Health data intelligent early warning method and system

Abstract

The invention relates to the technical field of digital health, in particular to an intelligent health data early warning method and system.The method comprises the steps that multi-channel signals such as electrocardio signals, pulse oximeter signals and continuous blood pressure signals which are synchronously filtered are injected into an erbium-doped annular cavity through light modulation to generate delay chaos, and sparse binary vectors are obtained through frequency domain hash coding; the vector predicts a future cardiovascular state through a spiking neural network and a reversible ordinary differential equation network; inputting the continuous prediction frame into a reversible tensor flow network, and obtaining a risk vector through anti-fact disturbance; splicing risks, states, codes and contexts to calculate risk energy, and outputting personalized intervention actions by means of a diffusion generation model; and synchronously updating model parameters and coupling coefficients according to the rewards. According to the invention, millisecond-level prospective early warning, low-power-consumption operation and interpretable intervention are realized, and the method is suitable for long-term wearable health monitoring.

Description

Technical Field

[0001] The present invention relates to the field of digital health technology, and in particular to a health data intelligent early warning method and system. Background Art

[0002] Health data early warnings can capture abnormal physiological trends before symptoms appear, providing an opportunity for "second-level" intervention for high-risk chronic diseases like cardiovascular disease, significantly reducing out-of-hospital sudden death rates, alleviating medical resource constraints, and spurring upgrades in the wearables industry. Existing solutions mostly rely on: Cloud-based algorithms based on thresholds or simple convolutional-recurrent networks, which often suffer from high false negatives due to their neglect of signal synchronization and physical properties; purely digital feature extraction, which requires massive floating-point multiplications and accumulations, resulting in high power consumption and latency on the wearable side; black-box deep models, which lack causal explanations and hinder precise intervention planning; and one-way data flow, which prevents real-time correction of sensor drift and leads to poor long-term stability. Summary of the Invention

[0003] In response to the many problems existing in the above-mentioned existing technologies, the present invention provides an intelligent early warning method and system for health data. The present invention uses optical chaos to map synchronous physiological signals into a sparse binary space, uses neuromorphic-reversible dynamics to predict future states, and then calculates the risk gradient with reversible tensor flow. Intervention actions are generated through a diffusion model, and the reward-gradient is synchronously written back to the light, pulse and dynamics layers.

[0004] A health data intelligent early warning method, comprising: The synchronously filtered ECG, photoplethysmography, continuous blood pressure, triaxial acceleration, skin temperature and ambient oxygen partial pressure signals are optically modulated and injected into an erbium-doped ring cavity to generate a time-delayed feedback chaotic waveform, and the chaotic waveform is subjected to frequency domain hash coding to obtain a sparse binary vector; The sparse binary vector is converted into an event sequence and input into a pulse neural network to generate a current vector, which is used as the control variable to input the reversible ordinary differential equation network integral to obtain the predicted cardiovascular state vector, and the coupling coefficient of the erbium-doped ring cavity is adjusted accordingly. The cardiovascular state vectors predicted by the most recent multiple frames are stacked and positionally encoded, then fed into a reversible tensor flow network to generate an encoding vector. The risk vector is obtained through counterfactual perturbations and a classifier, and the parameters of the reversible ordinary differential equation network are updated using the risk vector gradient. The risk vector, predicted cardiovascular state vector, encoding vector and contextual information are concatenated to form a state vector. The risk energy is calculated. The risk energy gradient is used in the diffusion generation model to generate an action latent vector and decode it into an intervention action. After the intervention action is executed, the diffusion generation model parameters, the risk weights of each cardiovascular event, the reversible ordinary differential equation network parameters and the erbium-doped ring cavity coupling coefficient are synchronously updated according to the reward function.

[0005] Preferably, pulse-second signal calibration is performed on the synchronously filtered electrocardiogram signal, photoelectric volume pulse wave signal, continuous blood pressure signal, three-axis acceleration signal, skin temperature signal and ambient oxygen partial pressure signal, and Butterworth filtering and wavelet denoising are performed in sequence after synchronization to obtain a denoised signal for light modulation.

[0006] Preferably, the frequency domain hash coding performs equally spaced quantization on the real part of the spectrum of the time-delayed feedback chaotic waveform, maps the quantized segments with positive signs to activation bits in a sparse binary vector, and maps the quantized segments with negative signs to zero bits to form a sparse binary vector of fixed length.

[0007] Preferably, the spiking neural network is composed of leaky integrate-and-release neurons, which dynamically adjust synaptic weights using time-dependent synaptic plasticity rules driven by event sequences and output current vectors through integration.

[0008] Preferably, the reversible ordinary differential equation network includes a reversible feedforward coupling structure, and a zero-curl regularization constraint is imposed on the vector field during the integration process to maintain bidirectional mapping consistency of the numerical integration.

[0009] Preferably, a preset number of continuous predicted cardiovascular state vectors are stacked in chronological order, position coding is added, and then input into a reversible tensor flow network to generate a coding vector.

[0010] Preferably, the counterfactual perturbation is generated by applying a fixed amplitude offset to a single channel of the encoding vector, while keeping the values ​​of the remaining channels unchanged.

[0011] Preferably, the diffusion generation model adaptively schedules the noise amplitude according to the risk energy in the forward diffusion stage, and completes the reconstruction of the action latent vector within a predetermined number of steps in the reverse denoising stage.

[0012] Preferably, the reward function consists of a cardiovascular event risk energy term and an intervention action cost term, and the synchronous update iteratively adjusts the diffusion generation model parameters, the risk weights of each cardiovascular event, the reversible ordinary differential equation network parameters and the erbium-doped ring cavity coupling coefficient through a proximal strategy optimization algorithm.

[0013] A health data intelligent early warning system, used to implement the health data intelligent early warning method, comprising: An optical encoding module is used to optically modulate the synchronously filtered ECG signal, photoplethysmography signal, continuous blood pressure signal, triaxial acceleration signal, skin temperature signal, and ambient oxygen partial pressure signal, and inject the optically modulated signal into an erbium-doped ring cavity to generate a time-delayed feedback chaotic waveform, and then perform frequency-domain hash coding on the chaotic waveform to obtain a sparse binary vector. a pulse prediction module, configured to convert the sparse binary vector into an event sequence and input the event sequence into a spiking neural network to generate a current vector, input the current vector as a control variable into a reversible ordinary differential equation network integral to obtain a predicted cardiovascular state vector, and adjust the erbium-doped ring cavity coupling coefficient according to the difference between the predicted cardiovascular state vector and a synchronously measured value; a causal assessment module, configured to stack the predicted cardiovascular state vectors of the most recent multiple frames in chronological order, add position encoding, and then input the stacked vectors into a reversible tensor flow network to generate an encoding vector; apply counterfactual perturbations to the encoding vectors and pass them through a classifier to obtain a risk vector; and use the gradient corresponding to the risk vector as backpropagation to update the parameters of the reversible ordinary differential equation network; An intervention generation module is configured to concatenate the risk vector, the predicted cardiovascular state vector, the encoding vector, and the context information to form a state vector, calculate the risk energy, use the gradient of the risk energy in a diffusion generation model to generate an action latent vector and decode it into an intervention action, and after executing the intervention action, synchronously update the diffusion generation model parameters, the risk weights of each cardiovascular event, the reversible ordinary differential equation network parameters, and the erbium-doped ring cavity coupling coefficient according to a reward function.

[0014] Compared with the existing technology, the advantages and beneficial effects of the present invention are: through the optical delayed feedback chaos + frequency domain hashing technology, high-dimensional amplification and sparse coding effects of multimodal physiological signals are achieved under milliwatt power consumption; through the pulse neural network coupled reversible ordinary differential equation network technology, the continuous cardiovascular state prediction effect of 20 seconds ahead at the microsecond level is achieved; through the reversible tensor flow counterfactual inference technology, an explainable event-level risk quantification effect is achieved; through the risk energy modulation diffusion generation technology, the risk-driven, cost-constrained personalized intervention action generation effect is achieved; through the gradient closed-loop update of the annular cavity coupling coefficient technology, the sensing-computing-physics three-layer self-calibration effect is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 Schematic diagram of the interaction of the system of the present invention. DETAILED DESCRIPTION

[0016] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0017] like Figure 1 As shown, a health data intelligent early warning method includes: The synchronously filtered ECG, photoplethysmography, continuous blood pressure, triaxial acceleration, skin temperature and ambient oxygen partial pressure signals are optically modulated and injected into an erbium-doped ring cavity to generate a time-delayed feedback chaotic waveform, and the chaotic waveform is subjected to frequency domain hash coding to obtain a sparse binary vector; In the present invention, the synchronously filtered ECG signal, photoplethysmography signal, continuous blood pressure signal, triaxial acceleration signal, skin temperature signal, and ambient oxygen partial pressure signal are collectively considered a multi-channel timing sequence. After clock calibration at the acquisition end, each channel enters the same analog front end, maintaining phase jitter at the microsecond level, meeting the synchronization requirements of subsequent optical modulation. To reduce hardware resource consumption, the present invention utilizes a single optical modulation path: first, the amplitude of each channel is normalized to the 0–1 range, and then a single-channel baseband drive signal is generated through a programmable weighted summation circuit. This baseband signal drives a silicon-based Mach-Zehnder modulator, mapping the amplitude to the intensity modulation depth of a continuous laser with a wavelength of 1550 nm. Phase modulation is not used in the present invention because intensity modulation maintains sufficient linearity for continuous cardiovascular monitoring and makes it easier to control the optical gain of the erbium-doped ring cavity.

