Diabetic foot closed-loop model construction method and system based on multi-source data

Abstract

The application discloses a diabetes foot closed loop model construction method and system based on multi-source data, and belongs to the technical field of medical information technology and intelligent health management technology. Through a sensor terminal deployed on the patient side, a clinical information integration interface and a mobile detection device, multi-source heterogeneous data is synchronously collected. On this basis, a dynamic data structure body containing a main node and multiple sub-nodes is constructed, multi-source data is mapped to the corresponding sub-nodes according to a timestamp alignment rule, and a time sequence correlation graph is established in each sub-node. The system listens to the data stream in real time, adopts an incremental learning logic with a time decay factor to update the graph, and adjusts the weight parameters of other sub-nodes through an intervention response matrix according to a cross-node influence rule. The application provides support for risk assessment and pathological evolution analysis of diabetes foot by constructing a dynamically evolving data structure and a cross-dimension correlation mechanism.

Description

Technical Field

[0001] This invention discloses a method and system for constructing a closed-loop model of diabetic foot based on multi-source data, belonging to the field of medical information technology and intelligent health management technology. Background Technology

[0002] With the rapid development of medical informatics and smart healthcare technologies, digital management of chronic complications has become a key aspect of improving the quality of clinical diagnosis and treatment. Diabetic foot, as a serious complication of diabetes, exhibits complex disease progression and high individual variability. Utilizing computer technology to assist in the assessment and management of patients' pathophysiological states has significant clinical value. In modern healthcare systems, the collection of massive amounts of physiological data through various sensor devices, information systems, and detection terminals provides a solid data foundation for constructing digital diabetic foot monitoring models.

[0003] Among these, diabetic foot condition modeling based on multi-source data fusion is a core direction for achieving precision medicine. This technology aims to integrate real-time physiological signals from wearable devices, structured medical records from electronic medical records systems, and point-like detection information from portable testing instruments to construct a comprehensive data logic that can map the trend of foot lesions in patients. To reflect the evolution of the condition in real time and accurately, the system needs to perform in-depth analysis and correlation of heterogeneous data from different sources, with different sampling frequencies, and different dimensions to achieve a digital simulation of the pathophysiological process.

[0004] In existing technologies, assessment methods for diabetic foot are often limited to independent analysis of a single data source, lacking a mechanism to organically integrate continuous physiological signals, structured medical records, and point-like detection data under a unified computer data structure. Traditional data processing methods often employ simple data splicing or format conversion, failing to fully consider the essential differences in temporal granularity between different data sources and the deep correlations at the pathophysiological level. This results in data models that are difficult to reflect the dynamic interactions between various influencing factors during disease progression. Furthermore, existing modeling methods generally suffer from rigid data structures and a lack of adaptive updating capabilities. They cannot automatically trigger adjustments to internal logical levels based on newly added clinical data, and even more so, they struggle to achieve linked updates of representation weights between different pathophysiological dimensions. In addition, the lack of design for endogeneous correlations between multi-source data in the temporal and spatial dimensions makes it difficult for the system to accurately track complex nonlinear pathological evolution processes when faced with continuously flowing dynamic data. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution: The method for constructing a closed-loop model of diabetic foot based on multi-source data includes the following steps: Step 1: Simultaneously collect multi-source heterogeneous data, including dynamic pressure sequence of the sole of the foot, skin surface temperature gradient, changes in local blood oxygen saturation, peripheral nerve conduction velocity, and laboratory biochemical indicators, through sensor terminals, clinical information integration interfaces, and mobile detection devices deployed on the patient side. Step 2: Construct a dynamic data structure containing a master node and multiple child nodes, initialize the hierarchical logical topology, the master node is used to maintain the patient's global pathophysiological state baseline, and the multiple child nodes are associated with different pathophysiological dimensions respectively. Step 3: Map data items in the multi-source data stream to corresponding child nodes according to the timestamp alignment rules, extract the absolute timestamp or relative offset time in each data frame, apply a sliding window-based synchronization alignment algorithm, map data items from different sampling frequencies to the corresponding child node storage space in the dynamic data structure, and establish a time-series correlation graph in each child node with data items as graph nodes and the time interval and numerical evolution slope between adjacent data items as edge weights; Step 4: Monitor the input status of multi-source data streams in real time. When a new data item is detected, determine the target sub-node based on the data source identifier, and update the temporal correlation graph within the target sub-node using incremental learning logic that introduces a time decay factor, and recalculate the temporal feature vector. Step 5: Adjust the weight parameters of data items in other child nodes according to the preset cross-node influence rules. Based on the pathophysiological correlation logic, establish a cross-node intervention response matrix. Trigger the fine-tuning of the weights of adjacent or related child nodes by the state evolution of the target child node.

[0006] Furthermore, step 1 also includes: For high-frequency sampled plantar pressure data, a pressure sensing array with a preset sampling frequency range is used for continuous acquisition. The pressure sensing array consists of multiple sensor units distributed in the first metatarsal bone, fifth metatarsal bone, arch, and calcaneus regions of the plantar surface. For low-frequency sampled biochemical test data, point data including glycated hemoglobin, white blood cell count, procalcitonin, C-reactive protein, and blood lipid indicators are obtained from hospital information through standard medical information exchange protocols; For the irregularly generated clinical description data, natural language processing technology is used to extract pathological entities from the chief complaint, present illness history and physical examination records in the electronic medical record. The pathological entities include redness and swelling of the soles of the feet, increased skin temperature and decreased tactile sensation, and are assigned corresponding time tags according to the time of generation of the pathology report or the time of issuance of the medical order. A Butterworth low-pass filter of a predetermined order is used to remove high-frequency noise generated by environmental electromagnetic interference, wherein the filter cutoff frequency is set at a preset multiple of the effective bandwidth of the signal. A median filtering algorithm is used to perform sliding processing on the signal to eliminate signal jump points caused by sensor momentary poor contact or motion artifacts. The sliding window size of the median filtering algorithm is dynamically adjusted according to the sampling frequency.

[0007] Furthermore, step 2 also includes: The pathophysiological dimensions include foot tissue perfusion dimension, local infection load dimension, biomechanical stress dimension, and neuropathy degree dimension. The master node is configured to maintain a global state vector obtained by weighted fusion calculation of the feature values ​​fed back by each child node. The global state vector is used to characterize the overall deterioration risk score of diabetic foot. Each child node has a local data buffer for temporarily storing the original sampled data. The global state vector maintained by the master node is defined as a multidimensional tensor. The value of this tensor is normalized and weighted by transforming the local feature vectors output by each child node through the feature mapping matrix, and is used to reflect the patient's overall foot health status within the current time window. Each child node's internal local data buffer uses a circular queue structure to temporarily store the original values ​​within a time series of a preset length; The feature mapping matrix inside the child node serves as a transformation operator, used to transform the original numerical data into quantitative features that reflect the trend of pathological evolution. The quantitative features include the coefficient of variation or the slope of numerical decline of physiological indicators within a specific time period. The dynamic data structure is organized in memory in the form of a balanced tree or a directed acyclic graph to support the retrieval of data for each physiological dimension.

