforms this pathological coupling pattern of increase and decrease into a significant positive energy contribution, greatly increasing the joint score and effectively capturing cooperative subthreshold abnormalities that cannot be identified by independent dimension-by-dimensional judgment. Conversely, for scenarios involving normal physiological compensation (such as heart rate and temperature increasing in the same direction under high temperature), the contribution pattern of the cross-coupling term differs fundamentally from the pathological pattern, and the joint score does not show an abnormal increase, thus naturally suppressing contextual false alarms at the algorithm level. Through this quadratic form aggregation operation, the system achieves a paradigm shift from independent dimension-by-dimensional judgment to globally coupled joint judgment.

[0046] Furthermore, from the physical meaning of this formula, quadratic form operations... This can be expanded into a superposition of diagonal and off-diagonal terms. The diagonal term corresponds to the sum of the products of the elements on the main diagonal of the physiological coupling weight matrix and the squared values ​​of the deviations of each dimension. Physically, it represents the weighted sum of the independent deviation energies of each physiological dimension, preserving the basic information of the degree of deviation of each dimension in traditional dimensional decision-making. The off-diagonal term corresponds to the sum of the products of the elements on the off-diagonal of the physiological coupling weight matrix and the deviation components of the corresponding two dimensions. Physically, it represents the weighted energy contribution of the coupling amplitude of the deviation synergy between any two physiological dimensions. This part is the core increment of this scheme compared to dimensional independent decision-making. When the heart rate dimension deviation component is positive and the blood oxygen dimension deviation component is negative, their product is negative. The corresponding off-diagonal elements in the physiological coupling weight matrix, after outer product operation, also encode the strength information of this negative coupling mode. The contribution after multiplication is transformed into a positive increase in the joint score through a quadratic accumulation mechanism, enabling this compensatory pathological coupling mode to achieve a significant numerical expression in the joint score scalar. When the heart rate dimension and body temperature dimension increase in the same direction, their product is positive. The encoding method of their coupling mode in the physiological coupling weight matrix is ​​fundamentally different from that of the compensatory pathological mode at the symbol level. After quadratic accumulation, it will not produce the same score increase effect as the pathological mode. Thus, the joint coupling abnormal score scalar naturally has the ability to distinguish between pathological synergistic abnormalities and normal physiological compensation.

[0047] In step S633, the joint coupling anomaly score scalar is compared with a preset joint coupling anomaly threshold using an indicator function to obtain an anomaly alarm flag. It should be understood that although the joint coupling anomaly score scalar aggregates the independent deviation energy of each dimension and the cross-dimensional cross-coupling energy into a single global anomaly degree metric, this scalar is still in continuous real number form and has not yet been converted into a binary judgment result that can directly trigger a terminal warning. The health warning function of the smart bracelet for the wearer needs to provide a clear yes or no judgment on whether there is a cross-dimensional collaborative pathological shift at the current moment, so as to promptly issue a warning reminder to the wearer when a joint anomaly is detected. Therefore, in the technical solution of this application, the joint coupling anomaly score scalar is further compared with a preset joint coupling anomaly threshold using an indicator function to obtain an anomaly alarm flag, thereby converting the continuous joint score value into a discrete binary alarm flag, completing the mapping from the global coupling anomaly score space to the final alarm decision space. In this way, the coupled score scalar can be used as a unified decision input to replace the decision method of independent comparison of each dimension. This allows the system to trigger an early warning by raising the coupled score even when none of the dimensions exceed the independent threshold, effectively solving the problem of missed detection of collaborative anomalies. At the same time, for single-dimensional cases that exceed the threshold due to normal physiological compensation, the joint score will not rise abnormally and will remain silent, thereby suppressing contextual false alarms.