[0018] The erbium-doped ring cavity is a closed-loop amplified optical fiber structure with a ring length of about 50m, corresponding to a basic optical delay of about 250ns. The highest effective frequency of human physiological signals is much lower than this delay, so the light field in the ring can map the multi-channel acquisition sequence to the high-dimensional phase space of the delayed feedback system and generate a chaotic waveform. The chaotic waveform has two major advantages: first, it has an amplifying effect on slight input perturbations, which can improve the instantaneous resolution at the physical level; second, the fiber distributed delay provides natural high-dimensional support for feature expansion, so that subsequent feature extraction can cover broadband information without the need for large-scale convolution kernels. In order to ensure the stability of the chaotic waveform, the present invention controls the ring coupling coefficient and the fiber amplifier pump current so that the light field is in the steady-state chaotic region rather than the self-oscillation region.

[0019] The chaotic waveform is sampled digitally at 1Gsps and then enters the fast Fourier transform process. Assume that the continuous sampling sequence is , which is expressed in the discrete frequency domain as To obtain a fixed-length and sparse binary vector, quantized hash coding is performed: the quantization step size is used. right Perform equal interval segmentation and round the segmentation results to get the index .like If the number is odd, the corresponding bit is set to 1; otherwise, it is set to 0, ultimately forming a sparse binary vector. This process requires only a single comparison operation to determine the active bit, making it ideal for low-power edge devices. Experimental results show that when the window length is 1024 points and the quantization step size is selected via grid search, the sparse binary vector length is 512 bits, and the average active bit range is 30–70, which maintains state discrimination while avoiding information loss due to excessive sparsity.

[0020] This sparse binary vector plays a dual role in the present invention. First, it is directly fed into the spiking neural network as the event sequence input. The sparse structure inherently reduces the pulse rate, significantly reducing the overall power consumption of the spiking neural network. Second, the Boolean structure of the sparse binary vector provides a robust initial condition for the reversible ordinary differential equation network, making it easier for the latter to learn differential relations involving multivariable coupling. Test results show that compared to convolutional feature methods, this method can maintain extremely low latency without increasing the computational effort.

[0021] This invention uses a sparse binary vector of activation bits to implement closed-loop control of the coupling coefficient of an erbium-doped ring cavity. When the activation bit is below a threshold, the coupling coefficient is slightly increased to enhance the chaotic amplitude; when the activation bit exceeds the threshold, the coupling coefficient is reduced to avoid gain saturation. The measured activation bit distribution fluctuates within a range of 30–70, maintaining long-term stable system operation.

[0022] From a principle perspective, the erbium-doped ring cavity provides physical-layer random mapping, while frequency-domain hash coding acts as an efficient sparsification operator. The two combine to form an optical neural front-end that can be deployed on edge devices, providing low-latency, high-recognition input for the neural dynamics prediction and causal inference modules. Effectively, this solution improves prediction accuracy, reduces power consumption, and maintains robustness against noise and synchronization errors, making it suitable for long-term online operation of wearable health monitoring terminals.

[0023] Preferably, pulse-second signal calibration is performed on the synchronously filtered electrocardiogram signal, photoelectric volume pulse wave signal, continuous blood pressure signal, three-axis acceleration signal, skin temperature signal and ambient oxygen partial pressure signal, and Butterworth filtering and wavelet denoising are performed in sequence after synchronization to obtain a denoised signal for light modulation.

[0024] Before entering the optical modulation link of the present invention, the synchronously filtered ECG signal, photoplethysmographic signal, continuous blood pressure signal, triaxial acceleration signal, skin temperature signal, and ambient oxygen partial pressure signal need to undergo a unified time-domain and frequency-domain cleaning process. First, a pulse-second signal is used for timing calibration. That is, a pulse generated by the hardware once per second is used as a reference to align the time scales of all six analog quantities in the analog-to-digital conversion circuit. Because the pulse-second signal itself remains locked to the timing pulse of the global navigation satellite system, the trigger error of each channel is controlled to the microsecond level after timing calibration. This synchronization accuracy is two orders of magnitude higher than the subsequent optical modulation delay scale, ensuring that phase drift does not occur when multi-channel information is mapped into a chaotic waveform in the erbium-doped ring cavity.

[0025] After the time calibration is completed, each signal undergoes two levels of cleaning: first, a Butterworth filter is used to remove high-frequency noise and baseline drift outside the hardware bandwidth, and then discrete wavelet threshold denoising is used to suppress transient artifacts. The Butterworth filter is chosen instead of a Chebyshev or Bessel filter because the Butterworth filter has a flat response within the passband and can maintain key morphological information such as the ECG peak and the rising edge of the pulse wave to the greatest extent. For ECG signals, the filter cutoff frequency covers the main energy band of cardiac discharge; for photoelectric volume pulse wave signals, the filter cutoff frequency is slightly lower to avoid mixing in high-frequency respiratory resonance components. Wavelet threshold denoising uses a "soft threshold" strategy to set the detail layer whose absolute value of the coefficient is lower than the threshold to zero, and then perform an inverse wavelet transform to restore the time domain signal. Threshold It is obtained by estimating the noise variance, and its core formula is:

[0026] in represents the noise standard deviation, The above formula is not intended to propose a new mathematical principle in this invention, but is used to illustrate the consistency of threshold selection; The estimation method of is consistent with the sampling quantization noise standard of each channel, which ensures that the denoising strength of different channels maintains the same dimension.

[0027] The tandem use of Butterworth filtering and wavelet denoising yields a dual effect: the Butterworth filter provides a smooth noise floor, preventing spurious peaks during subsequent wavelet decomposition; while the wavelet denoising suppresses sudden spikes while preserving the main signal shape, improving the output signal-to-noise ratio by approximately 10dB. This improvement is crucial for the linearity of subsequent optical modulation links. In a Mach-Zehnder modulator, the modulation depth is directly proportional to the input voltage. Excessive high-frequency spikes will cause the optical output to enter the nonlinear region, disrupting the stable chaotic state of the erbium-doped ring cavity. Experimental measurements have shown that when the input signal-to-noise ratio is no less than 30dB, the Lyapunov exponent of the ring cavity's chaotic attractor fluctuates within the expected range, demonstrating the system's long-term stability.

[0028] Example 1: A multi-day wear experiment was conducted, with 5 volunteers collecting data continuously for 24 hours during their daily activities. A 20-minute resting segment was selected as the baseline, and then compared with the exercise segment and the fast staircase segment. The results showed that after the above two-level cleaning, the 95% confidence interval of the pulse second signal timing error distribution fell within 2µs; the ECG signal The standard deviation of peak amplitude decreased by 15%; the coefficient of variation of the rising slope of the photoplethysmography signal decreased by 18%; and the amplitude of the respiratory interference component in the continuous blood pressure signal decreased by 12%. These improvements are reflected in downstream tasks: when the cleaned signal is directly fed into the spiking neural network, the event frequency fluctuation is less than 5%, and the reversible ordinary differential equation network converges to a stable trajectory within the first iteration, significantly reducing the model's cold start time.

[0029] In the overall architecture, two-stage cleaning not only improves signal quality but also reduces system power consumption. The Butterworth filter uses an analog active second-order superposition structure to avoid digital post-processing and reduce the cost of edge hardware. Wavelet denoising utilizes parallel computing on the on-chip digital signal processor, and the required multiplication and addition operations account for less than 3%, which has a negligible impact on battery life in daily wearable scenarios. Through this layered design of "analog pre-cleaning + digital fine denoising", the present invention ensures that subsequent optical modulation and chaotic mapping can operate stably, laying a reliable signal foundation for the overall health data intelligent early warning link.

[0030] Preferably, the frequency domain hash coding performs equally spaced quantization on the real part of the spectrum of the time-delayed feedback chaotic waveform, maps the quantized segments with positive signs to activation bits in a sparse binary vector, and maps the quantized segments with negative signs to zero bits to form a sparse binary vector of fixed length.

