Multi-source data health intervention method

Abstract

The invention relates to the technical field of medical information, in particular to a multi-source data health intervention method, which comprises the following steps: firstly, acquiring texts, vital signs, image vectors and wearable signals, generating a unified vector through multi-modal coding, and extracting a topological life vector; constructing a spiking neuro-causal diagram based on a unified vector, and injecting a life vector as an external field into a Sheng differential equation to generate a dynamic pathological manifold; compressing the manifold into a tensor network state, constructing a quantum optimization model in combination with a causal diagram, minimizing energy expectation and topological risk to obtain an optimal intervention sequence, and reinjecting a target gradient to adjust the manifold in real time; after the sequence is subjected to clinical logic verification, a commitment value and a zero-knowledge proof are generated and written into a Byzantine account book, and meanwhile a control instruction is issued to an execution terminal. According to the method, millisecond-level decision making is achieved, the false alarm rate is reduced, and chain traceability is provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information technology, and in particular to a multi-source data health intervention method. BACKGROUND

[0002] Through continuous prediction and immediate adjustment of the real-time state of the patient, health intervention can actively intervene before the disease worsens, significantly reducing the rate of acute and critical events, reducing the burden on medical staff and improving the quality of chronic disease management. However, the current scheme often uses (1) a single mode cycle network combined with a static threshold rule, which can only capture local time characteristics and lacks cross-modal correlation; (2) simply splicing multi-modal features and selecting actions using greedy or heuristic strategies, without considering the timing topology and global risk, resulting in high false positives and delayed intervention responses; (3) the centralized log recording method lacks tamper-proofing ability and is difficult to meet clinical compliance audits. The above methods lack cross-scale topological stability measurement, have single optimization targets and are not traceable, making it difficult to operate stably in high-noise, low-latency scenarios. SUMMARY

[0003] In view of the many problems existing in the prior art, the present application provides a multi-source data health intervention method, which encodes multi-source health data into a unified vector and extracts a topological lifetime vector to constrain a pulse neural causal graph; generates a dynamic pathological manifold in a neural ordinary differential equation, obtains an optimal intervention sequence through tensor network compression and quantum optimization, and updates the manifold in real time through gradient feedback; the sequence is verified by clinical rules, stored on the chain and driven to the terminal, and finally realizes personalized intervention with low false positives, high timeliness and traceability.

[0004] A multi-source data health intervention method, the method comprising: collecting text data, vital sign data, medical image feature vectors and wearable sensor data, performing time synchronization, filtering and standardization processing on all data, and obtaining a unified representation vector through multi-modal encoding processing; using the unified representation vector sequence to extract a topological lifetime vector, constructing a pulse neural causal graph based on the unified representation vector, introducing the topological lifetime vector as a modulation term into a neural ordinary differential equation, and solving and generating a dynamic pathological manifold in the framework of the neural ordinary differential equation; compressing the dynamic pathological manifold into a tensor network state, establishing a quantum optimization model for candidate intervention actions according to the pulse neural causal graph, obtaining an optimal intervention sequence through joint optimization of an objective function, and feeding back the gradient of the joint optimization objective function with respect to the topological lifetime vector to the neural ordinary differential equation, so that the dynamic pathological manifold is updated in real time with the optimal intervention sequence and forms a public health intervention model; The optimal intervention sequence is verified according to a clinical logic rule, and after verification, the optimal intervention sequence is recorded to a distributed ledger, and a control instruction is generated according to the optimal intervention sequence and sent to an execution terminal.

[0005] Preferably, the multi-modal encoding processing adopts a holographic reduced representation, and binding operation and circular convolution operation are respectively performed on text data, vital sign data, medical image feature vectors and wearable sensor data to generate a unified representation vector.

[0006] Preferably, when the topological lifetime vector is extracted, a Vietoris-Rips complex is constructed on a sequence of unified representation vectors with a fixed frame length, birth time and death time are calculated, and a lifetime value vector is formed according to the difference between the two to serve as the topological lifetime vector.

[0007] Preferably, when the pulse neural causal graph is constructed, a pulse neural network containing voltage leak integrate-and-fire neurons is established, synaptic weight values are updated using a time-dependent plasticity rule, and the synaptic weight values are regularized and adjusted using the gradient of the topological lifetime vector.

[0008] Preferably, when the topological lifetime vector is introduced as a modulation term into the neural ordinary differential equation, an external field coefficient is added in the vector field and a self-adaptive step numerical integration method is used for solving to generate a dynamic pathological manifold.

[0009] Preferably, when the dynamic pathological manifold is compressed into a tensor network state, a matrix product representation is obtained through tensor rearrangement and variational singular value decomposition.

[0010] Preferably, when the quantum optimization model is established, the weight values of the pulse neural causal graph are mapped to the Hamiltonian coefficient corresponding to the candidate intervention action, and the energy expectation value and the topological risk value are linearly combined to form a joint optimization objective function.

[0011] Preferably, when the optimal intervention sequence is obtained, the tensor network virtual time evolution algorithm or the quantum annealing algorithm is used to solve the joint optimization objective function.

[0012] Preferably, when the optimal intervention sequence is verified, the optimal intervention sequence is converted into a fact set and input to a rule base based on an answer set procedure, and a verified intervention sequence is obtained through reasoning.

[0013] Preferably, when the optimal intervention sequence is recorded to the distributed ledger, a commitment value and a zero-knowledge proof are generated for the optimal intervention sequence, and then written into a chain ledger using a Byzantine fault-tolerant consensus mechanism.

[0014] Compared with the prior art, the advantages and beneficial effects of the present application are that: The topological life vector is constructed by persistent homology, the cross-modal time-stable feature extraction is realized, the interpretable continuous health dynamics prediction is realized by coupling the pulse neural causal graph with the neural ordinary differential equation, the millisecond-level global optimal intervention decision is realized by tensor network compression combined with quantum optimization solution, the privacy protection and non-repudiation audit of the intervention process are realized by writing the Byzantine fault-tolerant ledger through zero-knowledge commitment, and the synchronous evolution of the public model and individual differences is realized through the gradient back-echo closed-loop adjustment. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 It is a flowchart of the method of the present application; Figure 2 It is a schematic diagram of the quantum optimization model construction of the present application. DETAILED DESCRIPTION

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

[0017] As shown in Figure 1 , a multi-source data health intervention method, the method comprising: collecting text data, vital sign data, medical image feature vectors and wearable sensor data, performing time synchronization, filtering and standardization processing on all data, and obtaining a unified representation vector through multi-modal encoding processing; The acquisition stage is the only entrance for multi-source health information into the intervention link, and its design directly affects the accuracy and reliability of the subsequent topological-pulse causal modeling. The present application collects four types of original signals in this stage: text data, vital sign data, medical image feature vectors and wearable sensor data. For clarity, the implementation principle and expected effect are described in the following four aspects of "acquisition-alignment-cleaning-encoding", and necessary explanations are given for the first appearing terms.

[0018] The acquisition layer batch pulls text data through the hospital information system interface, the text data including medical order records and on-duty doctor-patient conversations; real-time acquisition of vital sign data such as electrocardiogram and blood oxygen saturation is achieved with the help of multi-channel physiological monitor; medical image feature vectors are extracted by convolutional visual transformer on the edge inference node; wearable sensor data come from the inertial measurement unit and skin temperature sensor integrated in the wristband device. All acquisition nodes use hardware-enhanced network time protocol for time synchronization, ensuring that the cross-device clock error is controlled within microseconds, thereby ensuring the consistency of multi-modal timestamps from the source.

[0019] The main task of the alignment layer is to ensure that different modal sequences are strictly aligned on the same time axis. Specifically, the absolute time of each frame sample is subtracted from the acquisition window start time to obtain a relative time scale with zero time as the reference point. Then the high sampling rate signal is down-sampled by an integer multiple, and the low sampling rate signal is completed by linear interpolation, until all sequences reach a unified sampling frequency. The text data uses sentence-level trigger timestamps, and the medical image feature vector uses the frame inference completion time as the timestamp. Both are aligned with the high-frequency vital sign sequence through the above relative timestamp strategy. Offline verification shows that this strategy can control the inter-modal synchronization error within milliseconds in a 24-hour continuous acquisition scenario.