[0008] Furthermore, step 3 also includes: The sliding window-based synchronization alignment algorithm uses a variable window with a preset time span. When a drastic change in the patient's movement state is detected, the window step value is automatically adjusted. To address data loss caused by inconsistent sampling frequencies, missing sampling points are identified and filled using a third-order spline interpolation algorithm within the same logical time slice. A piecewise cubic polynomial is constructed to ensure continuous derivatives at connection points. When establishing a temporal correlation graph within each child node, the turning point data of disease evolution are identified by calculating the centrality index of the graph. The centrality index includes degree centrality or eigenvector centrality. When the centrality index of a node changes abruptly and the weight of its adjacent edges shows nonlinear fluctuations, the node is marked as a pathological risk trigger point.

[0009] Furthermore, step 4 also includes: Set a data disturbance threshold. When the deviation of the value of the newly added data item from the data item of the same dimension at the previous time exceeds the first preset threshold, the incremental update program is activated immediately. The update program monitors the status of each child node in real time through a background listening process, and quickly locates the target child node after identifying the data source identifier. When using incremental learning logic, the influence of historical data is exponentially reduced through a time decay factor, so that the model prioritizes the retention of the feature contributions of recent physiological fluctuations. During the update process, the node increment and edge weight reconstruction of the graph are completed first in the memory buffer. After the calculation is completed, the updated topology is synchronized to the persistent storage layer to ensure real-time response to multi-source data streams.

[0010] Furthermore, step 5 also includes: When the blood oxygen index in the foot tissue perfusion dimension sub-node decreases beyond the second preset threshold, the intervention response matrix triggers a logical judgment to automatically increase the representation weight of the inflammation index in the global state vector in the local infection load dimension sub-node. The weight adjustment range is determined by a predetermined adjustment ratio. The adjustment process of the weight parameters is associated with an adaptive feedback loop, which evaluates in real time the accuracy of the global state vector after weight adjustment for clinical risk prediction. If the residual between the predicted risk value and the actual clinical observation value is greater than the preset error threshold, the coupling coefficient in the intervention response matrix is ​​corrected through the backpropagation mechanism to achieve dynamic self-optimization of the internal logical relationship of the model.

[0011] Furthermore, the method also involves a deep learning feature extraction process: Based on the constructed dynamic data structure, a deep convolutional neural network is applied to extract multi-dimensional spatial features, and high-pressure plaque regions in plantar pressure distribution images are identified through alternating convolutional and pooling layers. By combining long short-term memory network analysis with the temporal evolution of the output of each sub-node, the long-term trends and periodic fluctuations in the foot temperature change sequence or blood oxygen change sequence are captured by the forgetting gate and memory gate mechanisms. By fusing multi-scale features of spatial and temporal characteristics, the risk probability of foot ulcers occurring or the predicted healing progress of existing wounds within a preset prediction time period can be calculated.

[0012] Furthermore, the method interacts with external intervention devices through a closed-loop control module: When the risk score output by the master node exceeds the preset safety threshold, and the foot pressure is continuously concentrated in a specific area and the local skin temperature rises above the preset temperature threshold, the inflation pressure adjustment command of the decompression brace is automatically triggered. This reduces the load on the pressure area by changing the support position, or pushes an intervention command to the medical staff. The dynamic data structure is persistently stored using a distributed graph database, and the data is sharded and stored on different physical server nodes to support high-concurrency relational queries for a large number of patient groups. The data quality verification process is executed. If the data loss duration of a certain sub-node exceeds the preset duration threshold, the historical average of similar patients or the inferred value of related sub-nodes are automatically called to offset the data, and the data credibility mark is recorded.

[0013] Furthermore, the method employs an arbitration mechanism when handling concurrent conflicts between multiple sensors: The master node initiates the arbitration module based on evidence theory, treating the feature outputs of each child node as independent sources of evidence. Each sub-node outputs a trust rating index based on the current data quality label, and the master node integrates contradictory information from different dimensions using the evidence synthesis rule. If the data stability of a certain child node is lower than the preset stability standard, the contribution weight of that data source to the global risk score will be automatically reduced, and it will be marked as a suspected abnormal interference in memory, while triggering a local self-check program. By performing task orchestration through a stream processing framework, data acquisition, mapping, updating, and weight adjustment are abstracted into stateful stream processing operators. A watermark mechanism is used to solve the problem of out-of-order arrival of data items, ensuring a closed-loop response to new data is completed within seconds.

[0014] According to a second aspect of the present invention, the present invention claims protection for a system for constructing a closed-loop model of diabetic foot based on multi-source data, comprising: One or more processors; A memory storing one or more programs, which, when executed by one or more processors, enable the processors to implement the method for constructing a closed-loop model of diabetic foot based on multi-source data.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. A highly integrated multi-source heterogeneous data fusion mechanism This invention completely solves the problem of independent and difficult-to-coordinate data from different sources in existing technologies by constructing a dynamic data structure containing a master node and multiple child nodes. By mapping continuous physiological signals, structured diagnostic records, and point-like detection data into a unified computer logic body, deep organic integration of data at the pathophysiological level is achieved. Compared to traditional simple splicing methods, this method can reflect the complex interaction relationships between various influencing factors of diabetic foot from a system-wide perspective, significantly improving the accuracy of data association.

[0016] 2. Excellent dynamic adaptation and real-time update capabilities The dynamic update mechanism and incremental learning logic designed in this invention transform the constructed model from a static accumulation of historical records into a digital organism that continuously evolves with the progression of the disease. Triggered updates in response to new data ensure a high degree of synchronization between the model's internal logical hierarchy and real-time clinical status. In particular, the introduction of cross-node influence rules enables the coordinated adjustment of weights across different pathophysiological dimensions, significantly enhancing the model's tracking accuracy and pathological simulation capabilities when faced with dynamic and continuously flowing data, resulting in a substantial reduction in prediction bias compared to traditional static models.

[0017] 3. Precise timeline correlation and in-depth pattern mining By constructing timestamp alignment rules and temporal correlation graphs, this invention solves the technical challenge of inconsistent temporal granularity in multi-source data at the underlying data structure level. The system can automatically reconstruct the chronological order and evolutionary chain of pathological events without the need for tedious manual timeline comparisons. This endogenous correlation design enables the automated and intelligent extraction of nonlinear correlation patterns hidden among massive multi-source data, providing clinicians with more intuitive and in-depth decision support information and significantly improving the efficiency of assisted assessment.