[0048] In a specific example of this application, after obtaining the joint coupling anomaly score scalar, the final anomaly judgment and report output need to be completed based on this unified score. First, the joint coupling anomaly score scalar output from the previous step is received. This scalar is a non-negative real number; the larger the value, the more severe the cross-dimensional joint anomaly at the current moment. Next, the joint coupling anomaly score scalar is compared with a preset joint coupling anomaly threshold. This threshold is determined based on the statistical distribution of joint scores in historical healthy individuals. Specifically, a large number of joint coupling anomaly score samples from healthy wearers under normal physiological conditions are collected, and the upper quantile of their statistical distribution is calculated as the basis for threshold setting. Then, when the joint coupling anomaly score scalar exceeds the threshold, the coupling anomaly alarm flag is activated to one using an indicative function, indicating a cross-dimensional synergistic pathological shift at the current moment. When the joint coupling anomaly score scalar does not exceed the threshold, the indicative function outputs zero, indicating that the joint shift pattern of each dimension is within the normal range at the current moment. Further, the multi-dimensional core health indicator group and the coupling anomaly alarm flag are encapsulated in a structured dictionary nesting, outputting a comprehensive health assessment and early warning report for end users. The calculation formula for the coupled sensing joint decision is as follows: in, The generated coupling anomaly alarm flag is a Boolean scalar with a value of 0 or 1, where 1 indicates that there is a cross-dimensional collaborative pathological offset at the current moment; This is an indicator function; it takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This is the joint coupling anomaly score scalar output from the previous step; The pre-defined threshold for joint coupling anomalies is determined based on the statistical distribution of joint scores in historical healthy individuals. The final output is a comprehensive health assessment and early warning report; Encapsulate operators for structured dictionaries; This is a multidimensional core health indicator group. From the physical meaning of this formula, the indicator function... The decision logic essentially sets a decision boundary on the one-dimensional axis of the jointly coupled anomaly scoring scalar. This boundary divides the scoring space into two mutually exclusive subsets: a normal region and an anomaly region. Jointly Coupled Anomaly Threshold As the location parameter of this decision boundary, its value directly affects the balance between the sensitivity and specificity of the system. Setting the threshold too low will cause normal physiological fluctuations to be misjudged as abnormal, thereby increasing the false alarm rate. Setting the threshold too high will cause mild synergistic anomalies to be undetectable, thereby increasing the false negative rate. In this example, the threshold is set based on the 95th percentile of the statistical distribution of the joint score of historical healthy people, ensuring that the probability of the joint score exceeding the threshold under normal physiological conditions does not exceed 5%.

[0049] Compared to the independent comparison method in the first embodiment, the improved joint threshold decision uses the coupled score scalar as a unified decision input. This allows the system to trigger an alarm by increasing the coupled score even when none of the dimensions exceed the independent threshold, effectively solving the problem of missed alarms for coordinated anomalies. In early cardiopulmonary function abnormalities, heart rate deviation and blood oxygen deviation are both within their independent thresholds, and independent judgment in each dimension will not trigger any alarms. However, after cross-dimensional coupling strength encoding and quadratic energy aggregation, the joint coupled anomaly score scalar increases significantly due to the positive energy contribution of the cross-coupling terms and exceeds the joint coupled anomaly threshold. The indicator function sets the coupling anomaly alarm flag to one, successfully capturing this coordinated subthreshold anomaly. Meanwhile, for single-dimensional threshold exceedances due to normal physiological compensation, such as a scenario where compensatory increases in heart rate under high temperature conditions trigger an alarm for independent judgment in the heart rate dimension, the joint coupled anomaly score scalar will not increase abnormally because the contribution pattern of the cross-terms in the joint score is fundamentally different from the pathological pattern. The indicator function outputs zero, and the system remains silent, thus suppressing contextual false alarms. Finally, the multidimensional core health indicator group and coupled anomaly alarm flags are integrated into a comprehensive health assessment and early warning report through a structured dictionary encapsulation operator. This report includes a complete snapshot of the wearer's current health status and a joint anomaly judgment result based on cross-dimensional coupled perception, serving as the final output of the multidimensional health data fusion analysis on the smart bracelet side.

[0050] By introducing a complete technical chain—including cross-dimensional physiological coupling weight matrix construction, quadratic joint coupling anomaly score calculation, and coupled perception-based joint threshold judgment—the improved mechanism achieves the following technical effects in multi-dimensional health monitoring scenarios of smart bracelets: In terms of collaborative anomaly detection, when multiple physiological dimensions are in sub-threshold states but jointly exhibit a pathological coupling pattern (such as compensatory increase in heart rate accompanied by a slight decrease in blood oxygen), the cross-term in the quadratic operation can transform this cross-dimensional collaborative shift into a significant increase in the joint score, thereby triggering an alert and effectively reducing the false alarm rate of collaborative anomalies. In terms of false alarm suppression, for single-dimensional over-threshold scenarios that belong to normal physiological compensatory responses (such as compensatory increase in heart rate under high temperature conditions), since the contribution pattern of the cross-term in the joint score is fundamentally different from the pathological pattern, the system will not generate false alarms, effectively reducing the contextual false alarm rate. Overall, the improved mechanism, by upgrading the original dimension-by-dimensional independent judgment paradigm to a globally coupled joint judgment paradigm, significantly improves the sensitivity and specificity of health anomaly warnings on the smart bracelet side without adding additional sensor hardware, enhancing the clinical auxiliary reference value of wearable devices in complex real-world physiological scenarios.