[0031] The time-delayed feedback chaotic waveform is a light intensity signal formed in an erbium-doped ring cavity. The ring cavity is essentially an optical delay system with an adjustable coupling coefficient and gain. The input is the driving current after optical modulation of multi-channel physiological data after synchronous filtering and denoising, and the output is a light intensity sequence that evolves over time. This light intensity sequence exhibits quasi-random characteristics, and the high-dimensional delayed feedback makes different time slices have weak correlation, so it can be regarded as a physical layer high-dimensional mapping of multi-channel physiological signals. In order for the subsequent digital processing link to run in real time on the edge hardware, this high-dimensional analog signal needs to be compressed into a fixed length and easy to digitally operate representation. The present invention uses frequency domain hash coding to transform the chaotic waveform into a sparse binary vector, which not only retains the high-dimensional mapping characteristics, but also significantly reduces the computational and storage overhead.

[0032] First, perform discrete Fourier transform on the chaotic waveform obtained by sampling. Assume that the sampling sequence in a single transformation window is ,in Represents the sampling index, and the window length is . Discrete Fourier transform obtains the spectrum sequence ,in In order to avoid complex conjugate operations, the present invention only uses the real part of the spectrum As the input of the subsequent hash coding. The real part amplitude reflects the in-phase component of the sampling sequence at a given frequency, and for the light intensity signal, the in-phase component contains the main energy information. Using the real part can also reduce the hardware bit width of the subsequent comparator. The key to hash coding is to quantize the continuous amplitude into a discrete binary space. The execution steps are as follows: First, set the quantization step size . Calculate the quantization index for each frequency point:

[0033] in Indicates floor rounding operation. Quantization result is an integer, if If is an odd number, then the corresponding position of the sparse binary vector is set to 1; if If is an even number, the corresponding position is set to zero. The length of the sparse binary vector obtained in this way is fixed to half the length of the Fourier window, that is, The variables are defined as follows: The chaotic waveform The voltage value of each sampling point; represents the discrete Fourier transform of The complex amplitude of each frequency point; express The real part of Represents the quantization step size, which is used to control the activation density; Indicates the The quantized index of each frequency point.

[0034] This "odd-even mapping" strategy has the advantages of both hardware friendliness and information compression. Its hardware friendliness is reflected in the simplicity of comparison operations: only one multiplication instruction and one rounding are required to get the index, and then a logical judgment is performed on the lowest bit of the index to determine the setting. Its information compression is reflected in the fact that the amplitude distribution of the random chaotic waveform is approximately symmetrical, and the probability of positive and negative values ​​occurring is equivalent, so the odd-even division can theoretically provide an activation probability of half the frequency points. When the quantization step size is reduced, the amplitude range is divided into more quantization segments, the odd and even segments are more evenly distributed, and the number of activation bits in the sparse vector approaches half the window length; when the quantization step size is increased, only the frequency points with larger amplitudes can fall into the odd segments, and the sparse vector becomes even more sparse. By adjusting during the training phase And evaluate the convergence speed and power consumption of the downstream model, and select a value that compromises between the amount of information and sparsity.

[0035] In the scenario of intelligent early warning of health data, the number of activation bits of a sparse vector can be regarded as the "chaos representation density", which has a significant impact on subsequent system performance. Sparse vectors with higher density have more activation bits and provide richer features, but they will also increase the firing frequency of the pulse neural network, resulting in increased power consumption; too low a density may cause the loss of useful information. Therefore, the present invention adopts a dynamic adjustment strategy: the pulse neural network calculates the activation rate of the current batch. If it is lower than the lower limit, the annular cavity coupling coefficient is fine-tuned to increase the chaos amplitude; if it is higher than the upper limit, the coupling coefficient is reduced to return the activation rate to the set range. This closed-loop control stabilizes the input distribution of the downstream model while ensuring the power consumption budget.

[0036] To illustrate the effect of frequency-domain hash coding, Example 2 is given below: The experimental platform uses a system-on-chip, which includes an analog-to-digital converter, a programmable logic unit, and a pulse neural network computing unit. The Fourier window length is selected to be 1024, and the quantization step size is determined by offline search, so that the average activation bit of the sparse vector accounts for approximately 12%. In 20 hours of continuous sampling, compared with the method of directly using the convolution features of the original waveform, the absolute error of the present invention in predicting the systolic blood pressure interval in the next 20 seconds is reduced by approximately 18%, the peak power consumption of the pulse neural network is reduced by approximately 30%, and the delay is still kept less than 1ms during high-intensity exercise.

[0037] It's worth emphasizing that frequency-domain hash coding has a natural perturbation-sensitive amplification property. For the chaotic waveform output by an erbium-doped ring cavity, small dynamic changes manifest themselves in the spectrum as alternating sign flips in the amplitudes of several frequency points. These flips, after parity mapping, directly cause certain bits of the sparse vector to change from one to zero or from zero to one, generating new spike events in the spiking neural network. Because the spiking neural network weights have adapted to this sparse pattern through sequential synaptic plasticity, they can immediately capture subtle changes in physiological signals and transmit them to the reversible ordinary differential equation network, guiding the subsequent adjustment of the prediction trajectory. This chain ensures the rapid response of the intelligent health data early warning system to sudden cardiovascular stress.

[0038] Regarding the concept of "frequency domain hash coding" that appears for the first time in this invention, it needs to be clarified that "hash" here is not a hash in the cryptographic sense, but a Boolean mapping based on a quantization index. Unlike traditional compressed sensing, this method does not require sparse reconstruction, but only serves as a means of feature compression and activation dynamics control. Compared with the commonly used binarization method, the hash coding of the present invention utilizes symbol information to avoid the high-cost comparison threshold table, and achieves controllable sparsity by adjusting the quantization step size. Since the probability density of the chaotic waveform amplitude of the ring cavity is approximately symmetrical, the two types of symbols are statistically evenly distributed, so the odd-even division will not be biased towards a specific frequency band on a macro scale.

[0039] The sparse binary vector is converted into an event sequence and input into a pulse neural network to generate a current vector, which is used as the control variable to input the reversible ordinary differential equation network integral to obtain the predicted cardiovascular state vector, and the coupling coefficient of the erbium-doped ring cavity is adjusted accordingly. In this invention, sparse binary vectors play a key role as a bridge from the optical neural front end to the neural dynamics prediction link. The sparse binary vector has a fixed length, the activation bit density is controlled by the quantization step size, and has a self-explanatory Boolean structure, which is convenient for direct conversion into an event sequence that can be recognized by neuromorphic hardware. The event sequence generation logic is: when the vector index When the value of is 1, a pulse event is triggered in the current time slice, which is recorded as If the value is 0, no pulse is generated. The zero-delay mapping from sparse binary vectors to event sequences can be completed by the on-chip asynchronous event encoder. The event sequence is fed into a spiking neural network built using a leaky integral-release model. The network is implemented at the circuit level by a capacitor-resistor array, and the neuron membrane potential The evolution of satisfies the differential equation:

[0040] in is the membrane time constant, is the synaptic weight, is the Dirac pulse function, Indicates the The triggering moment of an event. When the membrane potential exceeds the threshold, an output pulse is generated, and then the non-emission period begins. The output pulse is integrated by the current shaping circuit to form a current vector In hardware implementation, the current vector is an analog voltage obtained by accumulating the pulse emission frequency in a time window using a counter, which is then converted into a digital quantity. This preserves the sparsity of event-driven operation while providing a continuous control variable for the subsequent reversible ordinary differential equation network.

[0041] The reversible ordinary differential equation network is a deep learning implementation of the controlled dynamics model, whose state vector It consists of four dimensions: stroke volume, peripheral resistance, heart rate and venous capacity. The core of the network is the vector field The set of learnable parameters The state evolution formula is:

[0042] here Using a reversible coupling structure, each coupling block splits the state dimension in half, uses half as the conditional input, performs an affine transformation on the other half, and shares the same time step numerical integrator within the block. The reversible structure can perform gradient backpropagation without storing all intermediate activations, reducing the memory usage of edge devices. By introducing zero-curl regularization in the offline pre-training stage, the vector field can maintain symmetric gradients and reduce numerical integration drift. In the online inference stage, the current vector As an external driving force to modulate the cardiovascular latent state, the network integrates with millisecond steps to the 20-second prediction moment and outputs the predicted cardiovascular state vector.