[0018] 4. Robust system architecture and clinical applicability The system employs a distributed storage and closed-loop control module design, ensuring high performance and reliability in large-scale application scenarios. This invention not only enables accurate early warning of diabetic foot risks but also allows for proactive intervention in the treatment process through closed-loop feedback, achieving a complete management loop from monitoring to assessment to intervention. Through automatic data quality verification and hedging, the model maintains good assessment stability even in cases of missing or abnormal data, significantly advancing the digital management of diabetic foot. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the process of constructing a closed-loop model of diabetic foot based on multi-source data, as claimed in this embodiment of the invention. Figure 2 This is a second flowchart of the method for constructing a closed-loop model of diabetic foot based on multi-source data, as claimed in the embodiments of the present invention. Figure 3 The third flowchart is a method for constructing a closed-loop model of diabetic foot based on multi-source data, as claimed in the embodiments of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of those features. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications in the embodiments of this application, such as up, down, left, right, front, back, etc., are only used to explain the relative positional relationships and movements between components in a specific orientation as shown in the accompanying drawings. If the specific orientation changes, the directional indications will change accordingly. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0022] References to embodiments herein mean that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] According to a first embodiment of the present invention, the present invention requests protection for a method for constructing a closed-loop model of diabetic foot based on multi-source data, referring to... Figure 1 This includes the following steps: Step 1: Simultaneously collect multi-source heterogeneous data, including dynamic pressure sequence of the sole of the foot, skin surface temperature gradient, changes in local blood oxygen saturation, peripheral nerve conduction velocity, and laboratory biochemical indicators, through sensor terminals, clinical information integration interfaces, and mobile detection devices deployed on the patient side. Step 2: Construct a dynamic data structure containing a master node and multiple child nodes, initialize the hierarchical logical topology, the master node is used to maintain the patient's global pathophysiological state baseline, and the multiple child nodes are associated with different pathophysiological dimensions respectively. Step 3: Map data items in the multi-source data stream to corresponding child nodes according to the timestamp alignment rules, extract the absolute timestamp or relative offset time in each data frame, apply a sliding window-based synchronization alignment algorithm, map data items from different sampling frequencies to the corresponding child node storage space in the dynamic data structure, and establish a time-series correlation graph in each child node with data items as graph nodes and the time interval and numerical evolution slope between adjacent data items as edge weights; Step 4: Monitor the input status of multi-source data streams in real time. When a new data item is detected, determine the target sub-node based on the data source identifier, and update the temporal correlation graph within the target sub-node using incremental learning logic that introduces a time decay factor, and recalculate the temporal feature vector. Step 5: Adjust the weight parameters of data items in other child nodes according to the preset cross-node influence rules. Based on the pathophysiological correlation logic, establish a cross-node intervention response matrix. Trigger the fine-tuning of the weights of adjacent or related child nodes by the state evolution of the target child node.

[0024] In this embodiment, the specific implementation steps of the method for constructing a closed-loop model of diabetic foot based on multi-source data are as follows: Step 1, Synchronous Acquisition of Multi-Source Heterogeneous Data: On the patient side, an integrated data acquisition front-end is first deployed. This front-end includes sensor terminals embedded in patients' everyday wearable devices such as smart socks and insoles, which contain multiple miniature pressure sensors, temperature sensors, and near-infrared spectroscopy sensors. Simultaneously, a secure connection is established with the hospital's electronic medical records and laboratory information via a clinical information integration interface, using the HL7 or FHIR standard protocol. Mobile detection devices, such as portable Doppler ultrasound machines or handheld biochemical analyzers used by healthcare personnel, collect data at the bedside or in the outpatient clinic and upload it in real time via a wireless network.

[0025] The specific data types collected include: Dynamic pressure sequence of the foot: When the patient walks or stands, the pressure sensor array in the insole continuously records the pressure values ​​of various areas of the foot at a frequency of 50 to 200 times per second.

[0026] Skin surface temperature gradient: Multiple temperature sensor nodes on smart socks or foot patches simultaneously collect the temperature of different areas such as the back of the foot, sole, and between the toes, forming temperature gradient data.

[0027] Local oxygen saturation changes: By wearing a reflective oxygen probe on the ankle or dorsum of the foot, the tissue oxygenation is continuously monitored, and a continuous change curve of oxygen saturation is obtained.

[0028] Peripheral nerve conduction velocity: The conduction velocity and amplitude are recorded by electrically stimulating the patient's sural nerve or tibial nerve at specific time points using a nerve conduction detector connected to a mobile detection device.

[0029] Laboratory biochemical indicators: Through an integrated interface, data such as glycated hemoglobin, C-reactive protein, procalcitonin, white blood cell count, and complete blood lipid profile uploaded by the laboratory during the patient's most recent outpatient or inpatient visit are automatically captured.

[0030] Step 2, construct the dynamic data structure: A core data structure is created in memory, consisting of a master node and four child nodes. These four child nodes correspond to four pathophysiological dimensions: foot tissue perfusion, local infection burden, biomechanical stress, and neuropathy severity. The master node does not directly store raw data; instead, it maintains a global state vector containing static baseline information such as patient age, diabetes duration, and past foot disease history. During initialization, each child node allocates an independent dynamic memory space as a local data buffer and establishes a feature mapping table to convert subsequent input raw data into pathological features for that dimension.

[0031] Step 3, Data Mapping and Temporal Relationship Graph Construction: When multiple data streams continuously flow in, each data point is first aligned within a time-based sliding window based on its absolute timestamp and data type identifier. For example, a peak pressure data point collected at time T and a blood oxygen saturation drop data point collected at T+0.1 seconds are considered related data within the same time window. After alignment, the data is routed to the corresponding child node based on its data type. For instance, all pressure data is stored in the biomechanical stress child node, and all blood oxygen saturation data is stored in the foot tissue perfusion child node.

[0032] Within each child node, each aligned data item is defined as a graph node, with node attributes including numerical value, collection time, and data source. Then, adjacent data items are connected by directed edges in chronological order. The weight of this edge is determined by two factors: first, the shorter the time interval between the two data items, the higher the weight, indicating a stronger correlation; second, the more drastic the change in the slope of the numerical evolution, the higher the weight, indicating a faster change in physiological state. Thus, a dynamic temporal correlation graph reflecting the evolution of that dimension's physiological parameter over time is formed within each child node.

[0033] Step 4, Incremental learning and updating: A background listening thread continuously monitors the message queues of all data input ports in real time. Once a new data packet is detected, its data source identifier is immediately parsed. For example, if the new data is a new pressure spike from a pressure insole, the biomechanical stress sub-node is quickly located.

[0034] Subsequently, an incremental learning process is initiated. This process does not recalculate the entire graph, but only performs local updates based on new data. It first inserts the new data as a new graph node at the end of the graph and calculates the weights of its edges connecting to the previous node. During the calculation, a time decay factor is applied, meaning that the influence of older nodes and edges in the graph decreases exponentially, for example, by 5% per day. This means that when recalculating the temporal feature vectors describing the current state of the child node, such as recent stress peak trends and fluctuation frequencies, newer data contributes more, thus enabling the model to quickly adapt to the patient's recent condition.

[0035] Step 5, cross-node weight adjustment: It incorporates a set of cross-node influence rules based on pathophysiological knowledge. When the state evolution of any sub-node, such as the biomechanical stress sub-node, in step 4 is triggered and its internal map is updated, a preset intervention response matrix is ​​invoked. This matrix defines the direction and intensity of the mutual influence between different sub-nodes.

[0036] For example, the intervention response matrix records a rule: abnormal biomechanical stress, such as persistently high heel pressure, increases the risk of foot tissue perfusion and may lead to a decrease in local blood oxygenation. Based on this rule, the system actively searches for foot tissue perfusion sub-nodes associated with the biomechanical stress sub-nodes and fine-tunes the weighting parameters of certain data items within them, such as the blood oxygen saturation data corresponding to the heel region. This fine-tuning means that during subsequent global risk assessments by the main node, the blood oxygen data of the affected heel region will receive greater attention, and its numerical changes will have a greater impact on the final risk score.

