Multi-source health risk prediction method

Abstract

The invention relates to the technical field of digital health monitoring, in particular to a multi-source health risk prediction method, which comprises the following steps of: synchronously acquiring inertia, vital signs, near-infrared and visual events, and sending the events into a photon pulse network to generate an event tensor after quaternary wavelet and pulse density modulation; a topology signature is obtained through persistent coherence and optimal transmission, Hamiltonian power students are embedded into the topology signature to form risk energy, and the risk energy is stored on a chain and fed back to the network. The risk, topology and timestamps are input into a regular strategy optimization network, touch pressure, stride frequency, respiration and illumination control quantity are output to drive an actuator to perform closed-loop intervention, an award drives a model and dissipation coefficient to perform adaptive updating, and end-side high-precision prediction and personalized real-time intervention are achieved.

Description

Technical Field

[0001] This invention relates to the field of digital health monitoring technology, and in particular to a method for predicting multi-source health risks. Background Technology

[0002] Predicting health risks can provide early warnings before symptoms appear, helping individuals and medical institutions to intervene in advance, reducing disease progression and medical costs. Existing wearable systems mostly rely on a single signal (such as heart rate or steps) and use traditional statistical models to estimate risk, resulting in insufficient sensitivity and specificity. While medical-grade fixed monitoring is highly accurate, it lacks applicability to mobile scenarios and cannot cover high-risk periods in daily life. Therefore, the industry urgently needs a new solution that balances multimodal accuracy and real-time performance. Summary of the Invention

[0003] To address the numerous problems existing in the prior art, this invention provides a multi-source health risk prediction method. This invention maps multi-source synchronous pulses to event tensors, extracts persistent topological features, injects conserved dissipative dynamics to generate risk energy, and then uses a reinforcement learning network to output a four-dimensional intervention signal and uses reward feedback callback dynamics to achieve adaptive, real-time, and interpretable health risk management.

[0004] A method for predicting multi-source health risks includes the following steps: The synchronous inertial data is decomposed into four-element wavelet decomposition, and the synchronous vital signs data and near-infrared spectral data are pulse density modulated. The pulse density modulated is then input into the photonic pulse neural network along with the visual event pulses to generate an event semantic tensor. Based on the event semantic tensor, a persistent coherent barcode is calculated, and a topological signature is obtained through optimal transmission centroid alignment. The topological signature is embedded into the Hamiltonian constant differential equation to generate a risk energy sequence. The topological signature and the risk energy sequence are written into the edge blockchain, and the gradient of the risk energy sequence with respect to the topological signature is fed back to the photonic pulse neural network. A state vector is formed using risk energy sequence, topological signature, and timestamp. This vector is then input into a proximal policy optimization network with an optimal transmission regularization term to obtain pressure control, step frequency control, breathing control, and illumination control. These control variables are used to drive the actuator to implement intervention using pulse width modulation signals. Rewards are generated based on user compliance. The proximal policy optimization network is updated with the rewards, and the updated information is uploaded. The dissipation coefficient of the Hamiltonian constant differential equation is adjusted based on the rewards.

[0005] Preferably, the four-element wavelet decomposition includes: combining the three-axis acceleration signal and the three-axis angular velocity signal into a four-element vector, performing wavelet transform on the four-element vector on a multi-scale convolution kernel, and writing the resulting convolution coefficients into a tensor for processing by a photonic pulse neural network.

[0006] Preferably, the pulse density modulation includes: generating a one-bit pulse sequence from vital sign data and near-infrared spectral data using a cascaded oversampling modulator, wherein the interval between adjacent pulses maintains a monotonic mapping relationship with the instantaneous amplitude of the input signal.

[0007] Preferably, the photonic pulse neural network is composed of a Mach-Zehnder interferometer array, and the phase shifting element adjusts the refractive index through the thermo-optic effect. The refractive index is updated synchronously according to the pulse time difference dependent plasticity rule and the contrast reconstruction error.

[0008] Preferably, the persistent coherent barcode is obtained by constructing an Alpha complex within a fixed-length sliding window and extracting a one-dimensional barcode and a two-dimensional barcode. The barcode is then subjected to an entropy-regularized optimal transmission algorithm to calculate the centroid and generate a topological signature.

[0009] Preferably, the entropy-regularized optimal transmission algorithm performs a fixed number of iterations in each centroid calculation process to obtain a continuously differentiable alignment result as the topological signature.

[0010] Preferably, the gradient of the risk energy sequence to the topological signature is calculated by an automatic differentiation engine, and the control voltage of the phase shift element in the photonic pulse neural network is adjusted at a set update frequency.

[0011] Preferably, the pressure control quantity drives the pneumatic soft actuator to adjust the air chamber volume through a pulse width modulation signal, the step frequency control quantity drives the step frequency prompting device to output a beat signal, the breathing control quantity drives the breathing prompting device to output a breathing rhythm signal, and the illumination control quantity drives the illumination adjustment device to change the luminous intensity.

[0012] Preferably, the reward signal consists of the difference between the preceding and following values ​​of the risk energy sequence, the user compliance flag, and the optimal transmission distance between the topology signature and the individual steady-state topology signature.

[0013] Preferably, the uploaded update information includes a policy gradient and topological statistics processed with differential privacy noise, wherein the topological statistics are the mean and covariance of the topological signature.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By fusing multi-source pulse events using a photonic pulse neural network, low-latency extraction of cross-modal time-difference dependent features was achieved. Stable representation and differentiable mapping of multi-scale topologies were realized through persistent coherence and optimal transmission centroid alignment. A physically consistent closed loop between risk assessment and intervention actions was achieved by coupling reinforcement learning intervention with Hamiltonian frequent differential equations. Federated training was implemented through differential privacy gradients and topological statistics uploads, balancing model generalization with user privacy protection. Attached Figure Description

[0015] Figure 1This is a schematic diagram of the process of the present invention; Figure 2 The framework diagram of multi-source data fusion and closed-loop intervention in this invention. Detailed Implementation

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

[0017] like Figures 1-2 As shown, a multi-source health risk prediction method includes the following steps: The synchronous inertial data is decomposed into four-element wavelet decomposition, and the synchronous vital signs data and near-infrared spectral data are pulse density modulated. The pulse density modulated is then input into the photonic pulse neural network along with the visual event pulses to generate an event semantic tensor. In the first step of this invention, the multimodal raw signal is mapped to a unified event semantic tensor, laying the foundation for subsequent topological analysis and dynamic derivation. This step includes: (1) quaternary wavelet decomposition of the inertial signal; (2) pulse density modulation of vital signs signal and near-infrared spectral signal; and (3) temporal fusion of multiple pulses in a photonic pulse neural network.

[0018] Four-element wavelet decomposition, with triaxial acceleration and triaxial angular velocity sampled synchronously under a unified clock, forms a 6-channel raw sequence. The system encodes each sample point as a four-element vector:

[0019] in , , These are the linear accelerations in the X, Y, and Z directions, respectively. This represents the angular velocity amplitude. The Morlet wavelet mother function is used to... At multiple scales Perform convolution on top:

[0020] It is a conjugate wavelet. The quaternary wavelet coefficients retain complete information in both amplitude and phase, and can be directly fed into the pulse learning circuit without the need for post-conversion phase reconstruction.

[0021] Pulse density modulation is used, with the ECG sequence, photoplethysmography (PPG) curve, and near-infrared spectral sequence being continuous amplitude signals. A cascaded oversampling modulator outputs a one-bit pulse stream. Each input sample is integrated and compared to a zero threshold; the quantizer generates pulses based on the difference and adjusts the feedback. The pulse interval maintains a monotonic mapping to the original amplitude, facilitating direct accumulation by the photonic circuit.

[0022] Multi-stream pulse fusion is achieved, with inertial wavelet pulses, vital sign pulses, spectral pulses, and visual event pulses entering the photonic pulse neural network in parallel after being clocked by the same clock. This network consists of 256 Mach-Zehnder interferometer arms, and the reconstructible matrix multiplication is performed on-chip by propagating the optical signal. Phase-shifting elements utilize the thermo-optical effect to change the refractive index, acting as weights. The update rule conforms to pulse time-difference dependent plasticity: positive time differences lead to increased weights, and negative time differences lead to decreased weights. The global threshold is synchronously adjusted based on the contrast reconstruction error, forming a self-supervised loop. After 100 iterations, the pulse layer weights converge.

