Portable diabetes foot touch screening data management method and system

Abstract

The invention provides a portable diabetes foot tactile screening data management method and system, and relates to the technical field of medical information processing, and the method comprises the steps: obtaining diabetes parameters of a detected person, generating a stimulation sequence by using a Hash algorithm, and applying tactile stimulation; after receiving the sensing data, identifying and eliminating memory and guess pollution data to form a purified data set; mapping the spatial deviation to a foot nerve map to construct a lesion path map; and calculating a second derivative of the modulus length change rate of the differential vector to realize early warning. According to the invention, the accuracy of diabetic foot screening is improved, and early lesion early warning is realized.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a portable method and system for managing tactile screening data of diabetic foot. Background Technology

[0002] Diabetic foot is a common and serious complication of diabetes, primarily caused by peripheral neuropathy and vascular disease, which can lead to foot ulcers, infections, and even amputation. Early screening and monitoring are crucial for preventing the development of diabetic foot. Tactile sensory screening is an important tool for assessing nerve function in the foot of diabetic patients, enabling the early detection of signs of nerve damage.

[0003] Current diabetic foot screening primarily employs traditional methods such as the monofilament test and tuning fork vibration test, where doctors use handheld devices to perform tactile or vibrational stimulation tests on the patient's foot. With the development of information technology, some portable devices are used for diabetic foot screening; however, these devices are mostly independently operated and have limited data management capabilities. Existing tactile screening data acquisition and analysis systems for diabetic foot typically use fixed stimulation sequences, and the data analysis methods are relatively simple, making it difficult to effectively identify and handle interference factors during the testing process.

[0004] The main shortcomings and deficiencies of existing technologies are as follows: Traditional tactile screening methods for diabetic foot lack effective random stimulation mechanisms, making them susceptible to patient prediction and memory effects, leading to biased test results. Patients may respond through memory or guesswork rather than genuine sensory feedback, affecting screening accuracy; existing technologies struggle to identify and eliminate memory-contaminated and guesswork-contaminated data during the testing process, resulting in poor data quality and low reliability of analysis results. This contaminated data can mask the true neuropathy condition, affecting doctors' diagnostic judgment; existing diabetic foot screening systems focus primarily on static measurement results, lacking effective tracking and early warning mechanisms for dynamic disease progression. This inability to promptly detect trends of accelerated disease deterioration leads to missed optimal intervention opportunities, causing patients to miss the golden period of early treatment and increasing the difficulty and cost of later treatment. Summary of the Invention

[0005] This invention provides a portable method and system for managing tactile screening data of diabetic foot, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides a portable method for managing tactile screening data of diabetic foot, comprising: Obtain diabetes-related parameters from the subjects and determine the set of detection locations and standard stimulus intensity; A sequence seed is generated based on hash operation. A stimulus sequence and a time offset sequence are randomly generated based on the sequence seed. These are combined to construct a hidden stimulus matrix. The sensor array is then driven to apply tactile stimuli according to the hidden stimulus matrix. Receive the sensing location and sensing intensity, calculate spatial deviation, intensity deviation and response delay to form the raw dataset; A trajectory point sequence is constructed from the spatial deviation sequence of the original dataset, spatial memory attractors are labeled, memory contamination data is identified, the response delay sequence of the original dataset is symbolically encoded, guess contamination data is identified, memory contamination data and guess contamination data are removed to form a clean dataset. Spatial biases in the purified dataset are mapped to foot neural atlases to construct weighted graphs. Based on graph cut algorithm and depth-first traversal, lesion paths connecting abnormal nodes are extracted to generate lesion path graphs. The first detected clean dataset is stored as a baseline dataset. The second derivative of the rate of change of the difference vector magnitude between the subsequent clean datasets and the baseline dataset is calculated. When the second derivative is positive, an acceleration warning signal is generated. Upload the lesion path map and the accelerated early warning signal.