[0037] Furthermore, step 1 also includes: For high-frequency sampled plantar pressure data, a pressure sensing array with a preset sampling frequency range is used for continuous acquisition. The pressure sensing array consists of multiple sensor units distributed in the first metatarsal bone, fifth metatarsal bone, arch, and calcaneus regions of the plantar surface. For low-frequency sampled biochemical test data, point data including glycated hemoglobin, white blood cell count, procalcitonin, C-reactive protein, and blood lipid indicators are obtained from hospital information through standard medical information exchange protocols; For the irregularly generated clinical description data, natural language processing technology is used to extract pathological entities from the chief complaint, present illness history and physical examination records in the electronic medical record. The pathological entities include redness and swelling of the soles of the feet, increased skin temperature and decreased tactile sensation, and are assigned corresponding time tags according to the time of generation of the pathology report or the time of issuance of the medical order. A Butterworth low-pass filter of a predetermined order is used to remove high-frequency noise generated by environmental electromagnetic interference, wherein the filter cutoff frequency is set at a preset multiple of the effective bandwidth of the signal. A median filtering algorithm is used to perform sliding processing on the signal to eliminate signal jump points caused by sensor momentary poor contact or motion artifacts. The sliding window size of the median filtering algorithm is dynamically adjusted according to the sampling frequency.

[0038] In this embodiment, an array of 16 to 64 micro-film pressure sensor units is integrated into the smart insole worn by the patient. These sensors are not uniformly distributed, but rather precisely aligned with key anatomical weight-bearing areas of the foot: at least 2-4 sensor units are deployed below the first metatarsal head of the first phalanx, below the fifth metatarsal head of the fifth phalanx, in the arch and navicular region, and in the central region of the calcaneus and heel. Each sensor unit is continuously scanned at a sampling frequency of 100Hz to 250Hz to create a dynamic image of the plantar pressure distribution.

[0039] A separate interface service periodically initiates queries to the hospital's Medical Information System (LIS), for example, every 6 hours. The queries are limited to specific test results related to diabetic foot generated within the last 72 hours, including glycated hemoglobin (HbA1c) reflecting long-term blood glucose control, white blood cell count and neutrophil percentage reflecting systemic infection status, procalcitonin and C-reactive protein reflecting the severity of bacterial infection and inflammation, and lipid indicators such as total cholesterol, triglycerides, and low-density lipoprotein (LDL). The acquired data is in point data format, with each data point accompanied by the precise sampling and reporting times.

[0040] For unstructured text entered by doctors into electronic medical records, such as chief complaint of redness and swelling of the right third toe, elevated skin temperature for 3 days, present medical history of increased pain when walking, decreased tactile sensation upon physical examination, and a positive 10g nylon filament test, a natural language processing engine based on BERT fine-tuning is invoked. This engine first performs word segmentation and entity recognition on the text, extracting pathological entities such as redness and swelling, elevated skin temperature, and decreased tactile sensation. Then, it assigns a timestamp to each extracted entity based on the document type (admission record, progress note, etc.) and the document's creation time. For example, if the entity "elevated skin temperature" appears in a progress note created on October 27, 2023 at 08:30, then the timestamp for that data point is October 27, 2023 at 08:30.

[0041] All continuous signals collected from sensor terminals, such as pressure and temperature, pass through a digital signal preprocessing unit before entering the data fusion module. This unit uses a fifth-order Butterworth low-pass filter with a cutoff frequency set to 1.2 times the effective bandwidth of the signal. For example, if the effective frequency components of the plantar pressure signal are mainly concentrated below 10Hz, the filter cutoff frequency is set to 12Hz to effectively filter out 50Hz or 60Hz power frequency interference and harmonics caused by radiation from surrounding electronic devices such as mobile phones and Wi-Fi routers.

[0042] For the signal after low-pass filtering, a median filtering algorithm is further applied. This algorithm sets a dynamic sliding window, the length of which is inversely proportional to the signal's sampling frequency. For example, for a 250Hz pressure signal, the window size is set to 9 sampling points; for a 50Hz temperature signal, the window size is set to 5 sampling points. The algorithm sorts all values ​​within the window and takes the median value as the output value for the current sampling point. This method can effectively eliminate instantaneous spike signals caused by poor sensor contact due to sudden foot tremors in patients, or random drift points caused by sensor aging, making the signal curve smoother and more accurate.

[0043] Furthermore, step 2 also includes: The pathophysiological dimensions include foot tissue perfusion dimension, local infection load dimension, biomechanical stress dimension, and neuropathy degree dimension. The master node is configured to maintain a global state vector obtained by weighted fusion calculation of the feature values ​​fed back by each child node. The global state vector is used to characterize the overall deterioration risk score of diabetic foot. Each child node has a local data buffer for temporarily storing the original sampled data. The global state vector maintained by the master node is defined as a multidimensional tensor. The value of this tensor is normalized and weighted by transforming the local feature vectors output by each child node through the feature mapping matrix, and is used to reflect the patient's overall foot health status within the current time window. Each child node's internal local data buffer uses a circular queue structure to temporarily store the original values ​​within a time series of a preset length; The feature mapping matrix inside the child node serves as a transformation operator, used to transform the original numerical data into quantitative features that reflect the trend of pathological evolution. The quantitative features include the coefficient of variation or the slope of numerical decline of physiological indicators within a specific time period. The dynamic data structure is organized in memory in the form of a balanced tree or a directed acyclic graph to support the retrieval of data for each physiological dimension.

[0044] In this embodiment, during initialization, the four child nodes created are responsible for data management in the following pathophysiological dimensions: The foot tissue perfusion sub-node is responsible for receiving and processing data such as transcutaneous oxygen partial pressure, ankle-brachial index, and local blood oxygen saturation.

[0045] The local infection burden sub-node is responsible for receiving and processing wound secretion culture results, inflammatory markers CRP, PCT, and local skin temperature data.

[0046] The biomechanical stress sub-node is responsible for receiving and processing data such as dynamic pressure distribution of the foot, gait cycle, and foot deformity angle.

[0047] The neuropathy severity sub-node is responsible for receiving and processing data such as nerve conduction velocity, vibration sensory threshold, and 10g nylon filament test results.

[0048] The master node maintains a multidimensional global state vector, which is a four-dimensional tensor. The value of each dimension does not come directly from the original data of the child nodes, but is obtained by weighted summation of the local feature vectors periodically output by the child nodes after transformation by a pre-trained feature mapping matrix. The final value of this vector directly maps to a diabetic foot overall deterioration risk score between 0 and 100.

[0049] Each child node contains a local data buffer, implemented as a fixed-length circular queue. The queue length is set to store the raw values ​​from the past 7 days for high-frequency data or the most recent 10 for low-frequency data. For example, for the biomechanical stress child node, its buffer might hold the complete stress distribution matrix for every gait test within the last 7 days. When new data arrives, it overwrites the oldest data in the queue, ensuring that the buffer always stores the most up-to-date time-series information.