[0023] This invention utilizes quaternary wavelet decomposition to perform global convolution on inertial data at the vector level, eliminating the need for channel-specific registration and ensuring consistent time-frequency unfolding. Pulse density modulation directly outputs the event format, avoiding traditional analog-to-digital conversion, multi-threaded buffering, and timestamp alignment. The photonic pulse network completes weighted accumulation in the sub-millisecond range, with lower energy consumption than the electro-optical-electro-electrical three-stage scheme. The sparsity of the output event semantic tensor is increased by approximately 3 times, eliminating the need to construct a high-density adjacency matrix for subsequent persistent cohomology calculations, significantly reducing computation time.

[0024] In an example, in a wristband prototype, the inertial measurement unit and the photoplethysmography (PPG) curve are both sampled at a frequency of 500 Hz, and the near-infrared spectrum is sampled sequentially at eight wavelengths. After quaternary wavelet decomposition, the inertial coefficient tensor dimension is 128; after pulse density modulation, the vital signs pulse rate is approximately 20 kHz, and the spectral pulse rate is approximately 2 kHz. These pulses are fed into a 256-arm photonic network. After 100 rounds of self-supervised training, the event semantic tensor refresh noise is less than 5%, the tensor sparsity is improved by 3 times compared to the undecomposed baseline, and the subsequent risk extrapolation delay is reduced by approximately 20%.

[0025] Through the above scheme, the system can quickly compress multi-source signals into a unified tensor while maintaining time synchronization and cross-modal consistency, providing high-quality input for subsequent topology and dynamics modules.

[0026] Preferably, the four-element wavelet decomposition includes: combining the three-axis acceleration signal and the three-axis angular velocity signal into a four-element vector, performing wavelet transform on the four-element vector on a multi-scale convolution kernel, and writing the resulting convolution coefficients into a tensor for processing by a photonic pulse neural network.

[0027] Triaxial acceleration and triaxial angular velocity signals jointly record the linear and rotational motion of the human body in three-dimensional space. However, they differ significantly in dimensions, phase, and frequency bands, and direct splicing would lead to uneven feature distribution. This invention first strictly aligns the two sets of signals on the sampling time axis, and then constructs a quaternion vector sequence to uniformly represent instantaneous attitude and displacement trends. A quaternion consists of three imaginary basis vectors and one real scalar, and can describe rotational relationships using the closure and noncommutativity of multiplication; therefore, it is suitable as a composite carrier of inertial data. Let... , , These are the linear accelerations in the X, Y, and Z directions, respectively. , , To determine the angular velocity of the corresponding axis, first calculate the magnitude of the angular velocity. Then on the virtual base , , The following is written:

[0028] This represents a four-element vector flow, where each component corresponds one-to-one with time. This step preserves the directional information of linear acceleration and embeds the magnitude of angular velocity into the real scalar term, allowing subsequent convolutions to be performed in a single channel.

[0029] Next, wavelet transform is performed on the multi-scale convolution kernel. The wavelet mother function is of Morlet type, with its real part being a Gaussian-modulated sine and its imaginary part being a Gaussian-modulated cosine, exhibiting good time-frequency localization capabilities. (Scale set) Arranged in ascending order of frequency, corresponding to progressively expanded frequency bands. Convolution operation syntax:

[0030] in For the first Individual scale, central moment The conjugate wavelet function. Convolution coefficients. Still in quaternion form, the amplitude and phase will exhibit significant distribution differences depending on the scale. To establish a unified feature dimension, this invention concatenates the coefficients of all scales along the scale axis and normalizes the amplitude of the imaginary and real parts within each scale. The normalization constant is calculated from the local maximum within the sliding window, without relying on global statistics, ensuring that the transformation is implemented online.

[0031] When writing the transformation results into a tensor, it is necessary to keep them synchronized with other modes. Specifically, this is done at each time step... Define tensor fragments , dimension ,in For the scale number, Corresponding to the four-element vector components. (The rest of the text appears to be a list of keywords or tags and doesn The four components are filled in sequentially. .all Concatenate them along the time axis to form a three-dimensional tensor. , This is the number of time steps. It is then cached using a double-ended queue. The recent The frames, aligned with the timestamps carried by the pulse density modulated vital sign pulse stream, spectral pulse stream, and visual event pulse stream, are written into the input register of the photonic spiking neural network. A tiling strategy is used here: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Flattened into a one-dimensional vector and encoded as a pulse beam of equal width, the photonic network can receive inertial features and other modal pulses in the same time slot.

[0032] The photonic pulse neural network utilizes a Mach-Zehnder interferometer array to perform matrix multiplication. Tensor flow is distributed to the interferometer arms after passing through the input coupler, and the phase shifting elements of the interferometer arms are initialized with refractive indices according to the weight matrix. Each input pulse generates a superposition of light intensity at the interferometer node. The light intensity is converted into a peak voltage by the detector, and the voltage then controls the phase shifting elements to perform instantaneous weight adjustments. Weight updates follow pulse time-dependent plasticity: if the input pulse arrives at the detector earlier than the output pulse, the phase shift is increased; otherwise, it is decreased. To avoid weight drift, this invention introduces contrast reconstruction error as a global regularization. The error is calculated by the digital backend and periodically written to the thermo-optical controller. After several training cycles, the photonic network exhibits differentiated weights in response to inertial features at different scales. Low-scale convolution coefficients are mainly assigned higher gains to capture gradual attitude changes; the weights of high-scale coefficients are moderately suppressed to reduce the influence of high-frequency noise.

[0033] Compared to traditional methods, this invention offers advantages in the following aspects: First, the quaternion representation avoids the phase asynchrony problem of axis-specific combinations, maintaining the consistency of rotational information during convolution. Conventional three-axis independent wavelets require additional alignment steps, resulting in higher computational cost and errors. Second, multi-scale coefficients are uniformly incorporated into the photonic network and processed in parallel by an optical interference array, achieving throughput at the speed of light, far faster than circuit matrix multiplication. Third, time-difference-dependent plasticity is embedded in the hardware layer, allowing online training without interrupting the data stream and continuously generating adaptive updates to environmental changes. Fourth, once the convolution coefficients are written into the tensor, they share the clock domain with vital sign pulses and spectral pulses, eliminating the need for software time calibration and mitigating jitter errors at the hardware level.

[0034] In this example, on a laboratory verification platform, the inertial sampling frequency was set to 500 Hz, the number of scales was 8, and the sliding window length was set to 32 frames. The number of photonic network interferometer arms was 256, corresponding to the length of the fully unfolded tensor. After 50 training cycles, the response spectrum of the quaternary convolution coefficients showed that the passband weights of the low-frequency scales were about twice as high as the weights of the high-frequency scales. The system was subjected to three scenarios—jumping, walking, and vehicle driving—and 10 minutes of data were collected. The event semantic tensor obtained by this invention was compared with the direct pulse on-photonic network output without wavelet decomposition. The former improved the subsequent topological signature confidence index by 17% and reduced the computational latency by 22%.

[0035] Through the above design, the decomposition results not only provide high-resolution attitude features, but also complete the fusion with subsequent multimodal streams at the hardware level, laying the foundation for the overall real-time performance and accuracy of the multi-source health risk prediction link.

[0036] Preferably, the pulse density modulation includes: generating a one-bit pulse sequence from vital sign data and near-infrared spectral data using a cascaded oversampling modulator, wherein the interval between adjacent pulses maintains a monotonic mapping relationship with the instantaneous amplitude of the input signal.

[0037] In the multi-source health risk prediction chain, this invention converts vital sign data and near-infrared spectral data into pulse format, which is a crucial step in achieving "full-pulse" processing. The pulse density modulation method used is derived from the oversampling analog-to-digital conversion principle, but customized improvements have been made in implementation details and application objectives.

[0038] Vital signs data (ECG time-domain sequences and photoplethysmography pulse curves) and near-infrared spectral data (multi-wavelength transdermal reflectance sequences) are essentially continuous amplitude signals. Direct sampling and digitization require multiple parallel analog-to-digital converters, which not only increases power consumption but also introduces complex buffering and timing logic when aligning with visual event pulses and inertial wavelet pulses. This invention employs a cascaded oversampling modulator to map amplitudes to a one-bit pulse sequence. The "one bit" in "one-bit pulse sequence" refers to the fact that the pulse signal output at each sampling moment carries only a single binary information bit, with only two states: "pulse present" and "no pulse," equivalent to 1-bit quantization in traditional analog-to-digital conversion. In the cascaded oversampling modulator of this invention, continuous amplitude signals are converted into this one-bit pulse stream: amplitude information is expressed through a monotonic mapping of the interval between adjacent pulses, rather than using multiple numerical values ​​to represent the amplitude; the time stamp is marked by a unified hardware clock, ensuring precise alignment of different modal pulses within the event domain. Therefore, "one bit" emphasizes minimizing the quantization bit width, so that vital signs and near-infrared spectral signals can directly enter the photon pulse computing link in the form of time density, which is superior to multi-bit amplitude coding in terms of power consumption, bandwidth and optical domain accumulation efficiency.