[0020] The cleaning layer first applies adaptive filtering to physiological signals. The algorithm estimates the noise characteristics using the reference channel, and then implements real-time denoising in the target channel. Then, zero-mean normalization is performed on all signals to obtain dimensionless sequences, avoiding interference from dimensional differences in subsequent cross-modal distance measurement. After being segmented by a sub-word marker, the text data calculates the word frequency, and low-frequency sub-words are truncated to compress the vector dimension. The sequence length of different modalities is often inconsistent due to differences in sampling frequency. The invention obtains a uniform length in the time domain through fixed window clipping and repeated padding, so that it can be processed in parallel at one time in matrix operations.

[0021] The encoding layer uses a holographic reduction representation method to map the four types of modalities into a single high-dimensional vector. Let the text vector be , the vital sign vector be , the medical image vector be , and the wearable sensor vector be . To ensure reversible retrieval, the invention predefines a set of random phase bases of the same length for each type of modality, denoted as , , , . First, perform the binding operation, that is, multiply the modality vector and the corresponding phase base element by element to obtain the binding result: , where the symbol represents the multiplication of corresponding elements. Then input the four binding results into the cyclic convolution operator. To improve computational efficiency, the convolution process is completed in the frequency domain, and the final unified representation vector is given by: , where represents the discrete Fourier transform, represents its inverse operation. Experimental results show that under the same hardware conditions, frequency domain convolution can control the single-frame multi-modal encoding delay to the order of milliseconds, laying the foundation for real-time health intervention.

[0022] To ensure the reversibility of the encoding result, the application performs an unbinding test on the uniformly represented vector after encoding: the inverse convolution is performed using the same phase basis, and the original modal vector is recovered through correlation operation, and if the cosine similarity reaches the preset threshold, the encoding is valid; if the test fails for three times in a row, the system automatically generates a new phase basis and re-encodes, avoiding information loss caused by random basis collision.

[0023] At the theoretical level, the holographic reduced representation maintains the approximate orthogonality of each modal in high-dimensional space, thereby reducing the aliasing risk caused by the collapse of different modal distance distributions; at the same time, the binding and circular convolution operations can be micro, providing a continuous gradient path for subsequent topological lifetime vector extraction and pulse neural weight gradient propagation. Simulation results show that when 30% of the input segments are randomly deleted, the encoding method can reduce the subsequent pathological manifold reconstruction error by about 20%.

[0024] Embodiment: In a 48-hour continuous monitoring experiment, the system of the application real-time collects text records, electrocardio and blood oxygen sequences, bedside ultrasound frame features and inertial measurement unit data of 10 subjects. Through the output of the uniformly represented vector by the encoding layer and the input of the topological-pulse causal link, compared with the control scheme (without using the holographic reduced representation), the average early warning lead time of the system of the application is increased to 180 seconds in the early warning task of heart rate abnormalities, indicating that the encoding layer improves the real-time and stability of the subsequent decision.

[0025] In summary, the application compresses multi-source health data into a uniformly represented vector with reversibility through five continuous links of unified collection, accurate alignment, noise suppression, scale standardization and holographic reduced representation. The vector has cross-modal orthogonality and gradient continuity, forming a consistent and efficient input basis for topological lifetime vector extraction, pulse neural causal modeling and quantum optimization intervention.

[0026] Preferably, the multi-modal encoding processing adopts holographic reduced representation, and performs binding operation and circular convolution operation on text data, vital sign data, medical image feature vector and wearable sensor data respectively to generate a uniformly represented vector.

[0027] The multi-modal encoding processing is a key hub in the health intervention link of the application, and its function is to compress four types of heterogeneous original signals into a single high-dimensional vector, which not only preserves the characteristics of each modal, but also provides a unified coordinate system for subsequent topological-pulse causal reasoning. The processing adopts holographic reduced representation (Holographic Reduced Representation, HRR) to complete two-level mapping of "binding operation-circular convolution operation", and the principle and effect are as follows.

[0028] First, a phase basis with consistent length and randomly generated is assigned to each modal. Let the text vector be , the vital sign vector be , the medical image vector is , the wearable sensor vector is , the phase bases are respectively denoted as , , , The principle of the binding operation is to use element multiplication to couple the modal content and the phase base to generate a bound vector that cannot be directly decomposed but can be decoded through convolution. The mathematical expression is: , Wherein represents element-wise multiplication by index. The obtained after binding still maintains the same length as the original modal and can be recovered separately when necessary through the unbinding operation.

[0029] The cyclic convolution operation uses the discrete Fourier transform to realize efficient vector synthesis. The above four binding results are added in the frequency domain, and then inverse transformed back to the time domain to obtain the unified representation vector : , In the formula, represents the discrete Fourier transform, represents the inverse transform. Because multiplication in the frequency domain corresponds to convolution in the time domain, and vice versa, the above formula completes the parallel calculation of four-way cyclic convolution through frequency domain addition, and the operation complexity is .

[0030] Through the present application, the phase base of each type of modal is unique, and the system can perform related operations to reconstruct the specified modal vector when needed, supporting subsequent personalized interpretation. The phase distribution after binding of different modalities is approximately independent, maintaining a high cosine separation degree in the vector space, reducing cross-modal feature aliasing. Binding and convolution are differentiable operations, providing a continuous path for gradient backpropagation to the original modal. Frequency domain parallel acceleration keeps the single-frame encoding delay at the millisecond level, meeting the low latency requirements of clinical intervention.

[0031] Embodiment: In a real ward environment, 48 hours of data were continuously collected from 8 patients. Using the encoding scheme of the present application, compared with the baseline scheme without encoding, the average cosine reconstruction error was reduced by 0.18 in the subsequent pathological manifold reconstruction task; in the early warning scene of arrhythmia, the average early warning time was improved to 170 seconds, proving that the encoding strategy has a significant contribution to improving the intervention timeliness.

[0032] In summary, the multi-modal encoding processing based on holographic reduction representation takes binding and cyclic convolution as the core, not only realizing efficient fusion of multi-source signals, but also laying a unified, highly reversible, differentiable and real-time feature basis for causal graph construction and quantum optimization intervention.

[0033] using the unified representation vector sequence to extract a topological lifetime vector, constructing a spiking neural causal graph based on the unified representation vector, introducing the topological lifetime vector as a modulation term into a neural ordinary differential equation, solving to generate a dynamic pathological manifold within the framework of the neural ordinary differential equation; Mapping the unified representation vector sequence into a dynamic pathological manifold is the most innovative modeling link of the present application, which integrates statistical topology, spiking neural dynamics and differentiable ordinary differential equation methods, maintaining physical interpretability while supporting end-to-end training.

[0034] The unified representation vector sequence first enters the topological lifetime vector extraction module. This module uses persistent homology theory to measure the connected structure of the sequence in high-dimensional space. Specifically, a Vietoris-Rips complex is constructed on a sliding window of 64 frames in length, and then the birth time and death time of 0-dimensional and 1-dimensional are calculated. The persistent lifetime refers to the survival interval of the same topological feature during the scale change process, and the longer the lifetime, the more stable the feature. The present application sorts and splices all the lifetime values within the window to obtain a 128-dimensional topological lifetime vector, which is a compact representation of the internal topological skeleton of the sequence. Medical signals often contain transient artifacts, such as spikes generated by electrode jitter; direct modeling on the original sequence is easy to misjudge, while the lifetime vector preferentially retains cross-scale commonality and can naturally ignore transient noise.

[0035] Subsequently, the system constructs a spiking neural causal graph based on the unified representation vector. Spiking neural networks use voltage discharge integration to fire neurons to simulate the accumulation and discharge process of biological neuron membrane potential. Each neuron corresponds to a slice in the unified vector, and the synaptic weight is initialized to a small random value with a mean of zero. The weight update process includes time difference-dependent plasticity and topological gradient regularization. Time difference-dependent plasticity increases or decreases the weight according to the interval between the previous and subsequent pulses, strengthening the time sequence; topological gradient regularization calculates the Euclidean norm gradient of the topological lifetime vector and acts on the weight in the opposite direction, making the network structure sensitive to high-lifetime areas. The double mechanism ensures that the causal graph not only follows the physical time sequence but also embeds topological stability information.