[0007] In one optional embodiment, a sequence seed is generated based on a hash operation; a stimulus sequence and a time offset sequence are randomly generated based on the sequence seed; a hidden stimulus matrix is ​​constructed by combining the sequences; and tactile stimuli are applied to the sensing array according to the hidden stimulus matrix, including: The current detection timestamp and the detected person's identifier are concatenated to form an initial string. A hash operation is performed on the initial string to generate a sequence seed. A pseudo-random number generator is initialized based on the sequence seed to obtain a random number stream. Extract a random floating-point number sequence equal to the number of detection positions from the random number stream, normalize the random floating-point number sequence, map it to each detection position index, sort the mapping results using a sorting algorithm to generate a stimulus order sequence, and simultaneously extract a second set of random numbers from the random number stream and generate a non-uniform time offset sequence through exponential distribution transformation. A hidden stimulus matrix is ​​constructed by combining the stimulus sequence and the time offset sequence. The sensor array is driven to apply tactile stimulation according to the spatiotemporal coordinates of the hidden stimulus matrix. Only the foot outline is displayed on the interactive interface to mask the actual activation position.

[0008] In one optional embodiment, extracting a second set of random numbers from the random number stream and generating a non-uniform time offset sequence by exponential distribution transformation includes: Extract a second set of uniformly distributed random numbers from the random number stream, perform a negative logarithmic transformation on each random value in the second set of random numbers to obtain the original time offset of the exponential distribution, and form the original offset sequence. The expected reaction time window of the subject is calculated based on the nerve conduction velocity parameter in the diabetes-related parameters. The ratio of the expected reaction time window to the preset health baseline reaction time is used as the time scale scaling factor to proportionally adjust each original time offset in the original offset sequence. Set upper and lower boundary thresholds for the adjusted offset sequence, and truncate the offsets that exceed the upper and lower boundary thresholds to the boundary values ​​to generate a non-uniform time offset sequence that meets the physiological perception range.

[0009] In one optional embodiment, constructing a trajectory point sequence from the spatial deviation sequence of the original dataset, labeling spatial memory attractors, and identifying memory-contaminated data include: Spatial deviation sequence is extracted from the original dataset. The deviation values ​​at consecutive time points in the spatial deviation sequence are combined to construct a multidimensional vector. Each multidimensional vector is used as a trajectory point in the phase space to form a trajectory point sequence. Density clustering is performed on the trajectory point sequence to obtain cluster centers. The number of times each trajectory point returns to the preset neighborhood of the cluster center is calculated. The ratio of the number of returns to the total number of trajectory points is calculated to determine the trajectory return rate. When the trajectory return rate exceeds a preset return threshold, the cluster center is marked as a spatial memory attractor. Calculate the distance from each trajectory point to the spatial memory attractor, count the density peak points of the distance distribution, delineate the boundary of the attraction domain using the distance values ​​corresponding to the density peak points, and mark the data in the original dataset whose corresponding trajectory points fall within the boundary of the attraction domain as memory contamination data.

[0010] In one alternative embodiment, the response delay sequence of the original dataset is symbolically encoded to identify suspected contamination data, including: Response delay sequences are extracted from the original dataset. Multiple interval thresholds are set according to the numerical range of response delay values. Each response delay value in the response delay sequence is compared with the interval thresholds, and each response delay value is mapped to the corresponding symbol category to form a symbol sequence. The transition relationships between adjacent symbols in the symbol sequence are statistically analyzed, the number of transitions between each symbol pair is recorded, a symbol transition matrix is ​​constructed, the conditional probability distribution of symbol transitions is calculated based on the number of transitions in the symbol transition matrix, the conditional entropy is calculated based on the conditional probability distribution, the joint entropy is calculated based on the joint occurrence probability of symbols, and the response mode complexity is determined based on the ratio of the conditional entropy to the joint entropy. When the complexity of the response pattern is lower than a preset complexity threshold, the original data within the corresponding time period is marked as speculative contaminated data.