[0039] The shorter the pulse interval per unit time, the higher the corresponding input amplitude; the longer the interval, the lower the corresponding amplitude. The pulse occurrence time point naturally carries amplitude information, while retaining a high-resolution time stamp, facilitating seamless integration with the event triggering mechanism of photon pulse networks.

[0040] The modulator structure, a cascaded oversampling modulator, consists of an integrator, a quantizer, and a feedback path. Let the input signal be... The integrator output is The quantizer output is a single-bit pulse. The feedback signal, multiplied by the integrator gain, is returned to the previous stage. The integrator integrates the input signal at a high sampling rate, and the quantizer compares the integration result with the zero threshold, outputs a pulse, and resets the integration path. The feedback suppresses over-integration, creating a noise shaping effect that pushes the quantization noise power into the high-frequency region. When the pulse stream subsequently accumulates in the optical domain, the high-frequency noise can be naturally attenuated by the low-pass response.

[0041] In this multi-stage cascaded scheme, the dynamic range of the signal amplitude differs significantly from the original amplitudes of vital signs and near-infrared spectra. To avoid saturation of the quantizer in a single-stage modulator, this invention employs a two-stage cascaded architecture. The first stage has an oversampling ratio of 32 to refine the time resolution; the second stage has a shaping coefficient of 2 to enhance the high-frequency noise reduction effect. The quantization error from the preceding stage is fed back and reshaped by the integrator in the following stage, ensuring that the time difference accuracy of the output pulse meets the pulse collision avoidance requirements of the downstream photonic network.

[0042] The synchronization mechanism uses a single clock source (40MHz) shared by all modulators, inertial sensors, and event cameras. Within the modulator, each quantization result is encoded as either a "pulse" or a "gap." Each pulse carries an absolute timestamp tag before being routed to the distribution bus; the timestamp is directly generated by the clock counter, avoiding the need for later software patch alignment. Simultaneously, the system calculates the average pulse interval within each statistical cycle to estimate baseline drift in real time and perform automatic gain control; this adapts to amplitude shifts caused by factors such as wearing tightness and changes in skin light path.

[0043] The spectral signal is preprocessed; the raw near-infrared spectral signal is a multi-wavelength sequence. This invention employs time multiplexing, with eight groups of LEDs lighting up in a fixed sequence, each group's excitation pulse lasting 1 ms. The photodetector output is amplified by an analog amplifier before entering the modulator. Since different wavelengths correspond to different tissue depths and blood oxygen absorption characteristics, the multi-wavelength sequence is interleaved into the modulator, resulting in a periodic pulse time distribution. The photonic network can distinguish wavelengths and assign independent weights based on the pulse arrangement order, eliminating the need for additional channel identifiers.

[0044] The hardware implementation integrates the modulator and pulse timing logic into the mixing-on-chip system of the wristband motherboard. Each vital sign channel is bound to a secondary modulator, sharing a clock and power supply. The modulator core uses a capacitor-integrating operational amplifier and a comparator, with the output pulse driving a CMOS-level driver. After single-ended to differential conversion, the pulse enters the optical bus coupler. Power consumption budget: The static power consumption of a single-channel modulator is approximately 45µW. The transient power consumption of the pulse drive varies with the input amplitude, but averages below 60µW, meeting the requirements for 24-hour wearable operation.

[0045] This invention folds amplitude information, which originally required resolutions of 12 bits or more, into the time domain using a single-pulse format, significantly reducing the data width of the digital interface and minimizing high-speed parallel port connections. Through oversampling and noise shaping, high-frequency quantization noise is suppressed during the multiplication and accumulation process of the photonic network, eliminating the need for additional digital filtering. Marking pulses within the same clock domain avoids cross-domain alignment errors between respiration, heart rate, and inertial events, providing an accurate time base for subsequent persistent coherence and conservation dynamics calculations. Multi-stage cascading causes quantization noise to attenuate with the square of the frequency, improving the fidelity of high-resolution low-frequency content and ensuring the integrity of the low-frequency waveform of the electrocardiogram and the rising edge information of the pulse wave.

[0046] In an example, during performance verification experiments, the original ECG amplitude range was ±3mV, and after two-stage modulation, the average pulse flow density was approximately 18kHz. The pulse flow was reconstructed into an analog signal, and compared with the sampling results of a 14-bit successive approximation analog-to-digital converter, the signal-to-noise ratio difference was less than 1.2dB. When the same ECG pulse flow was input into a photonic network, after 50 iterations of time-dependent learning, the correlation coefficient between the generated event semantic tensor and the inertial convolution tensor increased by 0.15, indicating enhanced cross-modal fusion performance. In near-infrared spectroscopy experiments, when detecting arterial blood oxygen fluctuations, the pulse flow of this invention could be recognized by the photonic network within 120ms after the change occurred, which is higher than the 170ms response time of traditional analog-to-digital conversion and repackaging methods.

[0047] In summary, pulse density modulation (PDM) converts analog signals into time-domain event streams in a low-power hardware manner, maintaining the integrity of vital signs and spectral information while being deeply coupled with the computational model of photonic pulse networks, providing a high-time-accuracy and high-consistency input foundation for multi-source health risk prediction.

[0048] Preferably, the photonic pulse neural network is composed of a Mach-Zehnder interferometer array, and the phase shifting element adjusts the refractive index through the thermo-optic effect. The refractive index is updated synchronously according to the pulse time difference dependent plasticity rule and the contrast reconstruction error.

[0049] The photonic pulse neural network is the core computing unit for multimodal event semantic fusion in this invention. Its hardware framework is a two-dimensional Mach-Zehnder interferometer array. Each interferometer arm consists of an input coupler, a beam splitter waveguide, two phase-shifting elements, and an output coupler connected in series. Adjacent interferometer arms in the array are interconnected, enabling real-coefficient matrix multiplication of arbitrary dimensions. The input end simultaneously receives inertial convolution pulses, vital sign pulses, near-infrared spectral pulses, and visual event pulses. The output end integrates light intensity within a fixed refresh period, and the light intensity is converted into voltage by a photodetector and written into the event semantic tensor.

[0050] Phase-shifting elements utilize the thermo-optical effect to adjust the refractive index. When the control current... As heat flows through the aluminum heat-conducting channel to the top of the silicon waveguide, the temperature of the heated area increases, and the waveguide's refractive index changes with the temperature coefficient. Linear change, thus producing phase delay The relationship between delay and current can be written as:

[0051] in The scaling factor is determined by the waveguide geometry, thermal resistance, and material dispersion. The phase delay determines the output light intensity weight at the interference node and is the physical mapping of the trainable matrix in this invention. The implementation details of the weight update mechanism are given below.

[0052] Pulse time difference depends on plasticity rule, assuming the arrival time of the input pulse is... The output pulse arrival time is Define the time difference

[0053] According to the bio-inspired impulse learning principle, when A positive value indicates that the input pulse has a causal precedence effect on the triggering of the output pulse, and the phase delay needs to be increased to improve the weight; when... A negative value indicates that the input pulse lags behind the output pulse, and the phase delay needs to be reduced. The update quantity is written as:

[0054] in The learning rate is a constant. It is a time constant. This is a symbolic function. This expression contains only... This variable ensures that the hardware can directly generate the gain sign while determining the pulse sequence.

[0055] Compared to the reconstruction error rules, the photonic network sets up a self-feedback path at the output end, and the reconstruction error is calculated by the digital back-end:

[0056] in The target pulse density vector, This is the output of the network's self-encoding mechanism. The error, used as a global normalizer, is converted into voltage via a digital-to-analog converter. And broadcast to all phase-shifting elements. The mapping between voltage and phase changes is:

[0057] in This is the global shrinkage coefficient. The reconstruction error is first processed by the digital backend. Multiply by the proportionality factor Generate analog voltage increment Then, the thermo-optical phase shift element multiplies the voltage increment by the temperature-phase coefficient. Convert to phase change The two-level coefficients are combined and denoted as the global contraction coefficient. Therefore, we obtain the calculation expression. A negative sign indicates that the larger the error, the more the phase should be adjusted towards suppression; a negative sign also indicates that the larger the error, the smaller the phase delay, thus shrinking the overall weights and preventing the photon network from overfitting to the current sample. The final weight increment is:

[0058] Ensure coordinated adjustment of local temporal correlation and global reconstruction quality. Indicates phase, Indicates heating current. Indicates the pulse time difference. This indicates the reconstruction error.