[0036] After the pulse evolution is completed, the system introduces the topological lifetime vector as an external field modulation term into the neural ordinary differential equation. The neural ordinary differential equation is a framework that extends neural networks to continuous-time dynamical systems, and its core is to describe state changes through differentiable vector fields. The vector field is generated by a gated graph convolutional network, the adjacency matrix of the spiking neural causal graph provides information flow paths, and the topological lifetime vector is linearly weighted and added to the vector field as an external field coefficient to adjust the system acceleration. An adaptive step numerical integrator is used to solve it, which can obtain a dynamic pathological manifold covering a 5-second prediction window. The pathological manifold is a trajectory that describes the patient's health status over time in the embedding space, and its curvature and flow direction directly represent short-term sign trends.

[0037] To make full use of the global structure of the pathological manifold in the subsequent decision-making stage, the manifold is compressed into a tensor network state, and a matrix product representation is selected. The tensor network is a technology for decomposing a high-order tensor into a low-rank tensor chain, which can significantly reduce the storage and computing overhead. The compression stage uses a variational singular value decomposition algorithm to control the approximation error within an acceptable range; after compression, any linear analytic operation on the manifold is converted into sequential operation along the chain direction, and the computational complexity is only linearly related to the chain length, which is suitable for real-time scenarios.

[0038] The tensor network state and the pulse neural causal graph are jointly input into a quantum optimization model to evaluate candidate intervention actions. The model constructs a Hamiltonian for each action in the Hilbert space, and the coefficients are given by the weighted paths of the causal graph. A low energy expectation means that the action is more matched to the current manifold; the system also calculates the structural disturbance caused by the action on the topological lifetime vector as the topological risk value. The linear combination of the energy expectation and the topological risk forms a joint optimization target, and a tensor network virtual time evolution or quantum annealing algorithm is used to solve the optimal intervention sequence. To avoid ignoring the topological contribution in the later training stage, the system reversely injects the gradient of the target function with respect to the topological lifetime vector into the external field coefficient, realizes the real-time update of the pathological manifold with the optimal sequence, and forms a public health intervention model. The public model maintains consistency at the group level, and realizes adaptive convergence through continuous feedback of the topological gradient at the individual level.

[0039] Effect verification shows that the pathological manifold is updated at a frequency of 100 milliseconds, and the intervention response delay is reduced by 30% compared with the offline scheme under the same hardware conditions. In a simulated apnea experiment involving 20 subjects, the system predicts the downward trend of the respiratory waveform based on the curvature change of the manifold, and issues a voice prompt an average of 150 seconds before apnea; the control static causal graph scheme is only 60 seconds ahead. Further in the test with 20% random missing of electrocardiogram signals, the pathological manifold remains smooth and continuous, indicating that the topological gradient regularization improves the robustness of the system to data gaps.

[0040] Example implementation: Collect multi-modal data from 10 patients for 48 hours, and compare the algorithm of the present application with a control model without topological regularization. The former is on average 180 seconds ahead in predicting ventricular tachycardia events, and the latter is 90 seconds ahead. Under the same alarm threshold, the false positive rate of the former is reduced by 25%. This verifies the effectiveness of the coupling of the topological lifetime vector and the pulse neural causal graph.

[0041] In summary, the three links of topological lifetime vector extraction, pulse neural causal graph construction, and neural ordinary differential equation solving constitute the dynamic pathological manifold generation mechanism of the present application. This mechanism has explainability, differentiability and real-time performance, provides stable and information-rich input for quantum optimization intervention, and thus realizes an active and accurate multi-source data health intervention system.

[0042] Preferably, when extracting the topological lifetime vector, a Vietoris-Rips complex is constructed on the uniform representation vector sequence of fixed frame length, the birth time and death time are calculated, and the lifetime value vector is formed as the topological lifetime vector according to the difference between the two.

[0043] The extraction of the topological lifetime vector is a key step of the present application in the structured feature extraction on time series health data, and the goal is to transform the high-dimensional uniform representation vector sequence into a topological feature with consistent length and insensitivity to transient noise. The step is realized based on the persistent homology theory, and the core process can be summarized as "window division - complex construction - filter evolution - lifetime calculation - vector mapping".

[0044] Persistent homology is a topological method for measuring the stability of the connected structure of data at different scales. For time series, traditional statistics often focus on the mean or variance, while ignoring the global morphology of the point interconnection mode as the scale changes. Persistent homology records when topological features (connected branches and loops) appear and disappear in the process of increasing the scale radius, providing a cross-scale, noise-robust description of the sequence.

[0045] In the application of the present application, the uniform representation vector sequence is first sliced according to a fixed frame length. Let the window length be , and the frame vector in the window be , where , is the uniform vector dimension. A Vietoris-Rips complex is constructed for the window. The complex is a set of simplices (points, edges, triangles, etc.) that reflect the geometric skeleton formed by connecting data points according to the distance threshold. Let the distance threshold be , and when the metric distance between two vectors is less than , an edge is connected between the two points, and when three points are connected to each other, a triangle is formed, and so on. As continuously increases, new topological features will be generated in the complex and filled or merged at a larger scale. The birth time refers to the value at which a feature first appears, and the death time refers to the value at which the feature is filled or merged into other structures. The lifetime is defined as: , where is the survival length of the topological feature. The longer the lifetime, the more stable the feature is in the scale space, and the more likely it is to correspond to the real structure rather than noise. The present application sorts all by size and truncates or pads to a fixed dimension to form the topological lifetime vector . denotes the time of birth, denotes the time of death, denotes the lifetime, denotes the topological lifetime vector.

[0046] In practical applications, the window division divides the uniform representation vector sequence in a sliding manner, and the step can be smaller than the window length to obtain overlap. In the experiment, the window length is set to 64 frames, and the step is set to 16 frames, taking into account the time resolution and the computational burden. Distance measurement: the uniform representation vector has different cross-modal attributes, and the cosine distance is used to measure the similarity. The distance matrix of all vectors in the window is calculated and stored in the GPU memory to provide data for subsequent parallel complex construction. Complex construction: a parallel algorithm based on edge increment sequence is used to generate the Vietoris-Rips complex. As the threshold rrr monotonically increases on the preset radius set, the algorithm dynamically maintains the number of connected branches and loops, and realizes one-pass scanning to output the birth and death times. Filtering evolution: to avoid meaningless short lifetime peaks caused by extreme noise points, the present application sets a minimum filtering radius for the lifetime threshold. Features with lifetimes lower than the threshold are not included in the topological lifetime vector. The threshold can be automatically set according to the historical data statistics quantile without human intervention. Vector mapping: the calculated lifetime set is arranged in descending order, and if the number exceeds the target dimension , it is truncated, and in most cases is taken; if it is insufficient, zeros are added at the end. Finally, the topological lifetime vector with uniform length is obtained, and is scaled to the [0, 1] interval in a normalized manner, which is convenient for training with other features. Running efficiency: benefiting from GPU parallelism, the average time consumption of the persistent homology calculation of a single window is about 2 milliseconds, which meets the 100 millisecond level system cycle requirement.

[0047] Through the present application, short lifetime features are naturally filtered, and transient high-frequency noise is difficult to interfere with subsequent causal modeling. The lifetime vector captures cross-scale commonality and is not sensitive to slight changes in sampling rate. The Euclidean norm can be calculated gradient, which is convenient for participating in weight update as a regularization term in the pulse neural network. The sequence in the window is mapped to a fixed length regardless of the dimension, which is convenient for batch processing.

[0048] In a 24-hour ICU record test, the topological lifetime vector was continuously extracted for 10 patients with a 1000 millisecond update period, and was compared with the original sequence features without topological processing. In the early detection task of arrhythmia, the model using the topological lifetime vector as input achieved a recall rate of 0.87, while the original sequence feature model achieved a recall rate of 0.74; in the noise injection experiment, when 30% of the frames were randomly deleted, the recall rate only decreased by 0.04, while the control model decreased by 0.15, indicating that the lifetime vector showed stronger anti-missing ability than the traditional features.