[0011] In one optional embodiment, spatial biases in the purified dataset are mapped to a foot neural atlas to construct a weighted graph. Lesion paths connecting aberrant nodes are extracted using a graph cut algorithm and depth-first traversal to generate a lesion path graph, including: The location coordinates and spatial deviation values ​​of each tactile stimulus are extracted from the cleaned dataset. Each location coordinate is mapped to a neural node in a preset foot neural atlas. The corresponding spatial deviation value is used as the node attribute value. The degree of deviation of each node attribute value from the preset normal threshold is calculated. Neural nodes whose deviation exceeds the preset abnormal judgment threshold are marked as abnormal nodes. The neural connection weights between adjacent nerve nodes in the foot neural atlas are calculated. The weighted combination of the spatial deviation difference of each nerve node and the neural anatomical distance is used as the edge weight. A weighted graph containing all nerve nodes and edge weights is constructed. A node affinity matrix is ​​constructed and a graph cut algorithm is performed on the weighted graph to segment the abnormal node subgraph and the normal node subgraph. The boundary nodes of the abnormal node subgraph are extracted. A depth-first traversal is performed based on the boundary nodes to search for the set of paths connecting each abnormal node. The joint score of the cumulative edge weight of each path and the degree of node deviation is calculated. The path with the highest joint score is selected as the lesion path. The neural node sequence and edge weight sequence in the lesion path are visualized and mapped to generate a lesion path graph.

[0012] In one optional embodiment, a node affinity matrix is ​​constructed to perform a graph cut algorithm on the weighted graph to segment the abnormal node subgraph from the normal node subgraph, and the boundary nodes of the abnormal node subgraph are extracted, including: The spatial deviation values ​​of each neural node in the weighted graph are normalized. The attribute similarity between adjacent neural nodes is calculated based on the normalized spatial deviation values. A node affinity matrix is ​​constructed based on the attribute similarity and edge weights. The attribute similarity is used as the connection strength between nodes. In a weighted graph, a source node and a sink node are defined. All abnormal nodes are connected to the source node and assigned connection weights based on the node affinity matrix. All normal nodes are connected to the sink node and assigned connection weights based on the node affinity matrix. Capacity allocation is performed on each edge in the weighted graph. Maximum flow minimum cut calculation is performed to determine the edge cut set that maximizes the total capacity from the source node to the sink node. The edge cut set divides the weighted graph into an abnormal node subgraph containing the source node and a normal node subgraph containing the sink node. Extract the boundary edges connecting the abnormal node subgraph and the normal node subgraph in the edge cut set, and identify the neural nodes in the abnormal node subgraph connected by the boundary edges as boundary nodes.

[0013] A second aspect of the present invention provides a portable tactile screening data management system for diabetic foot, comprising: The parameter setting unit is used to acquire diabetes-related parameters of the subject and determine the set of detection locations and standard stimulus intensity. The stimulus execution unit is used to generate a sequence seed based on hash operation, randomly generate a stimulus sequence and a time offset sequence according to the sequence seed, combine them to construct a hidden stimulus matrix, and drive the sensor array to apply tactile stimulation according to the hidden stimulus matrix. The data acquisition unit is used to receive the sensing location and sensing intensity, calculate the spatial deviation, intensity deviation and response delay, and form the raw dataset; The data cleansing unit is used to construct trajectory point sequences from the spatial deviation sequences of the original dataset, label spatial memory attractors, identify memory contamination data, symbolically encode the response delay sequences of the original dataset, identify guess contamination data, remove memory contamination data and guess contamination data, and form a cleaned dataset. The path generation unit is used to map the spatial bias in the purified dataset to the foot neural atlas to construct a weighted graph. Based on the graph cut algorithm and depth-first traversal, the lesion path connecting abnormal nodes is extracted to generate a lesion path graph. An accelerated warning unit is used to store the first detected clean dataset as a baseline dataset, calculate the second derivative of the rate of change of the difference vector magnitude between the subsequent clean dataset and the baseline dataset, and generate an accelerated warning signal when the second derivative is positive. The data upload unit is used to upload lesion path maps and accelerated early warning signals.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] In this embodiment of the invention, by generating sequence seeds based on hash operations and constructing a hidden stimulus matrix, the randomization of stimulus order and diversification of time offset are achieved, effectively avoiding patients' memory and prediction of test patterns, and improving the authenticity and reliability of screening data. A spatial memory attractor labeling and response delay symbolic coding technique is proposed, which can effectively identify and eliminate memory-contaminated and guess-contaminated data, forming a high-quality clean dataset and significantly improving the data accuracy of tactile screening for diabetic foot. By mapping spatial deviations to a foot neuropathy atlas and constructing a weighted graph, combined with graph cut algorithms and depth-first traversal techniques, accurate extraction and visualization of lesion paths are achieved, providing clinicians with intuitive and accurate lesion distribution information. By calculating the second derivative of the rate of change of the magnitude of the difference vector in the dataset, an innovative accelerated early warning mechanism for foot neuropathy is established, which can detect the trend of disease deterioration early, providing a time window for clinical intervention and effectively reducing the risk of diabetic foot ulcers and amputation. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the portable diabetic foot tactile screening data management method according to an embodiment of the present invention. Figure 2 This is a flowchart for lesion path analysis. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the portable diabetic foot tactile screening data management method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Obtain diabetes-related parameters from the subjects and determine the set of detection locations and standard stimulus intensity; A sequence seed is generated based on hash operation. A stimulus sequence and a time offset sequence are randomly generated based on the sequence seed. These are combined to construct a hidden stimulus matrix. The sensor array is then driven to apply tactile stimuli according to the hidden stimulus matrix. Receive the sensing location and sensing intensity, calculate spatial deviation, intensity deviation and response delay to form the raw dataset; A trajectory point sequence is constructed from the spatial deviation sequence of the original dataset, spatial memory attractors are labeled, memory contamination data is identified, the response delay sequence of the original dataset is symbolically encoded, guess contamination data is identified, memory contamination data and guess contamination data are removed to form a clean dataset. Spatial biases in the purified dataset are mapped to foot neural atlases to construct weighted graphs. Based on graph cut algorithm and depth-first traversal, lesion paths connecting abnormal nodes are extracted to generate lesion path graphs. The first detected clean dataset is stored as a baseline dataset. The second derivative of the rate of change of the difference vector magnitude between the subsequent clean datasets and the baseline dataset is calculated. When the second derivative is positive, an acceleration warning signal is generated. Upload the lesion path map and the accelerated early warning signal.