[0059] The implementation process includes: the input pulse is coupled to the interference array via a silicon nitride waveguide; and the timing logic is recorded each time an output pulse is triggered. and And calculate immediately Hardware lookup table provides... Approximate value and determine Then the on-chip digital controller reads the reconstruction error from the previous refresh cycle. Incremental generation from both parts Current driver according to Adjust the heating voltage of the phase shifting element to complete the weight writing. Since the thermo-optical response time is on the order of microseconds, which is much smaller than the refresh cycle, it can be considered as an instant update.

[0060] Compared to electronically weighted matrices, photonic arrays eliminate the register writing bottleneck. Input event pulses are transmitted quasi-synchronously using single photons, and intensity interference occurs only within the waveguide of each interferometer arm, eliminating the need for cross-chip bus operations. Secondary self-supervised training (usually) (Less than 200), the phase matrix is ​​stable. At this time, the photonic network can map inertial wavelet coefficient pulses, vital sign pulses, and near-infrared spectral pulses to a unified semantic space, realizing cross-modal coupling.

[0061] Through this invention, a silicon-based Mach-Zehnder array is fabricated on a single sheet of... The chip implements 256×256 real coefficient matrix multiplication within a parallel optical path. Throughput is determined by the input pulse rate and is not limited by register write frequency. Thermo-optical refractive index tuning eliminates charge injection effects, resulting in a write lifetime superior to phase change material solutions, capable of withstanding over [durations not specified]. Sub-weight updates meet the long-term adaptive requirements of wearable scenarios. Time difference-dependent plasticity is calculated directly at the hardware layer without the need for central processing unit intervention; contrast reconstruction errors are generated in batches and broadcast by the digital backend, reducing single-weight global synchronization communication. Through adaptive weight learning, the photonic network can form a high-response channel for specific health risk features (such as the simultaneous occurrence of abnormal heart rate variability and posture imbalance), providing high-confidence input for subsequent persistent topology coherence calculations.

[0062] In this example, a 256×256 photon array prototype was built, using the pulse density modulation output mentioned above as the input pulse source, with a refresh period of 250μs. After 150 training cycles, the network's reconstruction error on the test set decreased by 73%, and the inference power consumption was 18mW. The trained network output was fed into a topology persistent coherence module. When tested with real-world wearing data, it was found that the detection lead for high-risk posture-based rapid heart rate combinations was 30ms higher than the baseline multimodal all-electronic network, meeting the real-time warning requirements for outdoor strenuous exercise scenarios.

[0063] Through the above design, the photonic pulsation neural network achieves fast weight adjustment, low inference power consumption, and cross-modal time consistency at the hardware level, providing a key computational foundation for multi-source health risk prediction.

[0064] Based on the event semantic tensor, a persistent coherent barcode is calculated, and a topological signature is obtained through optimal transmission centroid alignment. The topological signature is embedded into the Hamiltonian constant differential equation to generate a risk energy sequence. The topological signature and the risk energy sequence are written into the edge blockchain, and the gradient of the risk energy sequence with respect to the topological signature is fed back to the photonic pulse neural network. In this invention, the event semantic tensor is a three-dimensional structure: the time dimension records the refresh step size, the scale dimension records the distribution positions of inertial wavelets and vital sign pulses, and the channel dimension records the output amplitude of the photon pulse neural network after weight mapping. The system processes this tensor in a sliding window manner, forming four consecutive sub-steps: topological information extraction, topological vector alignment, dynamic deduction, and reliable write-back.

[0065] Persistent homology barcode extraction, assuming a window length of [value missing]. A frame flattens a tensor sheet within a window into a point cloud. Using Euclidean distance as a metric, within a threshold... Built-in Alpha complex; with As the complexity increases, connected components merge and two-dimensional holes appear. Using the Ripser topological algorithm, one-dimensional and two-dimensional homology are calculated to obtain the barcode set:

[0066] in Indicates the first The birth threshold of a topological feature. This represents the corresponding mortality threshold. The difference between the two values... The settling time, which reflects topological characteristics, can serve as a weight in subsequent alignment.

[0067] Optimal transmission centroid alignment, treating the barcode as a discrete measure:

[0068] Weight and Proportional, center of mass The metric is calculated using the entropy-regularized optimal transmission algorithm. The center of gravity makes the second-order Wasserstein distance The minimum is reached. After the Sinkhorn iteration stops, a continuously differentiable mass distribution is obtained, and the length is written in a fixed order. vector This is called a topological signature. The aligned signature has a fixed dimension, facilitating direct comparison between different windows and different individuals.

[0069] Hamiltonian frequent differential equation derivation, defining hidden state vectors (cardiac output, pulmonary ventilation, etc.) and conjugate momentum vector Hamiltonian:

[0070] Where the mass matrix It is a diagonal positive definite matrix; potential energy Composed of two fully connected networks Approximate and accept topological signatures As additional input, the combination of conservation and dissipation terms yields the differential equation:

[0071] Let be the dissipation matrix. The Sinhalese method is used at time step 1. Integrate to obtain the state trajectory. Define the risk energy:

[0072] in For the Sigmoid function, This represents the individual's steady-state vector. The risk energy forms a sequence over time and is used as input to the decision network.

[0073] Blockchain writing and gradient feedback will enable topological signatures. With risk energy sequence The data is packaged, hashed, and written to the edge blockchain to generate a timestamp, ensuring the traceability of subsequent model training. The writing process follows a lightweight Byzantine fault-tolerant protocol, with a single commit latency of less than 1 second. The gradient is then obtained through automatic differentiation.

[0074] The current of the phase shift element in the corresponding channel is adjusted according to a set ratio via shared memory to achieve cross-layer adaptive operation.

[0075] This invention utilizes topological barcodes to measure connectivity and holes within an event tensor in a homology sense, enabling the capture of cross-modal synchronization anomalies. Centroid alignment ensures the comparability of the topological signature across different windows and individuals, avoiding dimensional mismatches caused by differences in point cloud size. The potential energy function explicitly depends on the topological signature, allowing the dynamic model to obtain cross-modal structural information and resulting in a smoother risk energy curve. Blockchain records are augmented with immutable timestamps, and the gradient feedback chain is transparent, meeting compliance and auditing requirements in medical scenarios.

[0076] In an experiment involving 30 volunteers wearing the device continuously for 14 days, the window length was set to 6 seconds and the topological signature length to 64. The algorithm of this invention was compared with the baseline of the default topological embedding. For two types of events—running sprint and standing dizziness—the median rising lag time of the risk energy curve was 0.18 seconds in this invention, compared to 0.31 seconds at the baseline. The reinforcement learning intervention strategy converged in 28 iterations under the framework of this invention, compared to 40 iterations at the baseline, demonstrating the gradient stability brought by topological injection. The daily blockchain data writing volume is 90 kilobytes, which can be met by Bluetooth Low Energy transmission. Comprehensive verification shows that the process of this invention can calculate risks in real time on the device side, generate traceable records, and enhance the front-end feature extraction capability through gradient feedback closed loop.

[0077] Preferably, the persistent coherent barcode is obtained by constructing an Alpha complex within a fixed-length sliding window and extracting a one-dimensional barcode and a two-dimensional barcode. The barcode is then subjected to an entropy-regularized optimal transmission algorithm to calculate the centroid and generate a topological signature.

[0078] Event semantic tensors form high-dimensional point clouds within a sliding window. Directly comparing Euclidean distances often results in being overwhelmed by the different modal amplitude scales. This invention introduces topological data analysis, using locally connected structures rather than absolute coordinates as the criterion to avoid information loss caused by amplitude normalization. The implementation process consists of two levels: one-dimensional and two-dimensional topological feature extraction, and entropy regularization for optimal transmission centroid alignment.

[0079] Within a fixed-length sliding window, continuous Frame event semantic tensor flattened into point cloud For each radius threshold Constructing an Alpha complex: when the distance between two points is less than Connect edges as needed; when three points lie within a common sphere, add a surface; iterate in this way until a higher-order simplex is formed. Change This results in connected component merging and 2D toroidal hole closure; topological events can be represented using barcodes. The Ripser algorithm directly computes persistent cohomology on the distance matrix of the point cloud, outputting a set of one-dimensional barcodes. With two-dimensional barcode set .in Represents the time when topological features are generated , Indicating extinction Survival span It is a core indicator for measuring stability. The longer the survival span, the more the shape characteristics are maintained within a wider threshold range.