[0050] Each child node contains a dedicated feature mapping matrix. This matrix is ​​essentially a complex set of operators used to transform the raw time-series data within the buffer into clinically meaningful quantitative features. For example, for the neuropathy severity child node, its feature mapping matrix calculates the ratio of the standard deviation of the coefficient of variation to the mean of the three most recent nerve conduction velocity measurements, and fits the slope of the trend line between the two most recent measurements using the least squares method, thereby quantifying the speed and stability of neurological function deterioration.

[0051] To achieve fast retrieval and dynamic updates, the entire dynamic data structure, including the master node and all child nodes, is organized in physical memory as a red-black tree. The master node serves as the root of the tree, and the four child nodes are the direct children of the root. Under each child node, the internal temporal relationship graph is organized as a directed acyclic graph. The balanced nature of the red-black tree ensures that the time complexity of locating any child node remains at O(logn) regardless of the data volume, thus guaranteeing retrieval efficiency under high-concurrency data streams. Simultaneously, the directed acyclic graph structure avoids circular dependencies within the graph, making the direction of feature calculation and risk propagation clear and traceable.

[0052] Furthermore, referring to Figure 2 Step 3 further includes: The sliding window-based synchronization alignment algorithm uses a variable window with a preset time span. When a drastic change in the patient's movement state is detected, the window step value is automatically adjusted. To address data loss caused by inconsistent sampling frequencies, missing sampling points are identified and filled using a third-order spline interpolation algorithm within the same logical time slice. A piecewise cubic polynomial is constructed to ensure continuous derivatives at connection points. When establishing a temporal correlation graph within each child node, the turning point data of disease evolution are identified by calculating the centrality index of the graph. The centrality index includes degree centrality or eigenvector centrality. When the centrality index of a node changes abruptly and the weight of its adjacent edges shows nonlinear fluctuations, the node is marked as a pathological risk trigger point.

[0053] In this embodiment, a data aligner based on a variable time window is implemented. The aligner defaults to a sliding window of 3 seconds with a 1-second step. However, the aligner continuously monitors the patient's motion status from accelerometer readings or pressure center trajectories. When it detects a change in the patient's state from rest to walking, or a significant change in walking speed (e.g., a change in cadence exceeding 20%), the aligner automatically shortens the window length to 1 second and the step size to 0.5 seconds. This dynamic adjustment ensures that during critical periods of drastic changes in motion status, data from different sampling frequencies, such as high-frequency pressure and low-frequency blood oxygen, can be more closely aligned to the same logical time slice, avoiding misjudgments due to data misalignment.

[0054] During data alignment, if a certain type of data is missing within a time slice—for example, a 1-second window may contain pressure data but no blood oxygen data—an interpolation completion module is activated. This module identifies the two nearest valid data points before and after the missing data point and then constructs a third-order spline interpolation function between these two points. This function is a smooth curve passing through these two endpoints and ensuring the continuity of its first and second derivatives. Based on the timestamp of the missing point, the completed value is calculated from this curve. This method preserves the fluctuation characteristics of physiological signals better than simple linear interpolation.

[0055] Once the data is organized into a temporal correlation graph within the child nodes, a graph theory analysis is performed periodically, for example, every 15 minutes. One core task is to calculate the eigenvector centrality of each node in the graph. This metric considers not only the number of neighbor centralities a node has, but also the importance of its neighbors.

[0056] A monitoring logic is established: when the calculation detects that the eigenvector centrality of a node suddenly increases by more than a preset threshold (e.g., 50%) within two consecutive update cycles, and the weights of the edges connecting this node to its immediate and adjacent nodes all show a rate of change exceeding the normal fluctuation range (e.g., pressure value increases by 30% within 0.1 seconds, or blood oxygen value decreases by 15% within 1 second), this node will be marked as a pathological risk trigger point. This marking signifies that at this point in time, the patient's physiological state has experienced a non-linear, critical deterioration inflection point, and subsequent modeling will conduct a more in-depth analysis of this node and its surrounding area.

[0057] Furthermore, referring to Figure 3 Step 4 further includes: Set a data disturbance threshold. When the deviation of the value of the newly added data item from the data item of the same dimension at the previous time exceeds the first preset threshold, the incremental update program is activated immediately. The update program monitors the status of each child node in real time through a background listening process, and quickly locates the target child node after identifying the data source identifier. When using incremental learning logic, the influence of historical data is exponentially reduced through a time decay factor, so that the model prioritizes the retention of the feature contributions of recent physiological fluctuations. During the update process, the node increment and edge weight reconstruction of the graph are completed first in the memory buffer. After the calculation is completed, the updated topology is synchronized to the persistent storage layer to ensure real-time response to multi-source data streams.

[0058] In this embodiment, a dynamic data perturbation threshold is set for the key features of each child node. This threshold is not a fixed value, but is dynamically calculated based on the standard deviation of the node's recent data, such as the data from the past 24 hours, and is typically set to three times the standard deviation.

[0059] A separate, high-priority listening process runs in the background, directly bound to the event-driven I / O model of the operation. When a new data item is written to the data acquisition queue, the listening process immediately captures the event and parses its data source identifier. Subsequently, the process quickly reads the most recently updated data value of the target child node and calculates the deviation between the new and old values. For example, if the newly received C-reactive protein value of the local infection load child node is 75 mg / L, while it was 25 mg / L at the previous moment, the deviation reaches 200%, exceeding the preset dynamic threshold, then the listening process immediately activates the incremental update procedure for that child node.

[0060] The activated incremental update procedure does not traverse the entire historical data of the child node. It first locks the local data buffer of the child node and appends new data items to the head of the circular queue. Then, the procedure begins updating the temporal correlation graph. It only creates new graph nodes and edges connecting new nodes to their predecessors. When calculating the weights of new edges, the procedure introduces an exponential time decay factor. This factor acts on all existing nodes and edges in the graph, resulting in the contribution weight of each historical data point decaying exponentially with its existence time when calculating the current temporal feature vector. Specifically, if a data point has existed for t days, its weight in the current feature calculation is e^(-λt), where λ is the decay rate. This achieves the model's ability to forget historical data and focus on recent data.

[0061] To ensure real-time responsiveness to the data stream, all operations of the incremental update procedure—including the creation of new nodes, the calculation of edge weights, and the updating of temporal feature vectors—are preferentially completed in the memory buffer. Only after all these calculations are complete and the new temporal feature vectors have been successfully written to the global state vector of the master node, will a low-priority asynchronous thread be started in the background. This thread serializes the updated child node topology, i.e., the updated graph, and writes it to persistent storage such as an SSD. This design ensures that the data acquisition and model update process is not blocked by disk I / O operations, and can always respond to new data input with a latency of microseconds.

[0062] Furthermore, step 5 also includes: When the blood oxygen index in the foot tissue perfusion dimension sub-node decreases beyond the second preset threshold, the intervention response matrix triggers a logical judgment to automatically increase the representation weight of the inflammation index in the global state vector in the local infection load dimension sub-node. The weight adjustment range is determined by a predetermined adjustment ratio. The adjustment process of the weight parameters is associated with an adaptive feedback loop, which evaluates in real time the accuracy of the global state vector after weight adjustment for clinical risk prediction. If the residual between the predicted risk value and the actual clinical observation value is greater than the preset error threshold, the coupling coefficient in the intervention response matrix is ​​corrected through the backpropagation mechanism to achieve dynamic self-optimization of the internal logical relationship of the model.