[0021] One alternative implementation involves early screening and monitoring of diabetic foot neuropathy. Diabetes-related parameters of the subject are acquired, including blood glucose levels, glycated hemoglobin levels, nerve conduction velocity, disease duration, and age. These parameters are used for subsequent determination of detection locations and data processing. Based on these parameters, a set of foot detection locations is determined, typically including key nerve distribution points such as the big toe ball, medial arch, heel, and dorsum of the foot. A standard stimulation intensity is then set as a baseline tactile threshold.

[0022] To ensure the objectivity of tactile stimulus detection, a randomization strategy is employed to generate the stimulus sequence. The current detection timestamp and the subject's identifier are concatenated to form an initial string. This string is then subjected to a SHA-256 hash operation to generate a sequence seed. A pseudo-random number generator is initialized based on the sequence seed to obtain a random number stream. A sequence of random floating-point numbers, equal in number to the number of detection positions, is extracted from the random number stream. After normalization, this sequence is mapped to the indices of each detection position, and then sorted to generate a stimulus order sequence. Simultaneously, a second set of random numbers is extracted from the random number stream and transformed using an exponential distribution to generate a non-uniform time offset sequence.

[0023] A hidden stimulus matrix is ​​constructed by combining the stimulus sequence and the time offset sequence, forming the spatiotemporal coordinates of the detection location, stimulus time point, and stimulus intensity. During tactile stimulation, the interactive interface only displays the foot outline and not the actual activation location to reduce the influence of psychological expectations on the detection results. According to the spatiotemporal coordinates of the hidden stimulus matrix, the system drives the sensor array to apply tactile stimulation to each detection point on the foot.

[0024] When receiving data on perceived location and intensity, the system records the subject's response to each stimulus. Spatial deviation is obtained by calculating the Euclidean distance between the actual stimulus location and the perceived location; intensity deviation is obtained by calculating the difference between the actual stimulus intensity and the perceived intensity; and response delay is obtained by calculating the time difference from stimulus application to perceived feedback. These data are then aggregated to form the original dataset.

[0025] The original dataset often contains data contamination caused by subjective factors, requiring data purification. A trajectory point sequence is constructed from the spatial deviation sequence, combining the spatial deviation values ​​at consecutive time points into a multi-dimensional vector to form trajectory points in phase space. Density clustering is performed on the trajectory point sequence to obtain cluster centers. The number of times each trajectory point returns to a preset neighborhood of the cluster center is calculated, and the return rate is calculated. When the return rate exceeds a preset return threshold, the cluster center is marked as a spatial memory attractor. The distance distribution from each trajectory point to this attractor is calculated, and density peak points are identified. The attraction domain boundary is defined by these density peak points, and data falling within this boundary is marked as memory-contaminated data and removed from the dataset.