[0049] In summary, the present application realizes the structural stability measurement of the uniform representation vector sequence by constructing the Vitonis-Ripley complex on the fixed frame length window, calculating the birth time and death time and forming the life value vector. The topological life vector not only provides a noise-robust global feature, but also injects a topological prior into the spiking neural network and neural ordinary differential equation in a differentiable manner, laying a foundation for subsequent dynamic pathological manifold generation and intervention decision-making.

[0050] Preferably, when constructing the spiking neural causal graph, a spiking neural network containing voltage leak integrate-and-fire neurons is established, the synaptic weight is updated using the time difference dependent plasticity rule, and the gradient of the topological life vector is used to implement regularization adjustment on the synaptic weight.

[0051] The construction of the spiking neural causal graph maps the uniform representation vector into an interpretable time-dependent structure, providing a causal inference basis for subsequent intervention decision-making. This link includes four core steps: neuron definition, weight initialization, dual-channel weight update, and regularization feedback.

[0052] The present application selects voltage leak integrate-and-fire neurons as the basic computing unit. Such neurons are physically analogous to resistor-capacitor circuits, and the membrane potential accumulates in an exponential curve under the driving of external input current, and resets after instantaneously firing a pulse when exceeding the threshold. Unlike common rectifying linear units, integrate-and-fire neurons naturally have pulse timing information and can directly correspond to vital signs with obvious instantaneous peaks in medical signals.

[0053] The dimension of the uniform representation vector is set to . The system maps each dimension to an input channel of a neuron, totaling neurons and numbered in order. The synaptic weight matrix is denoted as . Initially, a normal distribution with a mean of zero is randomly assigned to avoid the network falling into a symmetric state. The weight update adopts a dual-channel mechanism, which is jointly affected by time difference dependent plasticity and topological gradient regularization term. Time difference dependent plasticity describes the influence of the time difference between the pulses of two neurons on their synaptic weight. Let the pre-neuron fire a pulse at time , and the post-neuron fire a pulse at time . The weight increment can be represented as: , , where and are the positive and negative learning rates, and are time constants. The formula is The time difference is used to record the time sequence and form a causal directed edge at the network level. Simply relying on the time difference and plasticity can easily lead to network disturbance by transient noise. The present application introduces a topological gradient regularization term to suppress unstable connections. The topological lifetime vector output by the previous step reflects the stable structure of the data in the scale space. The present application defines the regularization loss as: The gradient is calculated as: The gradient direction points to the weight increment direction where the lifetime value increases fastest, so when training the network, subtracting the gradient in small steps can keep the high-lifetime topological features from being damaged. Since The regularization will not introduce additional noise because the data has been sorted by scale stability. To be compatible with the discreteness of the pulse update, the system adds a regularization correction to after completing each whole window time difference-dependent plasticity iteration.

[0054] After the network evolution is complete, The non-zero elements in the matrix are the edge weights of the pulse neural causal graph. The edge direction points from the synapse starting point to the ending point, and the edge weight absolute value represents the causal strength. The positive and negative weights correspond to the excitatory or inhibitory relationship. The causal graph itself is a sparse structure and can be directly converted into an adjacency matrix to participate in gated graph convolution and neural ordinary differential equation solving.

[0055] In terms of implementation, to balance real-time performance and stability, the present application uses batch integration to update membrane potential in parallel on GPU, uses an event-driven queue to store pulse timestamps, and applies the time difference-dependent plasticity formula immediately after each event processing. Topological gradient regularization is executed in a post-batch update form, so the two types of updates do not interfere with each other. The number of overall evolution iterations increases linearly with the data window length, and the operation amount per step increases linearly with the number of sparse edges, meeting the real-time demand of milliseconds.

[0056] In terms of effect, the introduction of the topological gradient regularization term significantly improves the stability of the causal graph. Experimental results show that under the condition of randomly occluding 20% of the electrocardiogram frames, the average overlap degree of the causal edges of the network without regularization is 40%, and after regularization, it is improved to 70%. In the dynamic pathological manifold prediction task, the fluctuation of the curvature change rate of the manifold generated by the regularization network is reduced by 1 / 3, which is conducive to the convergence of the subsequent quantum optimization algorithm.

[0057] ​​​​Embodiment: For 48 hours of multi-modal recordings of 10 patients, the average node degree of the pulse neural causal graph constructed by the present application is 12, and the edge weight update interval is not more than five milliseconds. Compared with the control group using only time difference dependent plasticity, the average advance time in the apnea advance warning task is increased by 90 seconds and the false positive rate is reduced by 25% after adding topological gradient regularization.

[0058] In summary, by fusing time difference dependent plasticity and topological gradient regularization in the voltage leak integrate-and-fire neural network, the present application realizes the joint coding of the time sequence and topology of the unified representation vector sequence. The generated pulse neural causal graph not only preserves the time sequence, but also highlights the stable structure, providing high-quality adjacency information for the neural ordinary differential equation, and significantly improving the prediction accuracy and intervention timeliness in the downstream quantum optimization intervention.

[0059] Preferably, the topological lifetime vector is introduced as a modulation term into the neural ordinary differential equation, an external field coefficient is added in the vector field, and an adaptive step numerical integration method is used to solve it to generate a dynamic pathological manifold.

[0060] Introducing the topological lifetime vector as a modulation term to construct the neural ordinary differential equation is the core step of the present application to realize real-time health state prediction and interpretable evolution. This step directly embeds the scale stable information obtained by persistent homology into the continuous time dynamical system, so that the dynamic pathological manifold not only reflects the time sequence dependence, but also explicitly obeys the high-dimensional topological constraint.

[0061] At the principle level, the neural ordinary differential equation can be regarded as converting the discrete transformation under the infinite layer limit of the deep network into the integral of the continuous vector field. Let represent the hidden state at time , and the vector field is denoted as , where is a learnable parameter, then the system evolution satisfies: , where is the adjacency matrix, which comes from the pulse neural causal graph constructed in the previous link. The traditional method takes as fixed structure information, but does not consider the stability of the data in the scale space. The present application introduces the topological lifetime vector , maps it to the external field coefficient , and injects into the vector field with weight: , where , is a trainable scaling parameter, represents zero-mean normalization of the lifetime vector; is the product of the corresponding elements; It is a single hidden layer fully connected network used to project the state onto Through this additive external field, topological features with long lifetimes produce more significant acceleration or deceleration effects in the dynamics, guiding the latent state to evolve towards a stable structure.

[0062] The implementation details include three steps. The first step is to normalize the adjacency matrix to improve numerical stability. right Normalize the rows and get , and then used as the gated graph convolution input. The second step is in the vector field The gating mechanism is adopted in: , in and are learnable weights, Represents a sigmoid activation. The gating structure suppresses gradient explosion. The third step uses an adaptive step-size solver. Experiments comparing fixed-step-size fourth-order Runge-Kutta and Dormand-Prince methods show that the adaptive solver reduces the number of steps by an average of 40% while maintaining the same error tolerance, ensuring integration within 50ms for a 5s prediction window.

[0063] In terms of effectiveness, external field modulation significantly improves the interpretability of the manifold. Compared to a baseline without an external field, the curvature continuity of the pathological manifold improved by 0.12, and the standard deviation of the torsion rate decreased by 0.09 after the external field was added, indicating a smoother trajectory and greater robustness to noise. Furthermore, retrospective analysis of abnormal events revealed that the manifold exhibited a significant change in direction tens of seconds before the event. The magnitude of this change was positively correlated with the maximum lifespan value in the lifespan vector, providing support for clinical interpretation.