[0080] The number of barcodes varies with the point cloud density, leading to dimensional inconsistencies between different windows. To obtain a fixed-length vector, this invention treats the barcodes as weighted measures. Let the weights... center of mass Constructing a metric:

[0081] Indicated by Dirac measure with fulcrum, Simultaneously traverse both one-dimensional and two-dimensional barcodes. The goal is to... Aligned to reference measures in shared space To reduce random drift between windows, the alignment criterion uses second-order Wasserstein distance:

[0082] in Given a jointly distributed set, direct solution requires quadratic programming, which is time-consuming. This invention introduces an entropy regularization term. This is transformed into a Sinkhorn-based distributed iterable form. Let the iteration variable be a two-dimensional matrix. , Record the cost matrix . This represents the square of the second-order Wasserstein distance, used to measure two probability measures. and Differences in a geometric sense. This represents a discrete measure generated from persistent homology barcodes, where each barcode feature is represented by its centroid position and weight. It represents the optimal transmission centroid measure, which represents the common "average" shape of multiple windows or multiple individual topological distributions. Indicates all For marginal distribution, with It is the set of joint distributions of marginal distributions, also known as the set of feasible transportation plans. express A specific joint distribution, i.e., a transportation plan, is integrated over it. This means finding the solution with the lowest handling "cost" among all feasible transportation plans. Indicates taking from the measure respectively and The support point represents the centroid position of the two barcode features on the real number axis. express and The squared Euclidean distance is taken as the cost of transporting a single feature. Indicates in joint distribution The overall transportation cost is obtained by summing the expected values ​​of the handling costs.

[0083] The iterative steps only involve vector multiplication, scaling, and exponentiation, making it suitable for parallel implementation. The centroid measure is obtained after the iteration stops. Its discrete mass is sorted and written into dimensional vector Because entropy regularization maintains the absolute continuity of the measure, right It is continuously differentiable, laying the foundation for subsequent gradient backpropagation.

[0084] Obtain topological signature Subsequently, this invention injects it into the potential energy function of the Hamiltonian ordinary differential equation. In the network implementation, the potential branch consists of two layers of linear mapping plus rectified activation, with the input being a concatenated hidden state vector and... The weights are updated via edge training. The system calculates the risk energy for the hidden state trajectory:

[0085] In this formula For individual homeostasis, For the Sigmoid function, The proportionality constant is obtained through automatic differentiation. The gradient vector is obtained directly on a single SoC. The gradient is written back to the photonic pulse neural network, and the front-end weights for risk-related topology patterns are strengthened by adjusting the current bias of the phase shifting element.

[0086] To achieve traceability and tamper-proof characteristics, this invention packages the topology signature and synchronization window risk energy sequence into a structure, calculates a secure hash function to output a 256-bit digest, and then writes it to an edge blockchain based on a lightweight Byzantine fault-tolerant protocol. The submission process does not involve raw personal privacy data, only derived features and energy values, meeting privacy compliance requirements. When the decision-making network needs historical risk journeys in subsequent processes, the signed data can be read back from the chain to ensure that the decision-making basis has not been tampered with.

[0087] This invention utilizes persistent coherent barcodes to capture the ring-like co-occurrence structure of cross-modal events along the time axis, enabling the detection of the synchronicity between inertial displacement anomalies and dramatic changes in vital signs. Entropy regularization and optimal transmission centroid alignment map the barcode to a fixed dimension, avoiding the dimensional drift introduced by traditional splicing or sorting methods. Topological signatures, as input, make the potential function dependent on the event geometry, providing physical interpretability for the evolution of hidden states. Blockchain writing adds an immutable timestamp to the gradient feedback, providing an external, reliable log for algorithm parameter tuning.

[0088] Example: In an experiment where 20 subjects wore wristbands and data was collected continuously for 7 days, with a window length of 6 seconds, an average of 140 topological features were extracted from the barcodes. After alignment into a 64-dimensional topological signature, the risk energy curve was compared with the abnormal time periods marked by doctors, and the early detection rate reached 92%. Without entropy regularization alignment, the early detection rate decreased to 78%. After gradient feedback activation, the weight of the photonic network front end for dangerous mode pulses increased by 1.7 times within 100 training iterations, and the overall inference latency of the system was controlled within 150 milliseconds, meeting the requirements for real-time health intervention.

[0089] Preferably, the entropy-regularized optimal transmission algorithm performs a fixed number of iterations in each centroid calculation process to obtain a continuously differentiable alignment result as the topological signature.

[0090] In the multi-source health risk prediction chain, topological signatures are responsible for converting event semantic tensors. The task of compressing high-dimensional geometric structures into fixed-length vectors allows for direct use by subsequent dynamic models. The event semantic tensor, derived from a photonic pulse neural network, preserves the synchronization relationships between inertia, vital signs, near-infrared spectroscopy, and visual pulses, while maintaining cross-domain uniformity in amplitude. However, the tensor slides continuously over time, and the number and distribution of point clouds within the window change rapidly with the scene. Directly comparing Euclidean coordinates is easily overwhelmed by amplitude scale differences; using a fixed-grid histogram results in dimensional explosion and a lack of differentiable structure. This invention combines topological data analysis with optimal transmission theory, proposing an entropy-regularized optimal transmission centroid alignment algorithm. This algorithm maps variable barcodes to consistent-length topological signatures while maintaining gradient continuity, balancing computational efficiency and physical interpretability.

[0091] Obtaining persistent coherent barcodes, in length of In the sliding window of the frame, the continuous The sheet volume is flattened into point clouds ,in This represents the number of channels multiplied by the number of scales. Using Euclidean distance as a metric, the radius threshold is gradually increased. In each The following describes the construction of an Alpha complex; the birth and disappearance of connected components and holes in the point cloud at different thresholds; and the Ripser algorithm outputs one-dimensional and two-dimensional barcodes.

[0092] For the first A topological feature birth threshold, The difference between the two is the death threshold. Known as the survival span, it is a commonly used metric for measuring the saliency of features. One-dimensional barcodes typically reflect clusters of synchronous events, while two-dimensional barcodes correspond to circular interactions of different modalities.

[0093] Treating barcodes as a measure involves directly concatenating them, but the number of barcodes varies depending on the window content, which is inconvenient for network input. This invention transforms barcodes into discrete measures. ,in For the number of barcodes, weight Represents relative survival span; center of mass The typical threshold representing the occurrence of topological features. This yields a measure. It also includes stability weights and threshold positions, which can reflect the topological shape concentration within the window.

[0094] Entropy regularization and optimal transmission centroid alignment: To obtain a fixed-length vector, this invention employs optimal transmission centroid alignment: finding the measure Make and The second-order Wasserstein distance is minimized.

[0095] Classical optimal transmission requires solving a quadratic programming problem, with a computational complexity of O(n log n). It is difficult to run continuously on the edge. This is addressed by adding an entropy regularization term to the objective function. The distance metric becomes:

[0096] in It is a joint transport plan. This is the cost matrix. This represents the entropy regularity strength coefficient, used to control noise levels; the larger the value, the smoother the transportation plan. The Joint Transport Plan indicates that in the 19th century... The source centroid and the first The mass share allocated among the centroids of each target. Denotes the Shannon entropy of the joint distribution; and The multiplication is used as a regularization term to penalize overly sharp transport plans. This represents the squared second-order Wasserstein distance after incorporating entropy regularization, balancing geometric cost and distribution smoothness. Represents the source measure and the target measure, corresponding to the barcode measure and the centroid measure of the current window, respectively. This means that the marginal distributions are respectively equal to and The set of all joint distributions is the set of feasible transportation plans. This represents the specific transportation plan selected from the feasible set. This means finding the transportation plan with the lowest total cost among all transportation plans. It represents the pure transportation cost, summarizing the transport mass of each pair of support points between the two measures multiplied by the transport cost. Represents the elements of the cost matrix, the first... The source centroid and the first The square of the Euclidean distance between the centroids of the target objects. The centroid position of the barcode feature is represented by a real number, which indicates its coordinates on the distance axis.

[0097] After introducing the entropy term, the Sinkhorn iteration can be performed... An approximate solution is obtained within a time limit, and the transport plan is continuously differentiable, making it suitable for automatic differentiation at the end.

[0098] The iterative process includes: 1. Initializing two vectors 2. Calculate the kernel matrix. 3. Repeat the following updates. Second-rate: ,in and 4. Calculate the transportation plan, specifying the target row and column weights. .

[0099] Number of iterations Using a fixed value, stable results are usually obtained after 20 to 30 trials. Output metric. The discrete masses are arranged in ascending order of their centroids, and the first few are truncated. Each weight filling vector Because Sinkhorn is mapped in The time is continuously differentiable with respect to the input, therefore the topological signature is... Compared to barcodes The gradient can be calculated directly without explicit sorting and backpropagation.