[0051] Accordingly, in step S64, the multidimensional health indicator set and the abnormal alarm flag set are structurally encapsulated to obtain a health assessment and early warning report. It should be understood that since the multidimensional health indicator set and the abnormal alarm flag set originate from different processing stages, the multidimensional health indicator set contains detailed health status measurement data such as continuous physiological indicator prediction vectors and discrete state classification labels, while the abnormal alarm flag set contains independent judgment results for each dimension and joint abnormal judgment results based on cross-dimensional coupled perception. These two sets differ in data type and semantic level. If they are not standardized and integrated within a unified data structure, the smart bracelet's terminal display module and the cloud-based health management platform will be unable to simultaneously obtain the wearer's detailed health measurement values ​​and corresponding abnormal warning status using a standardized data interface. This results in fragmented presentation of health information and an inability to support complete health trend retrospectives and medical data integration. Therefore, in the technical solution of this application, the multidimensional health indicator set and the abnormal alarm flag set are further structurally encapsulated to integrate detailed health status measurement data and abnormal warning judgment results into a self-contained standardized report structure under a unified time index. This provides a unified and semantically complete final output data packet for the smart bracelet's terminal display interface, Bluetooth data transmission protocol, and cloud-based health management platform, ensuring that the complete data link from data collection to alert presentation forms a closed loop in this step.

[0052] More specifically, in a concrete example of this application, firstly, the absolute timestamp corresponding to the current system's decision is obtained as the primary key identifier for this health assessment and early warning report. This timestamp is consistent with the reference clock used in the spatiotemporal alignment preprocessing in the preceding steps, ensuring that all data fields in the report can be traced back to the same acquisition time. Next, a multidimensional health indicator group is written as a measurement data field into the structured report body. This field includes the real-time heart rate estimate and blood oxygen saturation estimate in the continuous physiological indicator prediction vector at the current time, as well as the heart rhythm status determination result and sleep stage determination result in the discrete state classification label, providing the wearer and medical personnel with a complete multidimensional health status snapshot. Then, an abnormal alarm flag set is written as an early warning status field into the same structured report body. This field includes dimension-by-dimensional independent decision alarm flags for each physiological dimension and coupled abnormal alarm flags generated based on cross-dimensional coupled perception, providing a clear alarm trigger basis for the terminal early warning module. Furthermore, the timestamp primary key, measurement data field, and warning status field are encapsulated into a complete structured data dictionary instance to obtain a health assessment warning report. This report serves as the final output of the multi-dimensional health data fusion and analysis process on the smart bracelet side. It can be transmitted to the paired mobile terminal via Bluetooth Low Energy protocol for interface rendering and historical storage, or uploaded to the cloud health management platform via wireless network for long-term trend analysis and remote medical data docking, thus completing a complete data processing closed loop from the acquisition of raw data from multiple sensors to the output of a structured health warning report.

[0053] In summary, a multi-dimensional health data fusion and analysis method according to embodiments of this application is clarified. First, it performs spatiotemporal alignment and frequency domain denoising preprocessing on multi-source heterogeneous physiological motion data concurrently collected by a smart bracelet. Then, while extracting multi-branch features, it performs real-time quality quantization of the pulse wave signal based on the ratio of fundamental energy to full-band energy, generating a dynamic quality index. During the fusion stage, this quality index is used as a dynamic scaling mask factor in a cross-modal attention mechanism. When the quality of the core physiological signal decreases, the attentional guidance weight of the high-quality auxiliary modality on the damaged physiological features is automatically increased, achieving semantic-level feature self-healing repair, rather than simple discarding or static weighting. Finally, a multi-task learning network maps multi-dimensional health indicators, and combined with an individual dynamic baseline, performs coupled perception for joint anomaly judgment, thereby significantly improving the accuracy and continuity of health analysis under strong interference environments without increasing hardware requirements.

[0054] Furthermore, a multi-dimensional health data fusion and analysis system is also provided.