[0063] In this embodiment, the preset intervention response matrix is ​​a set of rules describing the interaction between sub-nodes. One of the rules is: when the blood oxygen saturation index corresponding to the heel region in the foot tissue perfusion sub-node decreases by more than 10 percentage points within 30 consecutive minutes, for example, from 95% to below 85%, it is considered to have exceeded the second preset threshold.

[0064] At this point, the intervention response matrix automatically triggers a logical decision process. This process generates a weight adjustment instruction: increase the weight of inflammation-related indicators, such as C-reactive protein and local skin temperature, in the global state vector calculation within the local infection burden sub-node by 30%. This adjustment percentage is predetermined but will be dynamically corrected based on subsequent feedback. For example, if the original weight of inflammation indicators in the global risk score was 20%, it will be increased to 26% after the adjustment.

[0065] The aforementioned weight adjustments are not a one-time event. A closed-loop adaptive feedback loop has been established. After the weight adjustments take effect, the changes in the global risk score output by the master node will be continuously tracked over a period of time, such as within 48 hours, and compared with the actual clinical observation results. The actual clinical observation results may include wound conditions subsequently entered by the doctor, such as whether the wound has worsened, new imaging examination results, or laboratory re-examination results.

[0066] The residual between the predicted risk value and the actual clinical observation value is calculated. For example, if the model predicts an extremely high risk, but the actual clinical observation shows that the patient's condition is stable, the residual will be large. When the absolute value of this residual is greater than a preset error threshold, such as a score deviation exceeding 15 points, a backpropagation correction procedure is initiated. This procedure uses an optimization algorithm to fine-tune the coupling coefficient in the intervention response matrix that leads to this weight adjustment, i.e., the correspondence between a 10% decrease and a 30% increase in weight. After multiple iterations, the coupling coefficient gradually converges, allowing the internal logical relationship of the model to more accurately reflect the true pathophysiological evolution, thus achieving model self-optimization.

[0067] Furthermore, the method also involves a deep learning feature extraction process: Based on the constructed dynamic data structure, a deep convolutional neural network is applied to extract multi-dimensional spatial features, and high-pressure plaque regions in plantar pressure distribution images are identified through alternating convolutional and pooling layers. By combining long short-term memory network analysis with the temporal evolution of the output of each sub-node, the long-term trends and periodic fluctuations in the foot temperature change sequence or blood oxygen change sequence are captured by the forgetting gate and memory gate mechanisms. By fusing multi-scale features of spatial and temporal characteristics, the risk probability of foot ulcers occurring or the predicted healing progress of existing wounds within a preset prediction time period can be calculated.

[0068] In this embodiment, a dynamic plantar pressure sequence is obtained from the biomechanical stress sub-node. This sequence is reorganized into a series of temporally continuous 2D pressure distribution maps, each representing a plantar pressure cloud map at a specific time point. These 2D maps are then fed into a pre-trained deep convolutional neural network, such as a variant of ResNet-50. The network uses stacked convolutional layers to extract local pressure features, such as the shape and size of high-pressure plaques, and pooling layers to reduce dimensionality and extract key features, abstracting layer by layer, ultimately outputting a multi-dimensional spatial feature vector at the network's end. This vector quantitatively describes key risk areas in the plantar pressure distribution, such as persistent high-pressure plaques below the first metatarsal head.

[0069] Simultaneously, the time-series data output from each sub-node, such as the blood oxygen saturation change curve from the foot tissue perfusion sub-node and the skin temperature change curve from the local infection burden sub-node, are concatenated to form a multi-dimensional time series. This series is then fed into a Long Short-Term Memory (LSTM) network. The forgetting and remembering mechanisms within the LSTM network automatically learn and remember long-term trends in the sequence, such as the slow decline in blood oxygen over the past two weeks, and periodic fluctuations, such as the normal fluctuations in skin temperature every morning and evening, while ignoring occasional noise spikes. The output of the last time step of the LSTM network is a temporal feature vector that condenses the evolution patterns of all the input time series.

[0070] The spatial feature vector extracted by CNN is concatenated with the temporal feature vector extracted by LSTM to form a new, more informative multi-scale fused feature vector. This fused vector is then fed into a regressor or classifier consisting of fully connected layers. After training, the network can directly output a numerical value based on the input fused features. This value represents the probability (0-100%) of a new foot ulcer developing in the patient within the next 7 or 14 days, or, for patients with existing wounds, a predicted value for wound healing progress, such as the estimated healing time or healing percentage.

[0071] Furthermore, the method interacts with external intervention devices through a closed-loop control module: When the risk score output by the master node exceeds the preset safety threshold, and the foot pressure is continuously concentrated in a specific area and the local skin temperature rises above the preset temperature threshold, the inflation pressure adjustment command of the decompression brace is automatically triggered. This reduces the load on the pressure area by changing the support position, or pushes an intervention command to the medical staff. The dynamic data structure is persistently stored using a distributed graph database, and the data is sharded and stored on different physical server nodes to support high-concurrency relational queries for a large number of patient groups. The data quality verification process is executed. If the data loss duration of a certain sub-node exceeds the preset duration threshold, the historical average of similar patients or the inferred value of related sub-nodes are automatically called to offset the data, and the data credibility mark is recorded.

[0072] In this embodiment, the master node continuously calculates the global risk score. Simultaneously, an independent closed-loop control module monitors key data from the biomechanical stress and local infection load sub-nodes in real time. When the risk score output by the master node exceeds a preset red alert threshold, such as 80 points, and simultaneously the biomechanical stress sub-node shows that plantar pressure remains concentrated in the second metatarsal head region for more than 10 minutes, and the local infection load sub-node shows that the corresponding local skin temperature in this area rises by more than 2.0°C above the corresponding area on the healthy side, the closed-loop control module automatically generates a control command.

[0073] The command is sent via Bluetooth or WiFi to the patient's smart pressure-reducing brace, such as an adjustable air-cushioned shoe. The command instructs the patient to increase the inflation pressure of the air bladder beneath the second metatarsal head to 30 kPa, thereby altering the brace's support point and redistributing pressure to the surrounding area. Simultaneously, a structured alert is generated and pushed to the attending physician's mobile terminal via internal hospital messaging. The alert includes the message: "Patient XXX, high pressure combined with elevated skin temperature in the right second metatarsal head area; pressure-reducing brace automatically adjusted; please pay attention."

[0074] The dynamic data structure for all patients is not stored in a single database during persistent storage, but rather in a distributed graph database, such as a JanusGraph or Neo4j cluster. Data sharding is employed: data from different patients is stored on different physical server nodes; for the same patient, the data from its master node and all child nodes is further sharded and stored on different nodes within the same cluster. This sharding strategy, combined with the graph database's native relational storage capabilities, enables simultaneous online monitoring of thousands or even tens of thousands of patients and allows healthcare professionals to concurrently perform complex cross-patient, cross-dimensional relational queries, such as querying all patients with abnormal heel pressure and elevated C-reactive protein levels.