[0100] Coupled with the dynamic model, topological signature Entering the potential energy network :

[0101] This, in turn, affects the risk energy output by the Hamiltonian constant differential equation. The system obtains this energy through the automatic differential chain rule:

[0102] in Calculated by the adjoint method of ODE, Determined by the network weights, the gradient ultimately adjusts the weights of the photon pulse network, achieving end-to-end adaptation from the front end to the middle end and back end.

[0103] Through this invention, entropy regularization ensures strong convexity in the transport plan, avoiding non-unique solutions arising from discrete measure alignment. The continuous differentiability property allows topological signatures to be used for gradient backpropagation, solving the problem of non-differentiability of traditional topological features. The computational cost of Sinkhorn iteration decreases with... Linear growth is robust to window size, making it suitable for wearable devices. Fixed-length signature vectors provide stable input to the dynamics model, reducing dimensional fluctuations in the potential energy network.

[0104] In an example experiment involving 20 subjects, the window length was 6 seconds, and the average number of barcodes was 140. Entropy regularization temperature. With a value set to 0.1 and 25 iterations, the edge-side Tensor acceleration core took 9 milliseconds. Aligning the risk energy with the abnormal segments labeled by doctors, the early detection rate reached 92%. Compared to the baseline without centroid alignment, the early detection rate improved by 14 percentage points.

[0105] In summary, the entropy regularization optimal transmission centroid alignment algorithm obtains a differentiable topological signature by a fixed number of iterations, providing a stable and highly discriminative structured input for the multi-source health risk prediction link of this invention, and achieving overall adaptation through gradient loops.

[0106] Preferably, the gradient of the risk energy sequence to the topological signature is calculated by an automatic differentiation engine, and the control voltage of the phase shift element in the photonic pulse neural network is adjusted at a set update frequency.

[0107] In the multi-source health risk prediction chain, the risk energy sequence serves as both a dynamic quantitative indicator for assessing a user's current health status and a feedback signal driving the adaptive updates of the front-end feature extractor. This invention uses an automatic differentiation engine to convert the partial derivative of the risk energy sequence with respect to the topological signature into a physically executable control voltage. This voltage directly adjusts the phase-shifting elements in the photonic pulse neural network at a fixed update frequency, achieving an end-to-end closed loop.

[0108] The risk energy output by the Hamiltonian constant differential equation is defined as:

[0109] For a moment The hidden state vector. For individual steady-state vectors, This is the proportionality coefficient. The function is the Sigmoid function. The hidden state evolution process depends on the topological signature vector. Specifically, it is written into the potential energy network:

[0110] in It is a two-layer fully connected network. The gradient of the risk energy with respect to the topological signature can be expanded according to the chain rule:

[0111] First item Derived from the explicit formula, the second term The adjoint method is used to solve the ordinary differential equation on its trajectory. The automatic differentiation engine can obtain both terms simultaneously in a single backpropagation.

[0112] The calculation process includes: 1. Time step The state trajectory is obtained by integrating the Hamiltonian differential equation using the Sinusoidal method. .

[0113] 2. At the end time Calling the adjoint equation, from Integrating in reverse At the same time, accumulate The Jacobian vector product.

[0114] 3. with Multiplying them together yields the gradient vector. , For topological signature dimensions.

[0115] 4. Each completed Number of windows (system experience value) Update once, and change the current Mapped to phase increment , Step size, This corresponds to the 256 interferometer arms of the photon pulse network.

[0116] 5. Put Converted to control voltage increment The relationship is determined by the calibration coefficient. Give The final voltage update command is sent via on-chip... The bus sends the signal to the thermo-optical phase shifting element.

[0117] Hardware mapping is used; the phase-shifting element employs an integrated structure of a silicon waveguide and an aluminum heating arm, with the refractive index increasing linearly with temperature. Temperature is adjusted by controlling the current, and temperature changes are mapped to phase delay. The current driver has a 12-bit resolution and can generate a minimum step size of 30µA, mapped to a phase resolution of approximately 0.002π. To avoid overshoot, this invention sets the maximum single-step phase change to 0.05π; if... If the limit is exceeded, the entire gradient vector is scaled proportionally. The update frequency is coupled with the window refresh cycle, for example, triggering once every 20 frames to ensure that the gradient is smooth in time and matches the arrival speed of the event pulse.

[0118] This invention enables automatic differential linking to directly reflect the sensitivity of risk energy to the input topology, with a clear target for weight adjustment, eliminating the need for manual learning rate searching. Gradient calculation is performed on the edge using the adjoint equation, independent of cloud computing power and preventing the leakage of private data. Thermo-optical phase tuning is completed in microseconds, offering lower latency compared to electronic variable resistance matrices, ensuring real-time closed-loop performance. A fixed update frequency prevents weight oscillations caused by high-frequency noise while maintaining the ability to track slow physiological trends.

[0119] In an example, during a continuous wear experiment with 30 subjects, the system had a window length of 6 seconds, updating the phase every 20 windows. After gradient feedback was activated, the photonic pulse network increased the weights of pulse channels that synchronized with dangerous postures and abnormal heart rates by an average of 1.7 times, while decreasing the weights of low-correlation channels by 0.4 times. Compared to the version without feedback, the risk energy rise advance was increased to 180 milliseconds, while maintaining a power consumption of 22mW during everyday activities. This result verifies the feasibility and effectiveness of gradient-driven photonic weight adjustment at the edge.

[0120] This invention uses an automatic differential engine to calculate the gradient of risk energy with respect to the topological signature in real time, and maps it to the phase of the control voltage-adjusted photonic pulse neural network at a fixed update frequency. This achieves closed-loop adaptation from high-level risk quantification to low-level feature extraction weights, thereby improving the sensitivity and stability of the multi-source health risk prediction link in complex scenarios.

[0121] A state vector is formed using risk energy sequence, topological signature, and timestamp. This vector is then input into a proximal policy optimization network with an optimal transmission regularization term to obtain pressure control, step frequency control, breathing control, and illumination control. These control variables are used to drive the actuator to implement intervention using pulse width modulation signals. Rewards are generated based on user compliance. The proximal policy optimization network is updated with the rewards, and the updated information is uploaded. The dissipation coefficient of the Hamiltonian constant differential equation is adjusted based on the rewards.

[0122] In the closed-loop decision-intervention phase of this invention, the system needs to convert high-level evaluation results into multimodal intervention commands in real time and back-inject the intervention effectiveness into the dynamic model to maintain consistency between prediction and regulation. To this end, this invention constructs a reinforcement learning network based on near-end policy optimization, constrains the policy distribution by incorporating an optimal transmission regularization term, and directly drives the actuator with the four types of control variables output by the network in pulse width modulation format. The entire process is completed on the edge, without relying on cloud inference, thus shortening the feedback loop and reducing the risk of privacy leakage.

[0123] State vector construction, risk energy sequence From Hamiltonian constant differential equations in recent Output within each refresh step; Topological signature sequence Updated at the same time interval; timestamp sequence Returned from the edge blockchain. The system concatenates the three sets of sequences along the time axis into a matrix and then flattens it to obtain a matrix of length [length missing]. state vector To suppress dimensional differences, layer normalization is first performed on each subsequence, and then the whole sequence is multiplied by a scaling factor to make different modes have comparable magnitudes in the vector space.

[0124] The near-end policy optimization network consists of two fully connected mapping layers, with a hidden layer width of [missing information]. The activation function uses a gated exponential linear unit. The output branch gives the mean vector. With diagonal logarithmic variance Jointly determine the control distribution:

[0125] in It is a four-dimensional control vector. Indicates the state Lower generation control quantity The strategy distribution; Represents a multidimensional Gaussian distribution; the mean is... The covariance is a diagonal matrix. Classical near-end strategy optimization directly uses the pruning ratio to constrain the difference between the old and new strategies, while this invention introduces an optimal transmission regularization term into the loss:

[0126] The first-order Wasserstein distance. is the regularization coefficient. Used to measure the output distribution of the old and new strategies and The differences between them, by constraining policy changes at the feature distribution level, can avoid excessive actions on "rare high-risk patterns" in a single update. The optimal transmission distance is approximated using a sample dual form, with computational complexity linearly related to batch size, making it suitable for edge execution.

[0127] The mapping from control quantities to actuators is as follows: the control vectors correspond sequentially to touch pressure, step frequency, breathing, and light intensity. The system maps each component... Limit to Then linearly mapped to duty cycle The pulse width modulation carrier frequency is uniformly set to 200 to ensure that the response time of the software actuator and the audio actuator is consistent.