[0055] Figure 7 This is a block diagram of a multi-dimensional health data fusion and analysis system according to an embodiment of this application. Figure 7As shown, a multi-dimensional health data fusion analysis system 100 according to an embodiment of this application includes: a data acquisition and preprocessing module 110, used to concurrently acquire and spatiotemporally align preprocessing multi-source physiological motion data streams through multiple sensors built into a smart bracelet to obtain a pulse wave preprocessing sequence, an inertial motion preprocessing sequence, and a body surface temperature preprocessing sequence; a modal feature extraction module 120, used to perform modal feature extraction on the pulse wave preprocessing sequence, inertial motion preprocessing sequence, and body surface temperature preprocessing sequence through a multi-branch feature extraction network to obtain a pulse wave feature matrix, a motion rhythm feature matrix, and a body surface context feature vector; and a signal quality dynamic quantization module 130, used to perform pulse... The pulse wave preprocessing sequence undergoes dynamic signal quality quantization to obtain the pulse wave quality index; the cross-modal semantic repair and fusion module 140 uses the pulse wave quality index as a dynamic scaling factor to perform quality-aware cross-modal semantic repair and fusion on the pulse wave feature matrix, the motion rhythm feature matrix, and the body surface context feature vector to obtain a cross-modal fused feature tensor; the health status index inverse mapping module 150 uses a multi-task learning network to perform health status index inverse mapping on the cross-modal fused feature tensor to obtain a multi-dimensional health index group; and the deviation judgment module 160 performs deviation judgment on the multi-dimensional health index group based on the individual dynamic physiological baseline to obtain a health assessment early warning report.

[0056] Those skilled in the art will understand that the steps, measures, and schemes in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted; furthermore, other steps, measures, and schemes in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted; furthermore, the steps, measures, and schemes in the prior art that are similar to those disclosed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0057] The above-described embodiments are merely illustrative of several implementation methods of this disclosure, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent for the embodiments of this disclosure. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this disclosure, and these all fall within the protection scope of the embodiments of this disclosure. Therefore, the protection scope of the embodiments of this disclosure should be determined by the appended claims. As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and detail can be made without departing from the spirit and scope of the present invention as defined in the appended claims.

[0058] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for multi-dimensional health data fusion and analysis, characterized in that, include: S1 uses multiple sensors built into the smart bracelet to concurrently acquire and spatiotemporally align preprocessing of multi-source physiological motion data streams to obtain pulse wave preprocessing sequences, inertial motion preprocessing sequences, and body surface temperature preprocessing sequences. S2, through a multi-branch feature extraction network, modal features are extracted from the pulse wave preprocessing sequence, inertial motion preprocessing sequence and body surface temperature preprocessing sequence to obtain the pulse wave feature matrix, motion rhythm feature matrix and body surface context feature vector; S3, Dynamic quantization of signal quality is performed on the preprocessed pulse wave sequence to obtain the pulse wave quality index; S4. Using the pulse wave quality index as a dynamic scaling factor, the pulse wave feature matrix, the motion rhythm feature matrix, and the body surface context feature vector are fused together with cross-modal semantic repair based on quality perception to obtain the cross-modal fused feature tensor. S5 uses a multi-task learning network to perform inverse mapping of health status indicators on cross-modal fusion feature tensors to obtain a multi-dimensional health indicator set. S6, based on an individual's dynamic physiological baseline, makes a deviation judgment on a multidimensional health indicator group to obtain a health assessment and early warning report.

2. The multi-dimensional health data fusion and analysis method according to claim 1, characterized in that, The multi-source physiological motion data stream includes photoplethysmography (PPG) signals, six-axis inertial motion signals, and skin temperature signals.

3. The multi-dimensional health data fusion and analysis method according to claim 1, characterized in that, Step S1 includes: Using the high-frequency hardware clock timestamp of the six-axis inertial motion signal as the reference clock, asynchronous timestamp alignment and resampling mapping are performed on the photoplethysmography pulse wave signal and skin temperature signal to obtain the pulse wave aligned sequence, inertial motion aligned sequence and body surface temperature aligned sequence. By using an infinite impulse response digital bandpass filter, the pulse wave aligned sequence, the inertial motion aligned sequence, and the body surface temperature aligned sequence are subjected to forward and reverse zero-phase frequency domain noise suppression filtering respectively to obtain the pulse wave preprocessed sequence, the inertial motion preprocessed sequence, and the body surface temperature preprocessed sequence.

4. The multi-dimensional health data fusion and analysis method according to claim 1, characterized in that, Step S2 includes: Based on a one-dimensional temporal convolutional network, a sliding window convolutional feature mapping is performed on the pulse wave preprocessing sequence to obtain the pulse wave feature matrix. Multi-dimensional feature reduction and aggregation are performed on the inertial motion preprocessing sequence to obtain the motion rhythm feature matrix; The surface temperature preprocessed sequence is subjected to environmental context deep feature extraction to obtain the surface context feature vector.