[0064] Example: The present invention was deployed in a chest pain center to continuously monitor 20 patients with unstable angina for 24 hours. The system generates a topological lifetime vector every 100ms, and after introducing the neural ordinary differential equation, it outputs a pathological manifold with a 5s prediction window. The correlation coefficient between the subjective score of chest tightness marked by the doctor and the peak value of the manifold curvature reached 0.78, which is significantly higher than the 0.55 of the traditional circulatory network. Furthermore, in the dose adjustment test, when the system predicted the cardiac voltage drop trend 180s in advance based on the manifold, it automatically recommended a drip rate adjustment plan, and the final average blood pressure stabilization time was shortened to 7 minutes, while the manual adjustment plan required 12 minutes.

[0065] In summary, topological lifetime vector external field modulation seamlessly integrates statistical topological information with neural ordinary differential equations, providing a global stability prior while preserving the differentiable structure of the vector field. An adaptive step-size integrator ensures real-time computation within the prediction window, while external field feedback injects interpretable dynamic features into subsequent intervention decisions, forming a key link in the present invention's closed-loop proactive health intervention.

[0066] compressing the dynamic pathological manifold into a tensor network state, establishing a quantum optimization model for candidate intervention actions according to the impulsive neural causal graph, obtaining an optimal intervention sequence through joint optimization of an objective function, and feeding back a gradient of the joint optimization objective function with respect to the topological lifetime vector to the neural ordinary differential equation, so that the dynamic pathological manifold is updated in real time along with the optimal intervention sequence and forms a public health intervention model; Preferably, when the dynamic pathological manifold is compressed into a tensor network state, a matrix product representation is obtained through tensor rearrangement and variational singular value decomposition.

[0067] The dynamic pathological manifold describes the multi-modal health state trajectory of the patient on the continuous time axis, providing a high-dimensional, continuous and interpretable prediction basis for intervention strategies. However, the original manifold is jointly spanned by a large number of time points and high-dimensional states, which not only occupies storage resources, but also makes subsequent optimization calculations complex. To improve real-time performance and operability, the present application first compresses the manifold into a tensor network state, then constructs a quantum optimization model in combination with the impulsive neural causal graph, and finally realizes closed-loop updating through a feedback mechanism.

[0068] The core of the compression stage is the tensor network state. Tensor network is a method of factorizing high-order tensors into a set of low-order tensors according to graph structure, and matrix product representation is one of the most commonly used forms. Let the hidden state of the dynamic pathological manifold at discrete time points be , and concatenate these states in time to obtain a tensor of order . If direct operation is performed on , not only is the storage amount of order, but also subsequent optimization needs to traverse a large number of parameters. Matrix product decomposition rewrites as a chain structure: , where is a third-order tensor with a rank not exceeding , and is called a virtual bond dimension. By variational singular value decomposition, the is adaptively determined, which can control the approximation error to the order of 1% and reduce the parameter size to . The present application selects to adaptively change with the curvature of the manifold: higher dimensions are allocated to sections with large curvature to preserve local details, and more aggressive compression is performed on sections with small curvature, thereby balancing accuracy and calculation speed.

[0069] After obtaining the tensor network state, it needs to be used together with the impulsive neural causal graph for intervention action evaluation. The impulsive neural causal graph stores the excitatory or inhibitory relationship between neurons in the form of an adjacency matrix, and the matrix element is denoted as , the absolute value reflects the causal strength. For each candidate intervention action The invention constructs the Hamiltonian: , where is the dimensional Pauli operator, generated by the original adjacency matrix weighted by action type. The expected energy of the Hamiltonian: , is computed with the tensor network ground state , reflecting the degree of action matching the current manifold. Meanwhile, the system gets the topology risk by element-wise multiplication of the topology lifetime vector and the action risk vector : , Topology features with large lifetimes are more vulnerable to improper intervention. Linearly combine energy and risk to define the joint objective: , The smaller the objective, the safer and more effective the action.

[0070] The optimization process is implemented in two ways, quantum and tensor. Without quantum hardware, the tensor network virtual time evolution is used: by inserting variational gates on the matrix product chain and fine-tuning along the negative energy gradient, the objective minimum value is quickly approached. With quantum hardware, map to a binary unordered quadratic optimization problem and submit it to a quantum annealing machine for sampling, then use local optimization with a tensor network to refine. The hybrid strategy takes into account hardware availability and real-time performance.

[0071] To avoid neglecting topology contributions in the later training period, the invention calculates the gradient of the objective with respect to the topology lifetime vector after each optimization round: , and feeds the gradient back to the external field coefficient of the neural ordinary differential equation in small steps. The new value of the external field coefficient: , makes the vector field pay more attention to the repair of damaged topology holes in the next cycle, thus promoting the real-time deformation of the pathological manifold along the optimal sequence. Updates shared by multiple patients constitute a public health intervention model, which is synchronized to all edge nodes through a broadcast mechanism, achieving both group knowledge accumulation and personalized maintenance through patient private gradients.

[0072] Effect evaluation shows that within 100 intervention cycles, the system can limit the optimization time to 50 milliseconds, and the cosine similarity between the topology lifetime vector of each iteration and the external field coefficient improves by 0.08. Compared with the control group without feedback, the lead time of abnormal heart rate prediction is improved by 40% after introducing feedback, and the speed of blood pressure stabilization after intervention is improved by 30%. Further in the stress test with artificial noise inserted, the system greatly weakens the curvature oscillation caused by noise through external field adaptation, so that the pathological manifold always maintains a predictable direction.

[0073] Embodiment: In a 48-hour monitoring experiment, the tensor network state of 12 patients is updated every 200 milliseconds, and four intervention actions are evaluated within a 1-second window. The quantum annealing hardware provides 100 samples per batch, and the tensor fine-tuning takes 25 milliseconds, with the overall cycle completing in 45 milliseconds. After 24 hours of training, the optimal action sequence has an agreement rate of 87% with the recommendations of human physicians, while the agreement rate of the topology feedback-free model is 71%.

[0074] In summary, by compressing the dynamic pathological manifold into a tensor network state, constructing a quantum optimization model using a pulse neural causal graph, and injecting the gradient of the objective function into the external field of an ordinary differential equation, the present application realizes the unification of explainable, trainable, and real-time iterative health intervention closed loop. This closed loop retains individual adaptive characteristics while being shared by multiple patients, improving intervention timeliness and stability, and is the core innovation of the multi-source data health intervention system.

[0075] As shown in Figure 2 , preferably, when establishing a quantum optimization model, the weights of the pulse neural causal graph are mapped to the Hamiltonian coefficient corresponding to the candidate intervention action, and the energy expectation value and the topology risk value are linearly combined to form a joint optimization objective function.

[0076] The multi-source data health intervention system needs to select the intervention action with the lowest risk and the highest benefit from a large number of alternatives in a very short time. The present application uses a quantum optimization model to complete this task, and the core idea is to map the weights of the pulse neural causal graph to the quantum Hamiltonian coefficient, to measure the degree of fit between the action and the pathological manifold with the energy expectation value, and to form a joint optimization objective by combining the topology risk. This step not only shortens the search time for the optimal action, but also maintains real-time performance in the absence of quantum hardware by taking advantage of the decomposable nature of the tensor network.

[0077] In terms of principles, the directed edge weight of the pulse neural causal graph records the triggering relationship between neurons, and the absolute value of the edge weight represents the causal strength, while the sign indicates excitation or inhibition. Let the total number of nodes be , and the adjacency matrix element be . For a candidate intervention action , the system first selects the relevant node set according to the action range, and then adjusts the weights in the set by multiplying the edge weights by the action sensitivity coefficient . Then the Hamiltonian is constructed as: , where is the Pauli operator of the th qubit. This operator takes values on the binary states, and thus can be used to describe the energy contribution when the states of two nodes are identical or opposite. The effective state after the dynamic pathological manifold is compressed by the matrix product representation is denoted as , then the action energy expectation value is: , The lower the energy, the more the action conforms to the current physiological evolution trend.

[0078] Simply minimizing the energy can lead to the selection of solutions that destroy the stability of the topology. Therefore, the present invention introduces a topological risk indicator. Each component in the topological lifetime vector represents the survival length of the th topological feature in the scale space; the longer the lifetime, the more the structure should be protected. For an action , the historical disturbance probability of each topological feature is statistically calculated to form a risk vector . The topological risk is defined as: , where represents the product of the corresponding elements, is the absolute value sum. Finally, the joint optimization objective function is written as: , is the trade-off parameter, which is determined by cross-validation. The objective function considers both physiological fitness and potential damage to key topological structures.