[0043] In order to maintain the closed loop consistency between the optical input link and the neural dynamics prediction link, the present invention uses the prediction-observation error to update the erbium-doped ring cavity coupling coefficient. The Euclidean distance between the predicted cardiovascular state vector and the simultaneously measured state vector is recorded as the error scalar. . Using proportional-integral control law:

[0044] In the formula is the coupling coefficient of the In-doped ring cavity, 、 is the gain coefficient, which can be set during the system calibration phase. Increasing the coupling coefficient increases the feedback energy of the ring cavity light field, raising the chaotic amplitude and the activation rate of the sparse binary vector; conversely, it decreases. This adaptive adjustment maintains the activation density within the set range, stabilizing the event load of the spiking neural network and ensuring the consistency of the input distribution of the reversible ordinary differential equation network.

[0045] In Example 3, three subjects were selected to sample in an outdoor walking scene for 40 minutes, and the first 5 minutes were taken as the system warm-up period. The experiment showed that if the coupling coefficient is not adaptively adjusted, the activation bits of the sparse binary vector will drop sharply in the high sweat stage, resulting in a 25% reduction in the firing frequency of the pulse neural network and a lag of about 6 seconds in the prediction of the sudden increase in heart rate by the reversible ordinary differential equation network; after adopting the above closed loop, the activation bit density is always maintained in the range of 10% to 15%, the firing frequency is stable, and the response delay of the reversible ordinary differential equation network to the sudden increase in heart rate is shortened to less than 2 seconds. This result shows that the optical-neural-dynamic three-loop coupling can significantly improve the system's foresight for sudden cardiovascular risks.

[0046] This invention proposes for the first time a complete chain of "sparse binary vectors-event sequences-spiking neural networks-reversible ordinary differential equation networks-coupled optical modulation." Sparse binary vectors release information density in the pulse domain through event sequences, while spiking neural networks leverage temporal plasticity to highly match sparse activation patterns. Reversible ordinary differential equation networks analyze the time-varying response of the cardiovascular system at the dynamic level, while coupling coefficient feedback regulates optical input in real time to maintain system balance. This collaborative mechanism achieves comprehensive optimization of low latency, low power consumption, and high foresight, providing reliable technical support for intelligent early warning of continuous health data.

[0047] Preferably, the spiking neural network is composed of leaky integrate-and-release neurons, which dynamically adjust synaptic weights using time-dependent synaptic plasticity rules driven by event sequences and output current vectors through integration.

[0048] The spiking neural network is the key computing unit for the present invention to realize real-time health data intelligent early warning on edge wearable hardware. The spiking neural network is composed of several leaky integral release neurons, each of which can be implemented in the circuit by a capacitor-resistor parallel structure and a programmable threshold comparator. The membrane potential of a leaky integrate-and-shoo neuron is , the membrane potential evolves with time as follows:

[0049] Where, represents the membrane time constant, represents the synaptic weight, Indicates that it comes from The input event is mapped from a sparse binary vector to an on-chip asynchronous event encoder: if the vector index is 1, then at the current moment A digital pulse with a fixed trigger width; if it is 0, no pulse is generated. This pulse is reflected in the physical hardware as a short voltage step across the capacitor. The multiply-add array is injected into the target leaky integrate-and-spark neuron.

[0050] When the membrane potential Increase to preset threshold When , the leaky integrate-and-fire neuron immediately generates an output pulse signal The output pulses are converted into analog currents representing the average discharge frequency through a programmable amplifier and a low-pass shaping network. The accumulated results of multiple output pulses are integrated into a four-dimensional current vector by a current sink circuit, corresponding to the four physiological dimensions of stroke volume, peripheral resistance, heart rate, and venous capacitance. The current vectors are sampled by an analog-to-digital converter and injected into the reversible ordinary differential equation network as control variables.

[0051] In order to ensure that the pulse neural network can still accurately reflect the activation pattern of the input sparse binary vector under different physiological conditions, the present invention adopts the time-dependent synaptic plasticity rule to update the synaptic weight online. Time-dependent synaptic plasticity is an engineered implementation of the biological synaptic learning rule. Its core is to adjust the weight size according to the order of the input pulse and the output pulse. The present invention implements this rule in the neuromorphic chip in a charge pump manner: when the input pulse appears first and then the corresponding output pulse appears, the synaptic weight Increase; if the order is reversed, To prevent unbounded weight growth or weakening, the chip incorporates upper and lower limit comparators at the weight accumulation end to automatically clip overflow and underflow. This adaptive mechanism continuously tracks the statistical changes in the activation bits of the sparse binary vector over the long term, maintaining the match between the spiking neural network and the input features.

[0052] From the perspective of intelligent early warning of health data, spiking neural networks perform the role of rapid sparse recoding of chaotic mapping signals. While sparse binary vectors compress chaotic waveforms, they still lack timing information. Driven by event sequences, leaky integration and release neurons generate current outputs of varying amplitudes based on the time intervals and strength differences of input pulses. Because the analog-digital hybrid implementation can complete an integration and generate current results in sub-milliseconds, the overall latency of this invention is minimal, meeting the need for proactive early warning in rapidly evolving cardiovascular risk scenarios.

[0053] In Example 4, a spiking neural network array containing 1024 leaky integral-release neurons was used to test 30 minutes of dynamic motion experimental data. Compared with the traditional convolutional neural network baseline, the mean absolute error of the heart rate value predicted 20 seconds later by the spiking neural network-reversible ordinary differential equation network link was reduced by 16%, and the total power consumption of the system was reduced by about 25%. The power consumption advantage comes from two aspects: first, the spiking neural network only triggers calculations when an event arrives, and maintains static power consumption when idle; second, time-dependent synaptic plasticity is implemented in hardware using a charge accumulator, without the need for external storage media to read and write, and the entire process is completed in a single clock beat.

[0054] Adaptive regulation is achieved between the spiking neural network and the erbium-doped ring cavity via a closed-loop "error-coupling coefficient" system. The difference between the predicted cardiovascular state vector calculated by the reversible ordinary differential equation network and the real-time sensor observation vector is linearly mapped to generate an error signal. When the error amplitude exceeds a threshold, the system proportionally adjusts the ring coupling coefficient to increase or decrease the amplitude of the optical chaos waveform, thereby changing the activation density of the sparse binary vector. This change in activation density directly affects the pulse frequency, which in turn affects the input current vector of the reversible ordinary differential equation network, thus completing the closed-loop regulation. This mechanism avoids the problem of characteristic drift caused by sensor drift or changes in ambient optical conditions during long-term operation.

[0055] Preferably, the reversible ordinary differential equation network includes a reversible feedforward coupling structure, and a zero-curl regularization constraint is imposed on the vector field during the integration process to maintain bidirectional mapping consistency of the numerical integration.

[0056] The reversible ordinary differential equation network is the core prediction model used by this invention to simulate time-varying cardiovascular mechanisms. Unlike traditional recurrent neural networks or gated recursive networks, the reversible ordinary differential equation network directly models the evolution of the hidden state as a continuous-time dynamical system and implements bidirectional mapping of state variables through reversible coupling blocks. The basic idea is to transform the prediction problem into solving initial value differential equations and utilize a reversible structure to enable forward reasoning and backward derivation to share the same trajectory, thereby saving edge hardware storage and maintaining numerical gradient stability.

[0057] Assume that the cardiovascular hidden state vector is , including four components: stroke volume, peripheral resistance, heart rate and venous capacity. The current vector output by the pulse neural network is recorded as The vector field function in the network is express, is a learnable parameter. The continuous time evolution formula of the vector field is:

[0058] Where, Represents the derivative of the hidden state with respect to time. In order to ensure the reversibility of the solution, The network is divided into two halves, denoted as and The first reversible coupling block maintains Unchanged, use Calculate the affine transformation parameters acting on The next piece switches roles. Applying the same type of transformation. This structure ensures that the determinant of the overall Jacobian matrix is ​​always unity, making it inherently reversible. During the backpropagation phase, only a small number of affine parameters need to be recalculated to restore the forward intermediate activations, eliminating the need to store the entire process state, significantly reducing the static power consumption of the wearable terminal.

[0059] In order to keep the curl of the vector field zero during the numerical integration process, the present invention introduces a zero-curl regularization constraint during the training phase. , the curl is defined as:

[0060] in Represents the vector field Quantity. The squared sum of forms the regularization term and is weightedly combined with the primary task loss. The zero-curl constraint prevents the vector field from forming circulation structures in high-dimensional space, reducing the probability of explosion or divergence in numerical integration. In cardiovascular dynamics scenarios, there are physical conservation relationships between stroke volume and peripheral resistance, and between heart rate and venous capacity. Zero-curl regularization makes the coupling relationships learned by the vector field closer to real physiological mechanisms, enhancing the interpretability of the model.

[0061] The network's time integration utilizes an adaptive-step fourth-order Runge-Kutta solver. When implemented on edge hardware, it employs a fixed-step approximation and an embedded reversible coupling computation module. Compared to traditional fixed-step Euler integration, the Runge-Kutta method reduces error accumulation under the same step-size conditions while preserving the reversible nature of the vector field. Combined with event-driven input updates, the network triggers an integration after each pulse of the neural network output current vector. The number of iterations ensures a 20-second prediction window and millisecond latency.