[0128] The pressure control drives the solenoid valve of the pneumatic wristband; the higher the duty cycle, the longer the air chamber inflates, and the greater the wristband's contraction amplitude. The cadence control drives the beat vibrator, providing evenly spaced vibrations to prompt the user to adjust their cadence. The breathing control drives the bone conduction speaker to play breathing rhythm notes; the beat interval is inversely proportional to the duty cycle. The illumination control drives the LED array on the inner side of the wristband; the duty cycle determines the brightness, and color temperature modulation indicates relaxation or a warning.

[0129] Reward calculation and network updates: After the intervention signal is output, the system waits for a fixed delay. Recalculate risk energy With topological signature The rewards are calculated using a weighted sum:

[0130] in As a sign of user compliance, To measure topology deviation for optimal transmission distance, For individual steady-state signature, The coefficient is set by hyperparameters. When the user does not follow the prompts or the garment is worn unstable... The reward was significantly reduced. Network parameter updates used a near-end strategy to optimize loss weighting. The gradient is optimized using an adaptive moment estimation algorithm, and the learning rate employs a walk decay. Each data collection... Each trajectory triggers an update on the edge, simultaneously extracting policy gradients and topological statistics, which are then uploaded to the cloud for aggregation after differential privacy noise perturbation. This edge-cloud federated approach both continues individual learning and ensures overall diversity.

[0131] Dissipation coefficient write-back, reward signal is converted into adjustment quantity by exponential averaging. Directly modify the diagonal elements of the dissipation matrix ,in The matrix is ​​the identity matrix. Positive rewards reduce dissipation, encouraging the dynamical system to retain its existing energy trajectory; negative rewards increase dissipation, accelerating the return of the hidden state to steady state. Since the dissipation term is located in the coefficient matrix on the right-hand side of the ordinary differential equation, this modification immediately affects the state prediction in the next window.

[0132] In this example, the wristband device refresh cycle is 250 microseconds, the window length is 6 seconds, the policy network batch size is 32, and the update cycle is 20 batches. The measured time for a single policy forward pass and loss backward pass on the edge is 3.5 milliseconds, and the additional overhead of PPO and optimal transmission regularization accounts for less than 25%. In a 10km running scenario, the system's intervention latency for the "heart rate spike - gait instability" joint mode is an average of 320 milliseconds, and an external accelerometer verifies that the user's cadence returns to the target range within 12 seconds. After adding reward-dissipation write-back, the peak risk energy is reduced by 18% compared to the group without write-back, indicating a closer synergy between the motivation model and the intervention strategy.

[0133] The power consumption data shows that the average power consumption during the policy network inference and update phases is 8.5 milliwatts, and the average power consumption during actuator activation is 24 milliwatts, which meets the energy consumption budget for continuous 24-hour operation of the wearable device.

[0134] This invention achieves integrated decision-making and intervention by inputting a state vector composed of risk energy sequence, topological signature, and timestamp into a near-end policy optimization network with optimal transmission regularization. It utilizes reward write-back to adjust the dissipation coefficient, transferring the regulatory effect to the prediction model. Combined with differential privacy federated updates, it balances personalization and group diversity. Ultimately, it forms a closed loop from event extraction, risk inference, physical intervention to model adaptation, providing reliable technical support for real-time health risk management.

[0135] Preferably, the pressure control quantity drives the pneumatic soft actuator to adjust the air chamber volume through a pulse width modulation signal, the step frequency control quantity drives the step frequency prompting device to output a beat signal, the breathing control quantity drives the breathing prompting device to output a breathing rhythm signal, and the illumination control quantity drives the illumination adjustment device to change the luminous intensity.

[0136] The intervention layer of this invention outputs a control vector based on multi-source risk prediction. At its core, the system maps these parameters in real time into pressure control, cadence control, respiration control, and illumination control. These are then used to drive four types of actuators via pulse width modulation (PWM) to achieve closed-loop health intervention. The hardware implementation, signal conversion principles, and application effects of each control parameter are described below.

[0137] The pressure-controlled quantity is controlled by a pneumatic soft actuator, which consists of a 0.3mm thick silicone air bladder, a polyester fiber restraint layer, and a miniature solenoid valve, with a rated inflation pressure of 60kPa. The controller will... Linear mapping to duty cycle:

[0138] The pulse carrier frequency is fixed at 200Hz. In a single cycle... Inside, high level continues Low level for a continuous period of time During the high-level period, compressed air is injected into the air chamber via the solenoid valve. The volume of the air chamber is:

[0139] in , The increased volume causes the wristband to exert a positive normal force on the skin, which can be quantified as pressure.

[0140] The aeroelastic coefficient, This refers to the contact area. Adjustment is possible. Tactile stimulation can be smoothly controlled within the range of 2kPa-15kPa. If the blood perfusion index remains above 1.3 times the baseline after exercise, the system settings... This generates a duty cycle of 0.8, using high-intensity touch pressure to guide the user to remain still and relax; after the indicator recovers... Gradually reduce to -0.2, and the pressure intensity drops.

[0141] The cadence indicator uses a 10mm diameter vibration motor array, each motor with a rated frequency of 180Hz. The controller is set to... Calculate the vibration trigger interval, where The target step rate is 150 steps / min. If the user's actual step rate is detected to be more than 10 steps / min lower than the target, the policy network outputs... The vibration interval is compressed to 0.48s to guide faster strides; output is given when the stride frequency exceeds the target by 15 steps / min. The vibration interval was widened to 1.2 seconds to reduce the load. A six-minute run test showed that the step loss rate decreased from 8% to 3%.

[0142] The breathing indicator is a bone conduction speaker with a frequency response of 100Hz-8kHz. The controller will... Convert to beat cycle 2.5 seconds corresponds to 24 breaths per minute. During the interval sprint phase, when the heart rate rises to 190 bpm, the system outputs... Shorten the beat cycle to 1.25s and guide rapid breathing; output during the recovery phase. The beat cycle lengthens to 3.25 seconds, promoting deep breathing. Bone conduction cues are clearly perceptible at a sound pressure level of 70 dB without masking ambient sound.

[0143] The lighting adjustment device consists of 32 silicon-controlled blue-green LEDs, each with a maximum brightness of 1000 lux. The controller is based on... Set the duty cycle to drive the constant current source. Brightness and Linear correspondence: It outputs approximately 900 lux of bright cool white light for afternoon fatigue relief. A 200-lux warm light output simulates dusk, promoting sleep. In a 14-day office environment experiment, the subjective drowsiness score decreased by 1.8 points (out of 7) after the light intervention. The closed-loop algorithm and synergistic benefits are achieved by generating the four-dimensional control vector from the proximal policy optimization network, with the loss function... ,in For the first-order optimal transmission distance, The policy network updates at the edge every 20 collected trajectories; the policy gradient is perturbed with Gaussian noise (variance 0.05) before being uploaded to the cloud. Reward function. , For risk energy, For compliance, This is for topological signature deviation. Rewards are fed back into the execution process after exponential smoothing. The system adjusts the dissipation coefficient in real time to maintain consistency between dynamic prediction and intervention. A 14-day wearing experiment showed that after adding four-channel intervention and reward write-back, the peak risk energy was reduced by 20% compared to the baseline without intervention, and the wearer's subjective fatigue score decreased by 1.8 points, verifying the system's real-time performance and effectiveness.

[0144] Preferably, the reward signal consists of the difference between the preceding and following values ​​of the risk energy sequence, the user compliance flag, and the optimal transmission distance between the topology signature and the individual steady-state topology signature.

[0145] The reward signal is the core evaluation component of the reinforcement learning intervention module, used to measure the immediate contribution of a control action to the user's overall health. The system decomposes the reward into three quantitative indicators: risk energy change value, user compliance flag, and topological offset distance, and then synthesizes them into a single scalar with linear weights, ensuring both gradient continuity and differentiability, and mapping to clear physiological and behavioral meanings.

[0146] Risk energy is derived from Hamiltonian regular differential equations, and its numerical range is normalized to 0 to 1 using a sigmoid function. The system records the start and end risk energy every 6-second window, calculates the difference, and then performs exponential smoothing to reduce measurement noise. A negative difference indicates a decrease in risk, while a positive difference indicates an increase in risk. This item directly evaluates the physiological effect of the intervention.

[0147] User compliance indicators are calculated in real time through a motion detection loop. The detection module compares the target signals with the actual execution signals across four channels: pressure, cadence, breathing, and light exposure, and obtains the maximum relative error. A maximum error of no more than 10% is considered fully compliant, between 10% and 25% is considered partially compliant, and exceeding 25% is considered non-compliant. This indicator encourages users to perform actions that are easy for them to execute, and also allows for timely reduction of intervention intensity when compliance decreases, avoiding over-stimulation.