[0075] A built-in data quality monitor is implemented, setting a timer for the duration of data loss for each data stream. For example, if the plantar pressure data stream does not receive any new data packets for more than 10 minutes, the monitor determines that the data stream of that sub-node is out of service. In this case, instead of simply using a null value, a data hedging module is activated. This module first retrieves the historical average pressure data for the same patient group in the same age group, with the same diabetes duration, and the same foot type (e.g., high arches, flat feet) within the same time period from the database. Simultaneously, the module also calls the current data of the neuropathy degree and foot tissue perfusion sub-nodes, and estimates a possible pressure value under the current neurological and perfusion conditions through a correlation extrapolation model trained on historical data. The values ​​from these two sources are weighted and averaged, and used as temporary substitute data to populate the biomechanical stress sub-node. A low-confidence metadata tag is attached to this data item for reference in subsequent analysis and clinical decision-making.

[0076] Furthermore, the method employs an arbitration mechanism when handling concurrent conflicts between multiple sensors: The master node initiates the arbitration module based on evidence theory, treating the feature outputs of each child node as independent sources of evidence. Each sub-node outputs a trust rating index based on the current data quality label, and the master node integrates contradictory information from different dimensions using the evidence synthesis rule. If the data stability of a certain child node is lower than the preset stability standard, the contribution weight of that data source to the global risk score will be automatically reduced, and it will be marked as a suspected abnormal interference in memory, while triggering a local self-check program. By performing task orchestration through a stream processing framework, data acquisition, mapping, updating, and weight adjustment are abstracted into stateful stream processing operators. A watermark mechanism is used to solve the problem of out-of-order arrival of data items, ensuring a closed-loop response to new data is completed within seconds.

[0077] In this embodiment, when the master node is fusing information from various sub-nodes, if it discovers logical contradictions in the feature outputs from different dimensions—for example, the local infection load sub-node judges the infection to be severe based on skin temperature data, while peripheral blood count data shows a normal white blood cell count—the master node will initiate an arbitration module. This arbitration module operates based on the DS evidence theory, treating the feature outputs of the four sub-nodes—biomechanical stress, local infection load, foot tissue perfusion, and neuropathy severity—as four independent sources of evidence.

[0078] Each child node, while outputting its feature value, also outputs a trust score based on the completeness, continuity, and consistency with historical data of its internal data. For example, a trust score of 0.9 indicates good data quality, while a trust score of 0.6 indicates missing data. The arbitration module uses evidence synthesis rules to fuse and calculate these potentially conflicting pieces of evidence with varying trust scores, synthesizing a comprehensive basic probability allocation with minimal conflict. Ultimately, the master node determines the global risk score based on this synthesized probability allocation, rather than simply trusting any single source of evidence.

[0079] During the arbitration process described above, if the arbitration module finds that the data stability of a certain child node is lower than a preset standard—for example, if its continuous output trust level is consistently below 0.5 in the last 10 updates—it will automatically reduce the contribution weight of that data source in the global risk score calculation, for example, by reducing its weight coefficient from 0.25 to 0.1. Simultaneously, it will mark a suspected abnormal interference state for the data stream associated with this child node in memory. Subsequently, it will trigger a local self-check procedure at the source of this data stream, such as a specific sensor terminal, for example, instructing the sensor to perform self-calibration or restart to troubleshoot the fault.

[0080] The entire process of data acquisition, mapping, updating, and weight adjustment is abstracted into a stateful stream processing topology, running on stream processing frameworks such as Apache Flink or Kafka Streams. Each processing step, such as data mapping, graph updating, and risk calculation, is encapsulated as an independent stream processing operator.

[0081] To address the issue of out-of-order arrival of sensor data due to network latency, a watermark mechanism is introduced into the data stream. Each data event carries an event timestamp. A fixed out-of-order tolerance period, such as 5 seconds, is set. When the watermark, representing the current time, exceeds the event time of a certain data point, that data is considered late. For late data that is still within the tolerance range, a rollback and recalculation operation is triggered: the previous calculation results based on incomplete data are rolled back, and the late data is included, re-executing all operator calculations within its time window to ensure the accuracy of the final result. The goal of the entire processing chain is to control the end-to-end delay from sensor data generation to triggering intervention commands, such as adjusting the brace, to within seconds.

[0082] According to a second embodiment of the present invention, the present invention claims protection for a system for constructing a closed-loop model of diabetic foot based on multi-source data, comprising: One or more processors; A memory storing one or more programs, which, when executed by one or more processors, enable the processors to implement the method for constructing a closed-loop model of diabetic foot based on multi-source data.

[0083] In summary, this invention, by constructing a dynamic data structure containing a master node and multiple child nodes, not only solves the data silo problem but also simulates the complex pathophysiological network of the human body through a cross-node association mechanism. The system's adaptive update and closed-loop control functions enable it to continuously evolve as the condition progresses, significantly improving the digitalization and intelligence level of diabetic foot management.

[0084] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

[0085] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0086] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0087] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A method for constructing a closed-loop model of diabetic foot based on multi-source data, characterized in that, Includes the following steps: Step 1: Simultaneously collect multi-source heterogeneous data, including dynamic pressure sequence of the sole of the foot, skin surface temperature gradient, changes in local blood oxygen saturation, peripheral nerve conduction velocity, and laboratory biochemical indicators, through sensor terminals, clinical information integration interfaces, and mobile detection devices deployed on the patient side. Step 2: Construct a dynamic data structure containing a master node and multiple child nodes, initialize the hierarchical logical topology, the master node is used to maintain the patient's global pathophysiological state baseline, and the multiple child nodes are associated with different pathophysiological dimensions respectively. Step 3: Map data items in the multi-source data stream to corresponding child nodes according to the timestamp alignment rules, extract the absolute timestamp or relative offset time in each data frame, apply a sliding window-based synchronization alignment algorithm, map data items from different sampling frequencies to the corresponding child node storage space in the dynamic data structure, and establish a time-series correlation graph in each child node with data items as graph nodes and the time interval and numerical evolution slope between adjacent data items as edge weights; Step 4: Monitor the input status of multi-source data streams in real time. When a new data item is detected, determine the target sub-node based on the data source identifier, and update the temporal correlation graph within the target sub-node using incremental learning logic that introduces a time decay factor, and recalculate the temporal feature vector. Step 5: Adjust the weight parameters of data items in other child nodes according to the preset cross-node influence rules. Based on the pathophysiological correlation logic, establish a cross-node intervention response matrix. Trigger the fine-tuning of the weights of adjacent or related child nodes by the state evolution of the target child node.