[0062] Example 5, focusing on prediction accuracy: Ten subjects were selected for a running test. Both the reversible ordinary differential equation network of the present invention and a traditional gated recurrent network were used to predict their heart rate 20 seconds into the future. The results showed that the mean absolute error of the reversible ordinary differential equation network was 13% lower than that of the gated recurrent network.

[0063] In intelligent early warning applications for health data, reversible ordinary differential equation networks provide two major benefits. First, continuous-time modeling avoids the limitations of the sampling period on prediction resolution, enabling the system to update cardiovascular state estimates at any time. Second, bidirectional mapping ensures that gradients can be traced back to the input event level, enabling subsequent tensor causal networks and diffusion intervention modules to directly utilize gradient information for adaptive adjustments, achieving a fully closed-loop system from signal acquisition to intervention execution.

[0064] The cardiovascular state vectors predicted by the most recent multiple frames are stacked and positionally encoded, then fed into a reversible tensor flow network to generate an encoding vector. The risk vector is obtained through counterfactual perturbations and a classifier, and the parameters of the reversible ordinary differential equation network are updated using the risk vector gradient. In the health data intelligent early warning link of the present invention, the reversible tensor flow network plays a pivotal role in event-level risk inference and gradient feedback. First, the continuous time series information output by the reversible ordinary differential equation network needs to be reorganized into a high-order tensor suitable for tensor flow learning. The specific approach is: in each inference cycle, the cardiovascular state vector of the latest 6 frames is taken from the prediction buffer. Each frame consists of 4 components: stroke volume, peripheral resistance, heart rate, and venous capacity. The 6 frames are stacked vertically in chronological order to obtain a matrix of shape 6×4 In order to make the tensorflow network aware of the time sequence, a position code in the form of sine-cosine is added to the outermost dimension of the matrix. The amplitude of the position code depends on the frame index. , the frequency increases exponentially with the channel index. This operation obtains a third-order tensor with position information .

[0065] The reversible tensor flow network uses a five-level affine coupling stack. Each coupling block splits the tensor along the last dimension into two, with one half remaining identical and the other half updated by a learnable affine transformation. Since the Jacobian matrix of the coupling block is a lower triangular structure, the entire network is easy to calculate in the forward direction and is naturally reversible. The network parameter set is set to , the forward mapping is recorded as , then the encoding vector Defined as:

[0066] Here The dimension of is fixed to 24, which is the product of the number of input frames and the number of channels. The reversible tensor flow network training phase minimizes two losses simultaneously: the log-likelihood term ensures Approximate the standard normal distribution, so as to have the conditions for performing vector operations in the latent space; apply the Jacobi regularization in parallel to centralize the affine scale factor of each level of coupling block to prevent ladder explosion.

[0067] In order to achieve a rapid assessment of potential risks, the present invention performs counterfactual perturbations in the coding space. Cardiovascular event-like potential disturbance vector , defined as the encoding vector The amplitude of the dimension is The increment of , and the other dimensions remain zero. After perturbation, the new potential vector is obtained , through the inverse mapping Get the counterfactual tensor The counterfactual tensor passes through a three-layer fully connected classifier to output the event risk probability. The six types of events correspond to six disturbance vectors, and the final risk vector is a six-dimensional probability array.

[0068] In order to make the risk inference result affect the dynamic model in reverse, the present invention calculates the gradient of the risk vector to the predicted cardiovascular state vector. right The gradient of is multiplied by the inverse of the above Jacobian matrix to obtain right The gradient of the retrieval system is then traced back to the single-frame predicted cardiovascular state vector. This gradient numerically represents the sensitivity of various event risks to small changes in stroke volume, peripheral resistance, heart rate, and venous capacitance. This gradient is weighted and injected into the reversible ordinary differential equation network as an external error signal, causing the dynamic model to automatically bias prediction errors toward high-risk directions in the next training cycle, thereby achieving risk-driven model self-correction.

[0069] The advantages of introducing a reversible tensor flow network are manifested on multiple levels. First, the normalization of the encoding space gives different physiological patterns a unified metric, facilitating cross-population transfer. Second, the counterfactual perturbation operation has a clear and interpretable meaning: the increased risk of comparing dimensions with the same perturbation amplitude directly reflects the causal strength. Third, the reversible structure allows forward reasoning and reverse gradients to share a common path, which not only reduces edge device storage but also avoids the numerical amplification of high-order gradients caused by traditional reverse mode differentiation.

[0070] In Example 6, eight volunteers at high risk of coronary heart disease were selected for the experiment and underwent stress testing in the presence of a doctor. The system recorded 15 premature beats annotated by the doctor. In the baseline version without counterfactual gradient writeback, the reversible ordinary differential equation network triggered 11 warnings 6 seconds in advance; after adding tensor flow counterfactual writeback, 14 warnings were triggered 10 seconds in advance, with only one missed warning and no more than two false alarms, indicating that the risk gradient has a substantial gain on the dynamic model. Power consumption evaluation shows that the tensor flow network only accounts for 12% of the total energy consumption at the edge, mainly because the reversible structure has low computational complexity and gradient sharing. On a wearable smartwatch equipped with a micro graphics processing unit, the delay of tensor flow inference and gradient return is kept within 20 milliseconds, and the user cannot perceive the delay.

[0071] Compared to traditional approaches that directly perform long-short-term memory inference or one-dimensional convolutional feature judgment in the time domain, this invention's inference chain based on reversible tensor flow reduces feature engineering, obtaining explicit probability outputs and interpretable gradients. Furthermore, closed-loop gradient feedback ensures that the dynamic model automatically corrects over time, thereby maintaining sustained prediction accuracy in real-world scenarios. This architecture retains the advantages of end-to-end learning while incorporating the interpretability guaranteed by reversibility and causal perturbations, providing a solution for the field of intelligent health data early warning that balances real-time performance, robustness, and forensic traceability.

[0072] Preferably, a preset number of continuous predicted cardiovascular state vectors are stacked in chronological order, position coding is added, and then input into a reversible tensor flow network to generate a coding vector.

[0073] In order to fully exploit the time correlation in the intelligent early warning link of health data, the present invention continuously reads a preset number of predicted cardiovascular state vectors from the output end of the reversible ordinary differential equation network, performs high-dimensional reorganization and position encoding processing in the tensor domain, and then inputs the reversible tensor flow network to obtain the encoding vector.

[0074] The predicted cardiovascular state vector is denoted as , whose elements are stroke volume, peripheral resistance, heart rate, and venous capacity in order. The present invention selects the most recent frame prediction value. frame And stacked in time order to form a two-dimensional tensor The stacking operation ensures that the historical state continuously describes the current cardiovascular evolution trend, but the two-dimensional tensor itself does not contain explicit time position information. In order to maintain the identifiability of the time sequence in the deep model, the present invention adds sine and cosine position encoding to the outermost layer of the tensor. Let the position index be , the channel index is , the position encoding element is defined as:

[0075] is the encoding dimension, is the frequency base. and Splice by dimension to get a third-order tensor with time and position information After this processing, the frequency domain information and time position information are both explicitly presented in a tensor structure, providing input for the subsequent density mapping of the reversible tensor flow network. The reversible tensor flow network adopts a piecewise affine coupling structure. Each coupling block The last dimension of is divided into two equal length segments, and the scaling coefficient and offset coefficient are calculated on one segment and then applied to the other segment to achieve a learnable affine transformation. The reversibility ensures that the network forward mapping and reverse gradient share the same trajectory, significantly reducing the edge hardware's demand for intermediate activation storage. Let the network parameter set be ,Will Mapping to get the encoding vector In the training phase, the network is optimized by maximum likelihood estimation, so that It is statistically close to the standard normal distribution, which facilitates subsequent counterfactual disturbances.

[0076] On the software side, the continuous predicted cardiovascular state vector is generated by a sliding window method. Window length Through experimental tuning, the main time scales of heart rate variability are covered without prolonging the inference delay. , the experiment verifies that it can take into account both timeliness and accuracy. Position encoding dimension Take 4 to keep consistent with the single frame state dimension. Encoded tensor It is transmitted to the graphics processing unit via a serial channel and used for inference by a reversible tensor flow network.

[0077] On the hardware side, a reversible tensor flow network is deployed on tensor cores that support mixed precision, and coupled block weights are quantized layer by layer to reduce power consumption. The coupled structure of the tensor flow network enables streaming inference, releasing activations from the previous layer after each layer is processed, keeping peak end-side graphics memory usage below 4MB. Position encoding vectors are generated using a lookup table, avoiding trigonometric operations during inference.