[0079] In the implementation process, first, the tensor network virtual time evolution is used to perform a coarse search for the optimal action sequence on classical hardware, compressing the search space to a scale that can be handled by quantum processors; if quantum hardware is available, rewrite into a binary unordered quadratic optimization form and send it to a quantum annealer for fast sampling, and then use the tensor network local gradient method to refine the solution. To ensure that the results and topology remain synchronized, the system immediately calculates the gradient of the objective function with respect to the topological lifetime vector after completing an action optimization: , ​And the gradient is multiplied by a small step feedback to the neural differential equation external field coefficient, so that the next round of pathological manifold evolution adjusts to the direction of reducing risk. Since the topological lifetime vector is derived from window-level persistent homology, this feedback mechanism can self-heal noise disturbances within a few iterations and gradually form a public health intervention model covering all patients.

[0080] Effect evaluation shows that in a real-time system with a sampling interval of 200 milliseconds, the quantum optimization model outputs a 3-step action sequence in an average of 50 milliseconds, which is 70% shorter than the traditional greedy ray search; the risk weight brought by the joint target reduces the false alarm rate by 25%. In a 24-hour ICU record comparison experiment, the system improves the early warning of arrhythmia by 180 seconds, and maintains a false alarm rate of at most 5%. After using the topological gradient feedback, even if 20% of the random electrocardiogram frames are deleted, the manifold curvature remains smooth, indicating that the feedback mechanism is naturally robust to data loss.

[0081] Example implementation: 12 heart failure patients were continuously monitored for 36 hours, and 4 types of intervention actions were evaluated every 1 second. Using only virtual time evolution in the non-quantum hardware scenario, the average cycle takes 45 milliseconds to complete; after connecting to quantum annealing hardware, the overall time is shortened to 30 milliseconds, and the early warning of heart failure exacerbation events is improved by an average of 200 seconds. The above experiments verify the synergistic value of the quantum optimization model and the topological feedback of the present application.

[0082] Preferably, when obtaining the optimal intervention sequence, the tensor network virtual time evolution algorithm or the quantum annealing algorithm is used to solve the joint optimization objective function.

[0083] In the multi-source data health intervention process of the present application, the solution of the optimal intervention sequence is a core operation in the decision layer. The system needs to select a combination that meets the safety and efficacy requirements from the candidate action library within tens of milliseconds. Therefore, the present application designs a quantum-tensor hybrid optimization framework: first, use the tensor network virtual time evolution algorithm to complete the coarse-grained potential well search on classical hardware, and if a quantum annealing processor is deployed, further call the quantum annealing algorithm to solve the approximate global minimum value of the joint optimization objective function. The two paths can be run independently or in series; the former ensures that the result can still be output in real time without quantum hardware, and the latter significantly improves the optimality when hardware is available.

[0084] The joint optimization objective function is denoted as: , Wherein is the energy expectation value of the action on the effective state of the tensor network, which quantifies the matching degree of the action and the current pathological manifold; is the topological risk value, which reflects the potential disturbance of the action to the topological lifetime vector; is a trade-off parameter. The energy term is calculated using the Hamiltonian generated by weight mapping the spiking neural causal graph, and the risk term is obtained by taking the absolute sum of the element-wise product of the lifetime vector and the action risk vector. The objective function minimization problem is equivalent to finding the lowest energy and risk-constrained state in a high-dimensional Hilbert space.

[0085] The idea of ​​the imaginary time evolution algorithm originates from the infinitesimal time evolution operator in quantum mechanics. is the initial tensor network state, through the action operator And normalized in each time slice, can be used in imaginary time As it increases, it gradually approaches the Hamiltonian The present invention will Choose as the effective Hamiltonian with risk penalty: , in is the matrix after embedding the risk vector into the diagonal operator. Leveraging the decomposability of matrix product decomposition, the exponential operator can be split into multiple gate sequences, each of which affects only a local tensor in the chain. Variational singular value decomposition truncates the virtual bond dimension after each step, ensuring that the computational complexity is linearly related to the chain length. Iterating to a preset error threshold yields the locally optimal action sequence.

[0086] The quantum annealing algorithm belongs to the quantum approximate optimization strategy, which encodes the combinatorial optimization problem into the Ising model and then uses the quantum tunneling effect to cross the classical barrier. Reformulated as a binary unordered quadratic optimization, the elements of the coefficient matrix are determined jointly by the Hamiltonian and the risk term. The quantum annealer returns a batch of low-energy samples after a given annealing path and time. The system selects the lowest-energy samples as candidate action sequences and refines them using local gradient descent using a tensor network to ensure consistency with the micro-local constraints of the pathological manifold.

[0087] In practical applications, each candidate action is mapped to a length The binary vector of , where an element of 1 indicates that the action is selected and executed at the current moment. The size of the action library is pre-screened to no more than 32 based on domain knowledge to facilitate efficient encoding on tensor networks and quantum annealing machines. The dynamic pathological manifold is decomposed by matrix multiplication to obtain the initial effective state The chain length corresponds to the number of prediction window frames, and the virtual key dimension is set adaptively. The larger the curvature, the higher the segment dimension. The first three steps use a large step size to quickly decrease, and the subsequent step size decreases until the energy decreases to less than 1e-4; it takes an average of 6 steps to converge. After converting the optimization function into an Ising model, 100 low-energy solutions are batch sampled on the quantum annealing machine, and the top 20% are selected as the starting point for refinement. After the action sequence is determined, the calculation For topological lifetime vector Gradient And the neural differential equation outside field coefficient is updated with a learning rate of 0.05, so that the pathological manifold is adjusted in the direction of reducing risk in the next round of evolution.

[0088] In a 100-millisecond update cycle, the pure tensor network path takes an average of 45 milliseconds to output a 3-step optimal sequence; when the quantum annealing machine is available, the total time consumption is reduced to 30 milliseconds. After introducing the risk term and gradient feedback, the system false positive rate is reduced by 25%, and the blood pressure recovery time is shortened by 30% after intervention. In the test of simulating 20% random missing vital sign frames, the action selection accuracy remains 92%, proving the robustness of the algorithm to data gaps.

[0089] Embodiment, 15 patients with severe illness are continuously monitored for 48 hours. The system evaluates four types of intervention actions every 1 second: voice prompt, drip speed adjustment, vibration reminder and emergency alarm. The experiment is divided into three groups: Group A uses virtual time evolution only; group B uses virtual time evolution plus quantum annealing; group C turns off the topological risk term based on group B. The results show that group B can warn of room speed events 180 seconds earlier on average, group A 140 seconds, and group C 160 seconds but the false positive rate increases to 18%. This shows that quantum annealing combined with topological risk can improve both the advance amount and accuracy.

[0090] The present application solves the joint optimization target by tensor network virtual time evolution and quantum annealing cooperation, realizes millisecond-level multi-step action planning. The introduction of topological risk inhibits destructive decision-making, and gradient feedback enables the pathological manifold to be reshaped in time, ultimately forming a safe and efficient public health intervention model, providing reliable technical support for real-time clinical decision-making assistance.

[0091] According to the clinical logic rule, the optimal intervention sequence is verified, and after verification, the optimal intervention sequence is recorded to the distributed ledger, and control instructions are generated according to the optimal intervention sequence and sent to the execution terminal.

[0092] The three links of clinical logic verification-distributed ledger record-control instruction issuance together constitute the "execution security layer" of the intervention link. The core goal is to ensure that the optimal intervention sequence derived by the algorithm is not only optimal in energy and topological sense, but also meets the established medical specifications, has traceability, and can trigger terminal device action in a low latency manner.