5. The multi-dimensional health data fusion and analysis method according to claim 1, characterized in that, Step S3 includes: dynamically quantizing the signal quality of the pulse wave preprocessing sequence using the following formula: in, The pulse wave quality index, Time windows in pulse wave preprocessing sequences Inside The sum of the fundamental wave signal energy, For time windows Total energy across the entire frequency band of the internal pulse wave sequence. This is the fundamental frequency energy ratio threshold bias term. This is a hyperparameter for steepness.

6. The multi-dimensional health data fusion and analysis method according to claim 1, characterized in that, Step S4 includes: The motion rhythm feature matrix and the body surface context feature vector are concatenated along the feature channel dimension to obtain the concatenation result. The pulse wave feature matrix is ​​projected into a query matrix, and the concatenation result is projected into a key matrix and a value matrix, respectively. A dynamic masking factor constructed based on the inversion value of the pulse wave quality index is used to perform quality-aware dynamic masking modulation and semantic-level self-healing fusion on the query matrix, key matrix, and value matrix to obtain a cross-modal fusion feature tensor.

7. The multi-dimensional health data fusion and analysis method according to claim 1, characterized in that, Step S5 includes: Shared low-level feature extraction and multi-task branch prediction are performed on the cross-modal fusion feature tensor to obtain continuous physiological index prediction vectors and discrete state probability vectors; The discrete state probability vector is extracted using the maximum a posteriori probability index to obtain the discrete state classification label; The prediction vectors of continuous physiological indicators and the classification labels of discrete states are structured and encapsulated according to timestamps to obtain a multidimensional health indicator group.

8. The multi-dimensional health data fusion and analysis method according to claim 1, characterized in that, Step S6 includes: Historical physiological data is iteratively updated to obtain an individual's dynamic physiological baseline vector; The continuous physiological indicators in the multidimensional health indicator group are dynamically deviated from the individual dynamic physiological baseline vector to obtain the multidimensional physiological deviation feature vector. The abnormal alarm flag set is obtained by comparing the absolute values ​​of the multidimensional physiological deviation feature vector with the preset medical safety boundary threshold dimension by dimension using the indicator function. A health assessment and early warning report is obtained by structurally encapsulating a set of multidimensional health indicators and abnormal alarm flags.

9. The multi-dimensional health data fusion and analysis method according to claim 8, characterized in that, The abnormal alarm flag set is obtained by comparing the absolute values ​​of the multidimensional physiological deviation feature vector with the preset medical safety boundary threshold dimension by dimension using an indicator function, including: Cross-dimensional coupling strength encoding is performed on the multidimensional physiological deviation feature vector to obtain the physiological coupling weight matrix; Based on the physiological coupling weight matrix, the multidimensional physiological deviation feature vector is coupled and weighted to aggregate energy to obtain the joint coupled abnormality score scalar. The joint coupling anomaly score scalar is compared with the preset joint coupling anomaly threshold by an indicator function to obtain the anomaly alarm flag.

10. A multi-dimensional health data fusion and analysis system, characterized in that, include: The data acquisition and preprocessing module is used to concurrently acquire and spatiotemporally align and preprocess multi-source physiological motion data streams through multiple sensors built into the smart bracelet to obtain pulse wave preprocessing sequences, inertial motion preprocessing sequences, and body surface temperature preprocessing sequences. The modal feature extraction module is used to extract modal features from the pulse wave preprocessing sequence, inertial motion preprocessing sequence, and body surface temperature preprocessing sequence through a multi-branch feature extraction network to obtain the pulse wave feature matrix, motion rhythm feature matrix, and body surface context feature vector. The signal quality dynamic quantization module is used to perform dynamic signal quality quantization on the pulse wave preprocessing sequence to obtain the pulse wave quality index. The cross-modal semantic repair and fusion module is used to perform quality-aware cross-modal semantic repair and fusion on the pulse wave feature matrix, the motion rhythm feature matrix, and the body surface context feature vector using the pulse wave quality index as a dynamic scaling factor to obtain the cross-modal fused feature tensor. The Health Status Indicator Inverse Mapping Module is used to perform health status indicator inverse mapping on cross-modal fusion feature tensors through a multi-task learning network to obtain a multi-dimensional health indicator set. The deviation judgment module is used to make deviation judgments on a group of multidimensional health indicators based on an individual's dynamic physiological baseline in order to obtain a health assessment and early warning report.

Images