[0078] Tensor reorganization and positional encoding enable the model to simultaneously understand short-term changes and relative positions, improving risk discrimination sensitivity. In offline cross-validation, the sensitivity of the version without positional encoding to rapid heart rate spikes was only 0.72; with positional encoding, this increased to 0.83. The density estimation capability of the reversible structure makes the encoded vector an informative and differentiable latent representation, providing a numerically stable workspace for subsequent counterfactual perturbations. Using the same hardware budget, the TensorFlow version consumes approximately 18% less power than a solution based on a long short-term memory network, while maintaining the same prediction lead time.

[0079] In Example 7, volunteers wore a multi-sensor chest belt and bracelet to complete three-stage tasks on a treadmill: constant speed, acceleration, and sprint. The system continued to reason with a window length of 6 and a position encoding dimension of 4. About 5 seconds after the start of the sprint phase, the stroke volume and heart rate rose synchronously. The encoding vector produced a significant offset in the corresponding latent dimension. After amplification by counterfactual perturbation, the classifier output arrhythmia risk probability increased to 0.65, and triggered an early warning. In the same scenario, it was replaced by a network without position encoding, and it took 9 seconds to reach the same threshold. The experimental results show that the solution of the present invention has higher warning timeliness in the second time domain.

[0080] Preferably, the counterfactual perturbation is generated by applying a fixed amplitude offset to a single channel of the encoding vector, while keeping the values ​​of the remaining channels unchanged.

[0081] After completing the forward mapping of the reversible tensor flow network, the system obtains a coding vector "coding vector" of length 24, whose dimensions are approximately independent and satisfy the standard normal distribution after maximum likelihood learning. In order to evaluate the sensitivity of a certain cardiovascular event to the underlying physiological state, the present invention performs counterfactual perturbations in the coding space. The so-called counterfactual means assuming that a certain feature of the coding vector deviates slightly from the current measured value, observing the resulting risk changes, and measuring the causal strength of the feature on the occurrence of the event. In order to avoid the mutual cancellation of disturbances or the introduction of high-dimensional coupling, the present invention adopts a "single-channel fixed amplitude" scheme: only one dimension in the coding vector is modified at a time, the dimension value is increased by a constant offset, and the remaining dimensions remain unchanged. The formal description is:

[0082] in, represents the original encoding vector; represents the counterfactual encoding vector; Indicates the The unit basis vectors of dimensions; Indicates a fixed amplitude offset, which is set uniformly for all entities.

[0083] The selection of fixed amplitude offset follows the principle of “interpretable amplitude”: on the one hand, it must be small enough to ensure that it falls in the high-density area of ​​the encoding distribution during the training period, so that the counterfactual tensor generated by the inverse mapping still conforms to the physiologically feasible range; on the other hand, it must be larger than the statistical fluctuation of the encoding vector standard deviation, so that the risk probability will have a significant difference in value. Offline grid experiments show that when When it is approximately equal to 0.5 times the standard deviation of the encoding vector, the slope of the risk probability curve is the largest and the gradient value is stable. Therefore, the present invention solidifies this coefficient in the model hyperparameters to avoid inference delays caused by online parameter adjustment.

[0084] Compared with the traditional method of injecting multi-dimensional Gaussian noise at one time, single-channel perturbation has three advantages. First, it has strong causal interpretability: the risk change is triggered by only one potential factor, which can be directly mapped to the contribution of this factor to cardiovascular events. Second, the gradient path is unique, which can greatly simplify the chain derivative calculation; in hardware implementation, only one Jacobian column vector needs to be saved to complete the gradient backpropagation. Third, single-channel perturbation allows the direction of risk vector change to be quickly located, and at the policy level, it can be directly mapped to "one-dimensional controllable" intervention measures, such as adjusting the breathing rate or cadence, without the need to analyze high-dimensional interactions.

[0085] The counterfactual encoding vector is inversely mapped through a reversible tensor flow network to obtain a counterfactual tensor, which is then input into a three-layer classifier network to output the risk probability. The system assigns a dedicated dimension to each type of cardiovascular event, applies perturbations sequentially, and calculates the risk probability, ultimately forming a six-dimensional risk vector. Next, the chain rule is used to calculate the gradient of the risk vector with respect to the encoding vector, which is then multiplied by the reversible Jacobian matrix to obtain the gradient of the risk vector with respect to the predicted cardiovascular state vector. This gradient not only provides the directionality of each physiological variable in response to risk, but also provides a numerical signal for the online fine-tuning of the reversible ordinary differential equation network: if a variable has a positive gradient in response to increased risk, the reversible ordinary differential equation network will tend to reduce the prediction bias of that variable in the next round of training, bringing the prediction trajectory back to a safe range.

[0086] Example 8: 10 hypertensive volunteers were selected to collect data on their daily activities, and the system used a 24-dimensional encoding vector to represent short-term cardiovascular dynamics. Compared with the one-time Gaussian perturbation scheme, the single-channel fixed offset scheme increased the average recall rate of six types of events by 9%, reduced the number of false alarms by 2, and shortened the inference delay by 6 milliseconds. The power consumption data was investigated and found that the number of matrix-vector multiplications required in the gradient return phase was reduced by about two-thirds, which is equivalent to reducing the peak power consumption of the digital signal processor by 15%. More importantly, the doctor's review pointed out that the causal explanation produced by the single-channel perturbation is more intuitive. For example, "the increase in the heart rate dimension leads to an increased risk of ventricular premature beats" is easy to correspond to clinical experience, while Gaussian multidimensional perturbation is difficult to directly attribute.

[0087] The risk vector, predicted cardiovascular state vector, encoding vector and contextual information are concatenated to form a state vector. The risk energy is calculated. The risk energy gradient is used in the diffusion generation model to generate an action latent vector and decode it into an intervention action. After the intervention action is executed, the diffusion generation model parameters, the risk weights of each cardiovascular event, the reversible ordinary differential equation network parameters and the erbium-doped ring cavity coupling coefficient are synchronously updated according to the reward function.

[0088] In the decision-making closed loop of this invention, the risk vector, predicted cardiovascular state vector, encoding vector, and contextual information are spliced ​​into a unified state vector to drive the diffusion generative model to generate personalized intervention actions. The risk vector records the immediate probability of six cardiovascular events; the predicted cardiovascular state vector provides physiological trends for the next 20 seconds; the encoding vector retains high-dimensional latent factors; and the contextual information consists of environmental fields such as the current time period, geographic location, and schedule tags. Directly splicing these four types of data in the feature dimension avoids scale imbalance caused by independent normalization and provides a complete state description for subsequent risk energy calculations.

[0089] Risk energy is defined as the scalar of the risk vector linearly weighted by the risk weight:

[0090] in Indicates the Risk weights for cardiovascular events, Indicates the probability of the corresponding event. Risk energy essentially measures the current comprehensive risk intensity and can modulate the noise injection process of the diffusion generation model. In order to make the energy gradient easy to solve, the present invention constrains the weight to be positive and uses piecewise linear activation in the training stage. Since the state vector is differentiable, end-to-end gradient backpropagation can be achieved.

[0091] The diffusion generation model includes two stages: forward diffusion and reverse denoising. Forward diffusion starts from the state vector and gradually adds noise to it to push it into an isotropic Gaussian distribution. Unlike the traditional fixed noise schedule, the present invention adaptively scales the noise amplitude according to the risk energy: the higher the energy, the smaller the noise amplitude, ensuring that the action latent vector generated in high-risk scenarios is concentrated and easy to converge. The reverse stage uses a conditional network to gradually denoise and generate an action latent vector. The conditional network uses the state vector as a condition and outputs the noise residual that needs to be offset at each step. The action latent vector decoder consists of two fully connected layers. The first layer maps the latent vector to a low-dimensional action space, and the second layer uses a piecewise linear function to discretize the continuous value into four types of intervention actions: silent reminder, vibration reminder, breathing guidance, and emergency contact. In addition, several dimensional continuous outputs are mapped to pulse rhythm parameters, which are used to adjust the beat of vibration or breathing prompts.

[0092] After the intervention action is executed, the system collects two types of feedback: physiological feedback from real-time sensor measurements, and behavioral feedback from user confirmation or background diagnosis. The reward function consists of a risk energy term and an action cost term:

[0093] in is the cost coefficient, Constants are set based on the type and frequency of intervention. A larger reward indicates a more effective intervention and a lower burden. The conditional network is updated using a proximal policy optimization method, the core of which is to limit the drift in the action probability distribution between the old and new policies to stabilize online learning. Risk weights are updated using an exponential average method based on the cumulative error weighted by the reward, strengthening the system's focus on recent high-incidence events.