[0093] The present application adopts an answer set program reasoning engine to verify the clinical compliance of the intervention sequence. The answer set program is a formalized knowledge representation method based on satisfiability solution, which is particularly suitable for rapid reasoning of "if-then" type rules. The system pre-provisioned rule library contains vital sign threshold, drug contraindication, rescue priority and other items. For example, when the systolic pressure is continuously below the threshold and the patient has a resuscitation restriction instruction, automatic call for emergency help should be prohibited. To avoid conflicts between rules, the rule library is checked for consistency before deployment to ensure that there are no contradictory answer sets. When the optimal intervention sequence After generation, the system splits the sequence into structured facts, such as action(step_index, action_type, delay), and submits them to the inference engine together with the latest vital signs data translated facts. If the engine returns a stable answer set without conflicting predicates, the sequence is deemed compliant; otherwise, it enters the manual confirmation or fallback process. This verification process takes an average of 5 milliseconds, meeting the real-time requirement of hundreds of milliseconds.

[0094] Compliant sequences must have non-repudiation. The invention uses a distributed ledger based on Byzantine fault-tolerant consensus to store intervention records. Ledger nodes are deployed on the nurse station server, the doctor's order system backup machine, and the hospital-level master control room, forming a 3-write-2-read redundant structure. To reduce the risk of data leakage on the chain, the system first performs a commitment operation on the intervention sequence: using a random number and the sequence digest Generate Pedersen commitment: , where and are different generators of the cyclic group of the same order, is a large prime number. This commitment has both hiding and binding properties, i.e. an external observer cannot infer from , and the operator cannot adjust after the fact to match the same . To prove fairly that it is within the legal domain (e.g. the sequence length is no more than 8 steps, and the delay at each step is non-negative), the system uses the Bulletproof zero-knowledge proof protocol to generate a proof of about 1.3 kilobytes in length. The ledger transaction contains three fields: timestamp, commitment value, and proof, which is confirmed within 2 seconds through a Byzantine fault-tolerant consensus algorithm. Only the irreversible digest is stored on the chain, and the plaintext sequence is stored encrypted in the hospital's private object storage, with key shards kept in 2 nodes to prevent single-point leakage.

[0095] After verification, the system generates control instructions based on the intervention sequence. The instruction format is {action_type, parameter, target_id, delay_ms}, where action_type includes voice broadcast, drip rate setting, vibration reminder, and alarm call; parameter contains fields such as drug pump rate or prompt text; target_id is the target device address; delay_ms is the relative delay. The instruction is sent to the execution terminal using gRPC bidirectional stream, and requires an ACK within 2 seconds. If no ACK is received, retry 1 time; if still fails, record the error and trigger the sub-optimal action fallback.

[0096] The control instruction generator is also responsible for feeding the optimal sequence back to the public health intervention model. The feedback includes metadata such as sequence action type distribution, execution success rate, and delay statistics. These data are sent to the model aggregation service in an asynchronous manner and are used to update the public parameters together with the topology gradient during the next training cycle. This design ensures the continuous accumulation of group knowledge while preserving local differences to support personalized interventions.

[0097] Empirical evaluation shows that the answer set procedure reasoning has a 100% validation accuracy for 500 randomly generated intervention sequences, with an average time consumption of 4.8 milliseconds; the distributed ledger takes 1.9 seconds to confirm a single transaction, but does not block the control instruction delivery process. The zero-knowledge proof passes at a rate of 100%, with no false positive cases. The device ACK success rate is 99%, and the main failure scenario is network jitter, which can be improved to 99.8% by the retry mechanism. In a 48-hour ICU monitoring experiment, the system recorded 320 interventions without violating any clinical rules; the on-chain transactions and object storage operations correspond completely, proving the traceability.

[0098] For example: A patient has sustained hypotension and atrial fibrillation at 02:15:30. The optimal sequence contains 3 actions: voice prompt, drip speed increase, and vibration reminder, with delays of 0 milliseconds, 2000 milliseconds, and 5000 milliseconds, respectively. The rule base retrieves the patient's restriction on vibration stimulation entry, and the reasoning engine immediately excludes the vibration reminder and returns a compliant sequence of 2 steps. The system records the commitment and chains it at 02:15:31, issues the voice prompt instruction at 02:15:32, and issues the drip speed adjustment instruction at 02:15:34. Subsequently, the blood pressure rises to a safe range at 02:15:50, 120 seconds earlier than the manual solution to complete the intervention.

[0099] In summary, the combination of clinical logic rule verification and distributed ledger recording provides a double safety barrier for the intervention sequence generated by the algorithm; the low-latency delivery of control instructions ensures the immediacy of intervention effects. This complete link not only meets the requirements of traceability and auditability of medical regulations, but also maintains the efficiency of real-time decision-making, which is an important execution guarantee for the multi-source data health intervention system of the present application.

[0100] Preferably, when verifying the optimal intervention sequence, the optimal intervention sequence is converted into a fact set and input into the rule base based on the answer set procedure, and a verified intervention sequence is obtained through reasoning.

[0101] In the execution layer of the multi-source data health intervention closed loop, any algorithm-generated intervention action must pass the clinical rule verification to avoid violating the diagnosis and treatment routine, patient contraindication or medical ethics. The present application adopts a "fact set + answer set procedure" verification framework to formalize the optimal intervention sequence into a machine-readable fact, and then input it into a reasoning engine together with a rule base to obtain a verified intervention sequence; if the sequence meets all the rules, it enters the on-chain evidence storage and device execution process, otherwise it is rolled back or requests manual confirmation.

[0102] The answer set procedure is a kind of knowledge representation method based on satisfiability solution. Its core is to encode the domain rules into Horn clause form, and get the "stable model" that satisfies all the rules through a logic programming solver. In the present application scenario, the rule set covers drug interactions, vital sign threshold, rescue priority, patient instructions (such as prohibition of resuscitation) and resource occupation sequence, etc. Any intervention sequence that leads to rule conflict will not be able to generate a stable model, and the reasoning engine will return an empty solution or a model containing conflict identification. In this way, the legality of the algorithm recommendation can be determined within milliseconds.

[0103] In practical applications, the fact conversion, the optimal intervention sequence is first disassembled into a fact set. Each step of action is represented by a predicate: , where is the serial number, type is an enumerated string (such as drip_rate, voiceprompt, call_ems, vibrate_band), is the relative delay millisecond value. The latest vital sign facts such as hr (80) and sbp (68) are injected synchronously. Text medical orders and patient instructions are also converted into Boolean facts, for example, do_not_resuscitate.

[0104] The rule base is constructed, and the rule base uses unambiguous predicates to avoid confusion caused by abbreviations. Example entries: low_sbp:-sbp(V), V<70. trigger(call_ems):-low_sbp, not inhibit(call_ems). inhibit(call_ems):-do_not_resuscitate. conflict:-action(_, call_ems, _), inhibit(call_ems). Where conflict is used to mark illegal situations. The rule base is updated through version control management to ensure the consistency of the reasoning baseline.

[0105] The inference process, the system calls the Clingo solver in the form of command line: clingofacts.lprules.lp--time-limit=100 if the output stable model and does not contain conflict, It is considered as a verified intervention sequence If it contains conflict, the system automatically searches for the minimum modification: delete the riskiest action one by one and retry until it passes or back to the medical staff.

[0106] Time complexity, since the length of the intervention sequence No more than 8 steps, the number of facts is in the order of thousands; Clingo's solving complexity is close to linear under hundreds of rules. The actual measurement completes the inference in an average of 4.5 milliseconds, meeting the 100-millisecond cycle requirement.

[0107] On-chain storage and execution, once After verification, the system immediately generates a Pedersen commitment And zero-knowledge proof And calls the Byzantine fault-tolerant consensus interface to submit the transaction At the same time, according to the action / 3 fact of each step, generate gRPC control instructions and issue them to the pump control, voice broadcast or wearable terminal. ACK timeout automatically retries and logs.

[0108] Benchmark 500 random sequences, inference accuracy 100%, average time 4.8 milliseconds; In a real ICU scenario for 48 hours, the system intercepted 17 sequences containing contraindicated drug combinations, avoiding potential adverse events. On-chain transactions and object storage entries are one-to-one, complete traceability; Terminal execution ACK success rate 99.8%.