2. The method for constructing a closed-loop model of diabetic foot based on multi-source data according to claim 1, characterized in that, Step 1 also includes: For high-frequency sampled plantar pressure data, a pressure sensing array with a preset sampling frequency range is used for continuous acquisition. The pressure sensing array consists of multiple sensor units distributed in the first metatarsal bone, fifth metatarsal bone, arch, and calcaneus regions of the plantar surface. For low-frequency sampled biochemical test data, point data including glycated hemoglobin, white blood cell count, procalcitonin, C-reactive protein, and blood lipid indicators are obtained from hospital information through standard medical information exchange protocols; For the irregularly generated clinical description data, natural language processing technology is used to extract pathological entities from the chief complaint, present illness history and physical examination records in the electronic medical record. The pathological entities include redness and swelling of the soles of the feet, increased skin temperature and decreased tactile sensation, and are assigned corresponding time tags according to the time of generation of the pathology report or the time of issuance of the medical order. A Butterworth low-pass filter of a predetermined order is used to remove high-frequency noise generated by environmental electromagnetic interference, wherein the filter cutoff frequency is set at a preset multiple of the effective bandwidth of the signal. A median filtering algorithm is used to perform sliding processing on the signal to eliminate signal jump points caused by sensor momentary poor contact or motion artifacts. The sliding window size of the median filtering algorithm is dynamically adjusted according to the sampling frequency.

3. The method for constructing a closed-loop model of diabetic foot based on multi-source data according to claim 1, characterized in that, Step 2 also includes: The pathophysiological dimensions include foot tissue perfusion dimension, local infection load dimension, biomechanical stress dimension, and neuropathy degree dimension. The master node is configured to maintain a global state vector obtained by weighted fusion calculation of the feature values ​​fed back by each child node. The global state vector is used to characterize the overall deterioration risk score of diabetic foot. Each child node has a local data buffer for temporarily storing the original sampled data. The global state vector maintained by the master node is defined as a multidimensional tensor. The value of this tensor is normalized and weighted by transforming the local feature vectors output by each child node through the feature mapping matrix, and is used to reflect the patient's overall foot health status within the current time window. Each child node's internal local data buffer uses a circular queue structure to temporarily store the original values ​​within a time series of a preset length; The feature mapping matrix inside the child node serves as a transformation operator, used to transform the original numerical data into quantitative features that reflect the trend of pathological evolution. The quantitative features include the coefficient of variation or the slope of numerical decline of physiological indicators within a specific time period. The dynamic data structure is organized in memory in the form of a balanced tree or a directed acyclic graph to support the retrieval of data for each physiological dimension.

4. The method for constructing a closed-loop model of diabetic foot based on multi-source data according to claim 1, characterized in that, Step 3 also includes: The sliding window-based synchronization alignment algorithm uses a variable window with a preset time span. When a drastic change in the patient's movement state is detected, the window step value is automatically adjusted. To address data loss caused by inconsistent sampling frequencies, missing sampling points are identified and filled using a third-order spline interpolation algorithm within the same logical time slice. A piecewise cubic polynomial is constructed to ensure continuous derivatives at connection points. When establishing a temporal correlation graph within each child node, the turning point data of disease evolution are identified by calculating the centrality index of the graph. The centrality index includes degree centrality or eigenvector centrality. When the centrality index of a node changes abruptly and the weight of its adjacent edges shows nonlinear fluctuations, the node is marked as a pathological risk trigger point.

5. The method for constructing a closed-loop model of diabetic foot based on multi-source data according to claim 1, characterized in that, Step 4 also includes: Set a data disturbance threshold. When the deviation of the value of a newly added data item from the value of a data item of the same dimension at the previous time exceeds the first preset threshold, the incremental update program is activated immediately. The update program monitors the status of each child node in real time through a background listening process, and quickly locates the target child node after identifying the data source identifier. When using incremental learning logic, the influence of historical data is exponentially reduced through a time decay factor, so that the model prioritizes the retention of the feature contributions of recent physiological fluctuations. During the update process, the node increment and edge weight reconstruction of the graph are completed first in the memory buffer. After the calculation is completed, the updated topology is synchronized to the persistent storage layer to ensure real-time response to multi-source data streams.

6. The method for constructing a closed-loop model of diabetic foot based on multi-source data according to claim 1, characterized in that, Step 5 also includes: When the blood oxygen index in the foot tissue perfusion dimension sub-node decreases beyond the second preset threshold, the intervention response matrix triggers a logical judgment to automatically increase the representation weight of the inflammation index in the global state vector in the local infection load dimension sub-node. The weight adjustment range is determined by a predetermined adjustment ratio. The adjustment process of the weight parameters is associated with an adaptive feedback loop to evaluate in real time the accuracy of the global state vector after weight adjustment for clinical risk prediction. If the residual between the predicted risk value and the actual clinical observation value is greater than the preset error threshold, the coupling coefficient in the intervention response matrix is ​​corrected through the backpropagation mechanism to achieve dynamic self-optimization of the internal logical relationship of the model.

7. The method for constructing a closed-loop model of diabetic foot based on multi-source data according to claim 1, characterized in that, The method also involves a deep learning feature extraction process: Based on the constructed dynamic data structure, a deep convolutional neural network is applied to extract multi-dimensional spatial features, and high-pressure plaque regions in plantar pressure distribution images are identified through alternating convolutional and pooling layers. By combining long short-term memory network analysis with the temporal evolution of the output of each sub-node, the long-term trends and periodic fluctuations in the foot temperature change sequence or blood oxygen change sequence are captured by the forgetting gate and memory gate mechanisms. By fusing multi-scale features of spatial and temporal characteristics, the risk probability of foot ulcers occurring or the predicted healing progress of existing wounds within a preset prediction time period can be calculated.

8. The method for constructing a closed-loop model of diabetic foot based on multi-source data according to claim 1, characterized in that, The method achieves interaction with external intervention devices through a closed-loop control module: When the risk score output by the master node exceeds the preset safety threshold, and the foot pressure is continuously concentrated in a specific area and the local skin temperature rises above the preset temperature threshold, the inflation pressure adjustment command of the decompression brace is automatically triggered. This reduces the load on the pressure area by changing the support position, or pushes an intervention command to the medical staff. The dynamic data structure is persistently stored using a distributed graph database, and the data is sharded and stored on different physical server nodes to support high-concurrency relational queries for a large number of patient groups. The data quality verification process is executed. If the data loss duration of a certain sub-node exceeds the preset duration threshold, the historical average of similar patients or the inferred value of related sub-nodes are automatically called to offset the data, and the data credibility mark is recorded.

9. The method for constructing a closed-loop model of diabetic foot based on multi-source data according to claim 1, characterized in that, The method employs an arbitration mechanism when handling concurrent conflicts between multiple sensors: The master node initiates the arbitration module based on evidence theory, treating the feature outputs of each child node as independent sources of evidence. Each sub-node outputs a trust rating index based on the current data quality label, and the master node integrates contradictory information from different dimensions using the evidence synthesis rule. If the data stability of a certain child node is lower than the preset stability standard, the contribution weight of that data source to the global risk score will be automatically reduced, and it will be marked as a suspected abnormal interference in memory, while triggering a local self-check program. By performing task orchestration through a stream processing framework, data acquisition, mapping, updating, and weight adjustment are abstracted into stateful stream processing operators. A watermark mechanism is used to solve the problem of out-of-order arrival of data items, ensuring a closed-loop response to new data is completed within seconds.

10. A system for constructing a closed-loop model of diabetic foot based on multi-source data, characterized in that, include: One or more processors; A memory having stored one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method for constructing a closed-loop model of diabetic foot based on multi-source data according to any one of claims 1 to 9.

Images