[0094] The parameters of the reversible ordinary differential equation network and the diffusion generation model share a gradient signal: when an intervention successfully reduces the subsequent risk energy, the gradient is propagated back to the reversible ordinary differential equation network along the state vector, causing the dynamic prediction curve to align with the actual downward trajectory. The coupling coefficient of the erbium-doped ring cavity is also adjusted during the same update cycle: if the reward is positive, it indicates that the current coupling coefficient is well matched to the intervention effect, and the system maintains or slightly increases the coupling coefficient. If the reward is negative and stems from high risk that has not been suppressed, the coupling coefficient is reduced, reducing the physical layer activation density and avoiding overload of the spiking neural network.

[0095] Preferably, the diffusion generation model adaptively schedules the noise amplitude according to the risk energy in the forward diffusion stage, and completes the reconstruction of the action latent vector within a predetermined number of steps in the reverse denoising stage.

[0096] The diffusion generation model is a sequence generation framework introduced by this invention at the intervention decision-making level, used to convert health risk assessment results into actionable latent vectors. The model comprises two conjugate paths: forward diffusion and backward denoising. Its overall concept is to gradually map the high-dimensional state vector to a Gaussian prior distribution, then sample within the prior space based on risk energy orientation. Finally, a conditional network is used to restore the noise samples to action latent vectors within a finite number of steps. This mechanism retains the diffusion model's ability to represent complex distributions while also introducing risk-driven adaptive noise scheduling, ensuring a concentrated distribution of action latent vectors and accelerated generation when risk levels are high.

[0097] Forward diffusion with state vector As input, construct a noise schedule ( represents the diffusion step index). The traditional diffusion model uses a fixed noise enhancement scheme. The present invention uses the risk energy Adaptively adjust the noise amplitude. The risk energy is defined as:

[0098] in For the Risk weights for cardiovascular events, is the corresponding event probability. The greater the risk energy, the higher the overall risk, requiring more precise intervention. The present invention writes the noise schedule as:

[0099] 、 are the lower and upper bounds of the noise variance, 、 is an empirically determined energy threshold. This linear mapping reduces the noise injected by forward diffusion in high-risk scenarios, preserving more of the original semantics. In low-risk scenarios, the noise amplitude is amplified, allowing for greater diversity in action latent vectors and minimizing user disruption. The above formula serves only as the core calculation expression; the remaining derivation details are common knowledge for those skilled in the art and will not be elaborated here.

[0100] After the forward diffusion is completed, the terminal noise vector is obtained, and its distribution is close to the standard Gaussian. Inverse diffusion is performed within the step, and the conditional network at each step Take the current noise vector and state condition as input and output the residual estimate The reverse denoising iterative formula is:

[0101] in is the current noise vector. , the reverse denoising residual is also reduced accordingly, which accelerates the convergence speed of the action latent vector. The number of steps is fixed at 30, and the risk energy adaptive noise can be used to maintain the reconstruction accuracy within the given number of steps without dynamically increasing or decreasing the number of steps, thus simplifying the hardware pipeline design.

[0102] The action latent vector is mapped into discrete intervention categories and continuous rhythm parameters through a two-layer decoder. The first layer outputs a 4-dimensional logistic score corresponding to the four intervention categories: silence reminder, vibration reminder, breathing synchronization guidance, and emergency contact. The second layer outputs a 1-dimensional rhythm control variable, which is used to adjust the tempo of the vibration or breathing reminder. The decoder uses piecewise linear activation rather than softmax to avoid gradient saturation. It also allows action category decisions to be driven by small fluctuations in the latent vector, resulting in slightly different actions within the same risk level and reducing user habituation.

[0103] After the intervention action is executed, the system collects physiological feedback and user interaction feedback and calculates the immediate reward:

[0104] in is the action cost constant, is the cost trade-off coefficient. The reward is passed back to the conditional network through the proximal strategy optimization algorithm. , risk weight and the parameters of the reversible ordinary differential equation network. In proximal policy optimization, the loss function truncates the deviation in the action distribution between the new and old policies to ensure stable online learning. For the erbium-doped ring cavity coupling coefficient, the present invention employs reward-gated increase and decrease: if the reward is positive and the risk is reduced, the coupling coefficient is slightly increased in the next cycle; if the reward is negative and the risk energy does not decrease, the coupling coefficient is slightly decreased to suppress redundant activation rates.

[0105] Example 9: During a 24-hour continuous monitoring experiment, the system recorded 18 physician-annotated potential arrhythmia precursors. The diffusion model of the present invention triggered a vibration alert 17 seconds earlier on average, 16 of which did not develop into a definitive arrhythmia. The baseline strategy (fixed noise diffusion) triggered an alert 11 seconds earlier, with a probability of no development of 12. User questionnaire data showed that after implementing risk-energy adaptive noise scheduling, the average hourly alert frequency decreased by 30%, and the subjective intrusiveness score decreased by 25%. In the hardware power consumption evaluation, the diffusion generation model had an inference delay of 26 milliseconds, and the peak power consumption accounted for 12% of the entire device, impacting the smartwatch's battery life by less than 5%.

[0106] Preferably, the reward function consists of a cardiovascular event risk energy term and an intervention action cost term, and the synchronous update iteratively adjusts the diffusion generation model parameters, the risk weights of each cardiovascular event, the reversible ordinary differential equation network parameters and the erbium-doped ring cavity coupling coefficient through a proximal strategy optimization algorithm.

[0107] Reward function design, the risk energy term is based on the linear combination of the aforementioned six-dimensional risk vector and risk weight , For the The probability of cardiovascular events, is the corresponding risk weight; both values ​​range from zero to one. The greater the risk energy, the higher the current comprehensive risk, and the lower the reward should be. The intervention action cost item is recorded as , the cost of silent reminder is one, the cost of vibration reminder is two, the cost of breathing synchronization guidance is three, and the cost of emergency contact is ten. Cost coefficient Used to measure the system's tolerance for user interruptions. , Rewards represent the immediate benefit the system receives at the end of a decision cycle, with higher values ​​being better. Rewards are positive when the system successfully mitigates risk and uses low-cost interventions. Rewards are negative when high risk is not mitigated or high-cost actions are frequently triggered.

[0108] The proximal strategy optimization algorithm is an online update diffusion generation model condition network. The present invention adopts the proximal strategy optimization algorithm. Assume that the logarithmic probability of the old strategy for the action is , the new strategy is The advantage function is estimated to be , update the target: , It is a clipping threshold to prevent the update from being too large and causing the strategy to collapse. Used to update conditional network parameters Since the reverse denoising of the diffusion generative model is completed within thirty steps, the conditional network gradient can be obtained immediately after one inference cycle, meeting the real-time constraints of wearable devices.

[0109] Risk weights are adaptive, and risk weight updates are based on exponential sliding:

[0110] in For the events in the last few rounds The cumulative value of positive rewards obtained, is the negative reward cumulative value, is the learning rate. This formula dynamically tilts the weight toward high-risk events, ensuring that the model's focus automatically drifts with population differences and time.

[0111] The dynamic model is fine-tuned, and the vector field parameter set of the reversible ordinary differential equation network is denoted as The gradient comes from two parts: prediction error and risk gradient. Compound loss:

[0112] The first term is the traditional mean square error, and the second term is the risk gradient weighted term. is the balance coefficient. The reversible structure supports the reuse of forward calculation traces during back propagation, reducing storage. The edge hardware uses a spliced ​​automatic differentiation engine to complete the gradient solution and update .

[0113] Optical coupling self-calibration, the coupling coefficient of the erbium-doped ring cavity is denoted as The present invention adopts the proportional increment method , is the step length. The reward is positive Slightly increase the Chaos Amplitude, when the reward is negative Decrease to prevent the activation bits from being too dense and causing the spiking neural network to saturate.

[0114] like Figure 2 As shown, a health data intelligent early warning system is used to implement the health data intelligent early warning method, including: The optical encoding module optically modulates the synchronously filtered ECG signal, photoplethysmography signal, continuous blood pressure signal, triaxial acceleration signal, skin temperature signal, and ambient oxygen partial pressure signal, and injects the signal into an erbium-doped ring cavity to generate a time-delayed feedback chaotic waveform. This chaotic waveform is then subjected to frequency-domain hash coding to produce a sparse binary vector. The optical encoding module consists of a six-channel analog front-end, a programmable weighted summation chip, a silicon-based Mach-Zehnder modulator, an erbium-doped fiber amplifier, and a 50m ring fiber. Each analog front-end includes an anti-aliasing filter and a 12-bit successive approximation analog-to-digital converter with a configurable sampling frequency. The multi-channel digital stream enters a field-programmable gate array (FPGA), where amplitude normalization and weight synthesis are performed before driving the modulator. The modulated output light intensity forms a chaotic waveform within the ring cavity. The photodetector samples the signal at a rate of 1 GHz and sends it back to the FPGA for fast Fourier transform and comparison operations. A sparse binary vector of length 512 is output using a simple logic table lookup.