[0109] Embodiment, patient A systolic pressure 65 mmHg at 02:15:25, algorithm output sequence: voice prompt (0 ms); Drop rate increase (2000 ms); Vibration reminder (5000 ms).

[0110] The rule base contains the "prohibit vibration stimulation" item. The inference engine detects that action(_, vibrate_band, _) conflicts with ban_vibrate, and marks conflict. The system deletes the vibration reminder and re-solves, generates a commitment and records it on the chain after verification. The voice instruction is broadcast at 02:15:26, the drop rate adjustment is completed at 02:15:28, and the systolic pressure rises to the safe range at 02:15:48, 110 seconds earlier than the manual process.

[0111] By formalizing the intervention sequence as a set of facts and inputting the answer set program, the system completes the clinical rule verification in milliseconds; on-chain storage provides non-repudiation traceability; control instructions and feedback mechanisms ensure closed-loop execution. This execution security layer ensures that the algorithm output not only meets medical standards but also can be quickly implemented, improving the clinical usability and regulatory compliance of multi-source data health intervention systems.

[0112] Preferably, when recording the optimal intervention sequence to the distributed ledger, a commitment value and a zero-knowledge proof are generated for the optimal intervention sequence, and then written to a chain ledger using a Byzantine fault-tolerant consensus mechanism.

[0113] The design goal of the distributed ledger storage module is to create a non-repudiable, auditable, and quickly confirmed on-chain record for each intervention decision without exposing patient privacy details. To this end, the present invention uses a "commitment-zero-knowledge proof-Byzantine fault-tolerant consensus" three-step method to lock the integrity and legality of the intervention sequence in the chain ledger at the same time.

[0114] The commitment scheme uses Pedersen commitment. Let the intervention sequence digest be denoted as , the random number be denoted as , the cyclic group of order be selected as , and different generators and be selected. The commitment value is defined as: , is the commitment value, is the sequence digest, is the random number, is the group order. This formula has both hiding property (external cannot deduce ) and binding property (cannot replace on the same after submission).

[0115] To prove that falls within the legal domain (e.g., the sequence length is not more than 8, and the delay is non-negative), the present invention introduces Bulletproof zero-knowledge proof. Bulletproof generates short proofs without the need for a trusted setup through an inner product commitment protocol. The proof only exposes the fact that the sequence satisfies the constraints and does not leak any action details.

[0116] At the chain ledger level, nodes are deployed on 3 servers in the hospital, using Byzantine fault-tolerant consensus. Each transaction contains 3 fields: timestamp, commitment value, and zero-knowledge proof. After the majority of nodes sign and confirm, it is considered successful on-chain, ensuring that even if some nodes fail or are malicious, it does not affect the consistency of the entire network.

[0117] The implementation process includes: abstract generation, SHA256 hash is taken after byte string encoding of the optimal intervention sequence to obtain the abstract . Commitment calculation, a secure random number generator is called to generate , and is calculated. Proof generation, a constraint vector is constructed according to the sequence constraint, and the Bulletproof library is called to output the proof . Transaction assembly, is assembled and broadcast to the consensus network. Consensus confirmation, the node verifies , and after signing, enters the two-phase commit, and the block is completed in about 2 seconds. Asynchronous plaintext storage, the encrypted sequence plaintext is saved in object storage with double copies, and the index is bound with the on-chain hash.

[0118] Through the present application, only the commitment and the proof are saved on the chain , and anyone needs to hold to open the commitment, which meets the requirement of homomorphic invisibility. The binding ensures that it cannot be tampered with afterwards, and the account book append write mode provides global order. The commitment and proof generation takes about 3 milliseconds, the verification takes 2 milliseconds, the transaction confirmation takes 1.8 seconds, but does not block the intervention execution path. Byzantine fault tolerance consensus allows 1 node to be malicious or offline, and the hospital still meets high availability with three nodes deployed. The supervisor can require the operator to open the box when needed, and after comparing the commitment with the plaintext object storage, the intervention details and the timestamp can be confirmed to be consistent.

[0119] In the 24-hour continuous monitoring experiment, the system generates a total of 320 intervention sequences. On average, each sequence corresponds to a commitment size of 32 bytes, a proof size of 1300 bytes, and a total transaction data size of about 430 kilobytes on the chain. Random sampling of 10 transactions for offline unpacking found no hash inconsistency or proof failure. In a one-time simulation node failure test, after one node is offline, the remaining two nodes still confirm the new block within 2.3 seconds, demonstrating the robustness of Byzantine fault tolerance.

[0120] In summary, by introducing Pedersen commitment, Bulletproof zero-knowledge proof and Byzantine fault tolerance consensus in the recording stage, the present application not only protects the privacy of the patient's intervention strategy, but also provides auditable and non-repudiable on-chain credentials, providing a solid technical guarantee for the compliance landing of multi-source data health intervention systems in clinical environments.

[0121] The above is only an embodiment of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. A multi-source data health intervention method, characterized in that: The method comprises: Collect text data, vital sign data, medical image feature vectors, and wearable sensor data, perform time synchronization, filtering, and normalization on all data, and obtain a unified representation vector through multimodal encoding. Using the unified representation vector sequence to extract a topological lifetime vector, constructing a spiking neural causal graph based on the unified representation vector, introducing the topological lifetime vector as a modulation term into a neural ordinary differential equation, and solving the neural ordinary differential equation within the framework to generate a dynamic pathological manifold; Compressing the dynamic pathological manifold into a tensor network state, establishing a quantum optimization model for candidate intervention actions based on the spiking neural causal graph, obtaining an optimal intervention sequence by jointly optimizing an objective function, and feeding back the gradient of the joint optimization objective function with respect to the topological lifetime vector to the neural ordinary differential equation, so that the dynamic pathological manifold is updated in real time with the optimal intervention sequence and forms a public health intervention model; The optimal intervention sequence is verified according to clinical logic rules. After verification, the optimal intervention sequence is recorded in a distributed ledger, and a control instruction is generated based on the optimal intervention sequence and sent to the execution terminal.

2. The method according to claim 1, characterized in that Multimodal coding processing adopts holographic reduction representation, and performs binding operations and circular convolution operations on text data, vital sign data, medical image feature vectors and wearable sensor data respectively to generate a unified representation vector.

3. The method according to claim 1, characterized in that When extracting the topological lifetime vector, a Vitalis-Lipps complex is constructed on a uniform representation vector sequence of fixed frame length, the birth time and death time are calculated, and a lifetime value vector is formed based on the difference between the two as the topological lifetime vector.

4. The method according to claim 1, wherein When constructing a spiking neural causal graph, a spiking neural network containing voltage-discharge integral-release neurons is established, the synaptic weights are updated using the time-difference-dependent plasticity rule, and the gradient of the topological lifetime vector is used to regularize the synaptic weights.

5. The method according to claim 1, characterized in that When the topological lifetime vector is introduced as a modulation term into the neural ordinary differential equation, the external field coefficient is added to the vector field and the adaptive step-size numerical integration method is used to solve it to generate dynamic pathological manifolds.

6. The method according to claim 1, characterized in that When the compressed dynamic pathological manifold is a tensor network state, the matrix product representation is obtained through tensor rearrangement and variational singular value decomposition.

7. The method according to claim 1, characterized in that When establishing a quantum optimization model, the weights of the pulse neural causal graph are mapped to the Hamiltonian coefficients corresponding to the candidate intervention actions, and the energy expectation value and the topological risk value are linearly combined to form a joint optimization objective function.

8. The method according to claim 1, characterized in that When the optimal intervention sequence is obtained, the tensor network virtual time evolution algorithm or quantum annealing algorithm is used to solve the joint optimization objective function.

9. The method according to claim 1, characterized in that When verifying the optimal intervention sequence, the optimal intervention sequence is converted into a fact set and input into the rule base based on the answer set program, and the verified intervention sequence is obtained through reasoning.

10. The method according to claim 1, characterized in that When recording the optimal intervention sequence to the distributed ledger, a commitment value and a zero-knowledge proof are generated for the optimal intervention sequence, and then written into a chain ledger using a Byzantine fault-tolerant consensus mechanism.

Images