[0148] Topological offset distance measures the degree of deviation of the current multimodal cooperative pattern from the individual steady state. The system first maps persistent coherent barcodes to fixed-length topological signatures using an entropy-regularized optimal transmission algorithm, and then calculates the weighted Manhattan distance between the current signature and the benchmark signature. The weights are adaptively adjusted using an exponential moving average to make high-risk relevant dimensions more sensitive. After the distance is normalized to 0 to 1, it enters the reward function, generating positive incentives when the structural deviation decreases.

[0149] The core reward formula is written as , For risk energy, For compliance, This is a topological signature deviation. The system provides immediate scalar rewards, with positive values ​​encouraging the current strategy and negative values ​​incurring penalties; the weights can be fine-tuned online based on the characteristics of the user group.

[0150] The edge implementation process consists of five steps: collecting risk energy and topology signatures, calculating differences and distances, detecting compliance, normalizing three indicators, synthesizing rewards, and writing them into the policy replay buffer. All calculations use fixed-point arithmetic instructions, with a total latency of 4 milliseconds and an average power consumption of 5 milliwatts, accounting for less than 15% of the wristband system's power consumption.

[0151] In a 14-day continuous wear experiment, the introduction of this reward design reduced the number of convergence rounds of the policy network from 40 to 28, decreased the peak risk energy by 20%, and reduced the average subjective fatigue score of the subjects by 1.8 points (out of 7). The results demonstrate that multidimensional coupled rewards can stabilize gradients and effectively improve intervention effects, providing an efficient, interpretable, and adaptive feedback mechanism for the closed loop of multi-source health risk prediction.

[0152] Preferably, the uploaded update information includes a policy gradient and topological statistics processed with differential privacy noise, wherein the topological statistics are the mean and covariance of the topological signature.

[0153] After each round of policy optimization is completed on the device side, the system needs to securely aggregate the local learning results with cross-user statistical features to improve the model's generalization ability and comply with privacy constraints. To this end, the wristband only sends two types of values ​​that have undergone differential privacy processing during the upload phase: the policy gradient and the topological statistics (i.e., the topological signature mean and covariance).

[0154] First, the edge-side policy network obtains a noisy gradient vector after completing one near-end policy optimization. To ensure that weight updates between different users do not leak individual health information, the system uses a Gaussian mechanism to add noise. Specifically, this involves adjusting the clipped gradient... Superimposed zero-mean Gaussian noise:

[0155] in The standard deviation of noise. It is a unit matrix of the same dimension. The noise amplitude is automatically set based on the privacy budget and the gradient L2 norm pruning threshold, which satisfies the differential privacy definition while avoiding complete distortion of the gradient direction. The pruning threshold is automatically adjusted by continuously monitoring the percentile of the gradient distribution to ensure convergence after noise injection.

[0156] Secondly, topological statistics are used to allow the cloud to understand the multimodal collaborative structure distribution of different users without exposing the complete topological signature. The wristband accumulates several topological signature samples in a local sliding window and calculates the vector mean. With covariance matrix To prevent indirect leakage of statistical characteristics, controllable noise is also added to these two items: [towards...] Superimpose independent Gaussian vectors to make the expected error of each dimension zero; for A symmetric noise matrix is ​​applied, ensuring the result is positive semidefinite. The mean and covariance, after noise processing, are packaged together with the noisy gradient and uploaded using a binary protocol.

[0157] The uploaded data packets contain a timestamp, quantization factor, noise scale, and two types of noisy tensors. The timestamp ensures the correct update order across different batches; the quantization factor compresses floating-point tensors into 16-bit fixed-point numbers, keeping the communication load below 120 kilobytes per day, complying with Bluetooth Low Energy bandwidth constraints. To defend against replay attacks, the data packets are signed with a message authentication code calculated by the edge security element, and the cloud only aggregates them after verifying the signature and privacy budget.

[0158] During the cloud aggregation phase, the server first performs a weighted average of the received noisy gradients, and then dynamically adjusts the learning rate according to the aggregation round. Unbiased estimation is used to recover the cross-user mean and covariance of the topological statistics, thereby updating the batch normalized layer statistical parameters in the potential energy network. Since the noise has zero mean, the aggregation error decays according to the square root law as the sample size increases, without affecting long-term performance.

[0159] The example demonstrates that in a 14-day continuous wear experiment with 30 subjects, the variance of the differential privacy noise was set to 0.05, and the window length of the topological statistics was set to 100. Compared to sending the original gradient, the noise-adding scheme only increased the number of convergence steps of the model in the cloud by 9%, while reducing the maximum amplitude of the individual risk energy curve by 21%, proving a good balance between privacy protection and model performance. Throughout the entire process, the user's original vital signs, movement trajectory, and topological signature details never leave the local device, meeting medical data compliance requirements. Simultaneously, cross-population collaborative improvement is achieved through statistical sharing, providing secure, efficient, and scalable federated learning support for multi-source health risk prediction.

[0160] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting multi-source health risks, characterized in that, Includes the following steps: The synchronous inertial data is decomposed into four-element wavelet decomposition, and the synchronous vital signs data and near-infrared spectral data are pulse density modulated. The pulse density modulated is then input into the photonic pulse neural network along with the visual event pulses to generate an event semantic tensor. Based on the event semantic tensor, a persistent coherent barcode is calculated, and a topological signature is obtained through optimal transmission centroid alignment. The topological signature is embedded into the Hamiltonian constant differential equation to generate a risk energy sequence. The topological signature and the risk energy sequence are written into the edge blockchain, and the gradient of the risk energy sequence with respect to the topological signature is fed back to the photonic pulse neural network. A state vector is formed using risk energy sequence, topological signature, and timestamp. This vector is then input into a proximal policy optimization network with an optimal transmission regularization term to obtain pressure control, step frequency control, breathing control, and illumination control. These control variables are used to drive the actuator to implement intervention using pulse width modulation signals. Rewards are generated based on user compliance. The proximal policy optimization network is updated with the rewards, and the updated information is uploaded. The dissipation coefficient of the Hamiltonian constant differential equation is adjusted based on the rewards.

2. The method according to claim 1, characterized in that, The four-element wavelet decomposition includes: combining the three-axis acceleration signal and the three-axis angular velocity signal into a four-element vector, performing wavelet transform on the four-element vector on a multi-scale convolution kernel, and writing the resulting convolution coefficients into a tensor for processing by a photonic pulse neural network.

3. The method according to claim 1, characterized in that, Pulse density modulation includes generating a one-bit pulse sequence from vital sign data and near-infrared spectral data using a cascaded oversampling modulator, wherein the interval between adjacent pulses maintains a monotonic mapping relationship with the instantaneous amplitude of the input signal.

4. The method according to claim 1, characterized in that, The photonic pulse neural network is composed of a Mach-Zehnder interferometer array, and the phase-shifting element adjusts the refractive index through the thermo-optic effect. The refractive index is updated synchronously according to the pulse time difference-dependent plasticity rule and the contrast reconstruction error.

5. The method according to claim 1, characterized in that, The persistent coherent barcode is obtained by constructing an alpha complex within a fixed-length sliding window and extracting one-dimensional and two-dimensional barcodes. The barcode is then subjected to an entropy-regularized optimal transmission algorithm to calculate the centroid and generate a topological signature.

6. The method according to claim 5, characterized in that, The entropy-regularized optimal transmission algorithm performs a fixed number of iterations in each centroid calculation process to obtain a continuously differentiable alignment result as the topological signature.

7. The method according to claim 1, characterized in that, The gradient of the risk energy sequence to the topological signature is calculated by an automatic differentiation engine, and the control voltage of the phase shift element in the photonic pulse neural network is adjusted at a set update frequency.

8. The method according to claim 1, characterized in that, The pressure control quantity drives the pneumatic soft actuator to adjust the air chamber volume through the pulse width modulation signal; the step frequency control quantity drives the step frequency prompting device to output a beat signal; the breathing control quantity drives the breathing prompting device to output a breathing rhythm signal; and the illumination control quantity drives the illumination adjustment device to change the light intensity.

9. The method according to claim 1, characterized in that, The reward signal consists of the difference between the risk energy sequence, the user compliance flag, and the optimal transmission distance between the topology signature and the individual steady-state topology signature.

10. The method according to claim 1, characterized in that, The uploaded update information includes policy gradients and topological statistics processed with differential privacy noise, wherein the topological statistics are the mean and covariance of the topological signature.

Images