[0115] The pulse prediction module converts the sparse binary vector into an event sequence and inputs it into a spiking neural network to generate a current vector. This current vector is then integrated into a reversible ordinary differential equation network as a control variable to obtain a predicted cardiovascular state vector. The coupling coefficient of the erbium-doped ring cavity is adjusted based on the difference between the predicted cardiovascular state vector and the synchronously measured value. In the pulse prediction module, the sparse binary vector is mapped via an asynchronous event encoder to the on-chip neuromorphic chip Loihi2, which integrates thousands of leaky integrate-and-spark neurons and a plastic synaptic array. Neuronal firings converge on an on-chip current integrator to generate a four-dimensional current vector, which is then transmitted to the main control system-level chip via a low-power analog-to-digital converter. The main control system integrates a vector digital signal processor that runs the Euler-Maruyama solver of the reversible ordinary differential equation network. Network weights are stored in on-chip static random access memory and can be computed in parallel using vector instructions. The predicted-observed difference is written back to the ring cavity coupling drive circuit via a serial peripheral bus, completing the physical-digital closed loop.

[0116] The causal evaluation module stacks the predicted cardiovascular state vectors from the most recent frames in chronological order, adds positional encoding, and then feeds the resulting vector into a reversible tensor flow network to generate an encoding vector. Counterfactual perturbations are applied to the encoding vector and passed through a classifier to obtain a risk vector. The gradient corresponding to the risk vector is then used for backpropagation to update the parameters of the reversible ordinary differential equation network. The causal evaluation module caches the prediction vectors for the most recent six frames, assembles the tensor directly in on-chip static random access memory, and appends positional encoding. The encoded tensor is then fed into a small graphics processing unit (GPU) with an integrated tensor core, which runs a five-layer reversible affine coupled network. Network parameters are stored in mixed precision to conserve on-chip static random access memory. Counterfactual perturbations are performed within the tensor core via vector addition, eliminating the need for external memory access. The risk classification head is a three-layer fully connected network, also residing on the GPU. Back gradients are returned to the SoC via a bus and written to the weight registers of the reversible ordinary differential equation network.

[0117] The intervention generation module concatenates the risk vector, the predicted cardiovascular state vector, the encoding vector, and contextual information to form a state vector, calculates the risk energy, and uses the gradient of the risk energy in the diffusion generative model to generate an action latent vector and decode it into an intervention action. After executing the intervention action, the diffusion generative model parameters, the risk weights of each cardiovascular event, the parameters of the reversible ordinary differential equation network, and the coupling coefficient of the erbium-doped ring cavity are synchronously updated according to the reward function. In the intervention generation module, the state vector is concatenated from the system-on-chip and written to a high-efficiency tensor processor to run thirty steps of conditional diffusion inference. Forward diffusion noise is generated by a hardware random number generator and adaptively scaled according to the risk energy. The weights of the backward denoising conditional network are stored in eight-bit fixed-point format, and each step is completed within 20ms. The decoder outputs the intervention category via Bluetooth Low Energy to drive a linear resonant motor or speaker. Emergency communication is transmitted via a standard HTTPS request via the cellular communication module. The reward is calculated within the microcontroller, and the incremental gradient is synchronized to the tensor processor, neuromorphic chip, and ring cavity driver digital-to-analog conversion channel via direct memory access, enabling joint online updates across the four hardware layers: optical, pulse, tensor, and diffusion.

[0118] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A health data intelligent early warning method, characterized in that: include: The synchronously filtered ECG, photoplethysmography, continuous blood pressure, triaxial acceleration, skin temperature and ambient oxygen partial pressure signals are optically modulated and injected into an erbium-doped ring cavity to generate a time-delayed feedback chaotic waveform, and the chaotic waveform is subjected to frequency domain hash coding to obtain a sparse binary vector; The sparse binary vector is converted into an event sequence and input into a pulse neural network to generate a current vector, which is used as the control variable to input the reversible ordinary differential equation network integral to obtain the predicted cardiovascular state vector, and the coupling coefficient of the erbium-doped ring cavity is adjusted accordingly. The cardiovascular state vectors predicted by the most recent multiple frames are stacked and positionally encoded, then fed into a reversible tensor flow network to generate an encoding vector. The risk vector is obtained through counterfactual perturbations and a classifier, and the parameters of the reversible ordinary differential equation network are updated using the risk vector gradient. The risk vector, predicted cardiovascular state vector, encoding vector and contextual information are concatenated to form a state vector. The risk energy is calculated. The risk energy gradient is used in the diffusion generation model to generate an action latent vector and decode it into an intervention action. After the intervention action is executed, the diffusion generation model parameters, the risk weights of each cardiovascular event, the reversible ordinary differential equation network parameters and the erbium-doped ring cavity coupling coefficient are synchronously updated according to the reward function.

2. The method according to claim 1, characterized in that The synchronously filtered ECG signal, photoplethysmographic signal, continuous blood pressure signal, triaxial acceleration signal, skin temperature signal and ambient oxygen partial pressure signal are calibrated by pulse-second signal. After synchronization, Butterworth filtering and wavelet denoising are performed in sequence to obtain a denoised signal for optical modulation.

3. The method according to claim 1, characterized in that Frequency domain hash coding performs equally spaced quantization on the real part of the spectrum of the time-delayed feedback chaotic waveform, maps the positive quantization segments to active bits in a sparse binary vector, and maps the negative quantization segments to zero bits to form a sparse binary vector of fixed length.

4. The method according to claim 1, wherein The spiking neural network is composed of leaky integrate-and-release neurons, which dynamically adjust synaptic weights using time-dependent synaptic plasticity rules driven by event sequences and output current vectors through integration.

5. The method according to claim 1, wherein The reversible ordinary differential equation network contains a reversible feedforward coupling structure and imposes a zero-curl regularization constraint on the vector field during the integration process to maintain the bidirectional mapping consistency of the numerical integration.

6. The method according to claim 1, characterized in that A preset number of continuous predicted cardiovascular state vectors are stacked in chronological order, position encoding is added, and then input into the reversible tensor flow network to generate the encoding vector.

7. The method according to claim 1, characterized in that The counterfactual perturbation is generated by applying a fixed-amplitude offset to a single channel of the encoding vector, while keeping the values ​​of the remaining channels unchanged.

8. The method according to claim 1, characterized in that The diffusion generation model adaptively schedules the noise amplitude according to the risk energy in the forward diffusion stage, and completes the reconstruction of the action latent vector within a predetermined number of steps in the reverse denoising stage.

9. The method according to claim 1, characterized in that The reward function consists of a cardiovascular event risk energy term and an intervention action cost term. The synchronous update iteratively adjusts the diffusion generation model parameters, the risk weights of each cardiovascular event, the reversible ordinary differential equation network parameters and the erbium-doped ring cavity coupling coefficient through a proximal policy optimization algorithm.

10. A health data intelligent early warning system for implementing the health data intelligent early warning method according to any one of claims 1 to 9, characterized in that: include: An optical encoding module is used to optically modulate the synchronously filtered ECG signal, photoplethysmography signal, continuous blood pressure signal, triaxial acceleration signal, skin temperature signal, and ambient oxygen partial pressure signal, and inject the optically modulated signal into an erbium-doped ring cavity to generate a time-delayed feedback chaotic waveform, and then perform frequency-domain hash coding on the chaotic waveform to obtain a sparse binary vector. a pulse prediction module, configured to convert the sparse binary vector into an event sequence and input the event sequence into a spiking neural network to generate a current vector, input the current vector as a control variable into a reversible ordinary differential equation network integral to obtain a predicted cardiovascular state vector, and adjust the erbium-doped ring cavity coupling coefficient according to the difference between the predicted cardiovascular state vector and a synchronously measured value; a causal assessment module, configured to stack the predicted cardiovascular state vectors of the most recent multiple frames in chronological order, add position encoding, and then input the stacked vectors into a reversible tensor flow network to generate an encoding vector; apply counterfactual perturbations to the encoding vectors and pass them through a classifier to obtain a risk vector; and use the gradient corresponding to the risk vector as backpropagation to update the parameters of the reversible ordinary differential equation network; An intervention generation module is configured to concatenate the risk vector, the predicted cardiovascular state vector, the encoding vector, and the context information to form a state vector, calculate the risk energy, use the gradient of the risk energy in a diffusion generation model to generate an action latent vector and decode it into an intervention action, and after executing the intervention action, synchronously update the diffusion generation model parameters, the risk weights of each cardiovascular event, the reversible ordinary differential equation network parameters, and the erbium-doped ring cavity coupling coefficient according to a reward function.

Images