[0026] Simultaneously, the response delay sequence is symbolically encoded, and interval thresholds are set to map delay values ​​to symbol sequences. The symbol transition relationships are statistically analyzed, a symbol transition matrix is ​​constructed, and the conditional entropy, joint entropy, and their ratio are calculated to obtain the response pattern complexity. When the complexity is below a preset threshold, it indicates that there is a regularity in the data; data within the corresponding time period are marked as speculative contamination data and removed.

[0027] After purification, the spatial biases in the purified dataset are mapped to a foot neural atlas to construct a weighted graph. Abnormal nodes are marked based on the degree of deviation of the spatial bias value from a preset normal threshold. The neural connection weights between adjacent neural nodes are calculated, taking into account both the difference in spatial bias values ​​and the anatomical distance between the nodes. A weighted graph containing all node and edge weights is constructed, forming a node affinity matrix.

[0028] When executing the graph cut algorithm, abnormal nodes are connected to the source node, and normal nodes are connected to the sink node, thus dividing the graph into an abnormal node subgraph and a normal node subgraph. Boundary nodes are extracted, and from these boundary nodes, a depth-first traversal is performed to search for the set of paths connecting all abnormal nodes. The cumulative edge weights and node deviations of each path are evaluated, and the path with the highest score is selected as the lesion path, generating the lesion path graph.

[0029] The lesion path diagram visually illustrates the spatial distribution characteristics of neuropathies, including the lesion origin, propagation path, and severity. Each node in the diagram represents a nerve detection point, the intensity of the node's color indicates the severity of spatial deviation, the edges connecting the nodes represent the nerve conduction path, and the thickness of the edges represents the strength of the nerve connection.

[0030] The clean dataset from the initial test is stored as the baseline dataset. In subsequent periodic tests, the magnitude of the difference vector between the new clean dataset and the baseline dataset is calculated to obtain the overall deviation value. By calculating the rate of change of the deviation value and its second derivative over multiple consecutive tests, the disease progression trend can be monitored. When the second derivative is positive, it indicates that the disease is deteriorating rapidly, and the system generates an accelerated warning signal.

[0031] By uploading lesion pathway maps and early warning signals to the medical information system, doctors can adjust treatment plans in a timely manner based on this data, and patients can also intuitively understand the status of foot neuropathy. Through long-term follow-up monitoring, treatment effectiveness can be evaluated, and the occurrence of serious complications such as diabetic foot ulcers can be prevented.

[0032] In this embodiment, a random stimulation strategy and data purification process are employed to reduce the impact of subjective factors on the detection results and improve data reliability. Through neural atlas mapping and graph theory analysis, a visual representation of the spatial distribution of lesions is achieved, providing an intuitive basis for clinical diagnosis and treatment. The second-derivative early warning mechanism provides an early warning of accelerated disease deterioration, facilitating timely intervention and reducing the risk of complications.

[0033] In one optional implementation, a sequence seed is generated based on a hash operation; a stimulus sequence and a time offset sequence are randomly generated based on the sequence seed; these are combined to construct a hidden stimulus matrix; and the sensor array is driven to apply tactile stimuli according to the hidden stimulus matrix, including: The current detection timestamp and the detected person's identifier are concatenated to form an initial string. A hash operation is performed on the initial string to generate a sequence seed. A pseudo-random number generator is initialized based on the sequence seed to obtain a random number stream. Extract a random floating-point number sequence equal to the number of detection positions from the random number stream, normalize the random floating-point number sequence, map it to each detection position index, sort the mapping results using a sorting algorithm to generate a stimulus order sequence, and simultaneously extract a second set of random numbers from the random number stream and generate a non-uniform time offset sequence through exponential distribution transformation. A hidden stimulus matrix is ​​constructed by combining the stimulus sequence and the time offset sequence. The sensor array is driven to apply tactile stimulation according to the spatiotemporal coordinates of the hidden stimulus matrix. Only the foot outline is displayed on the interactive interface to mask the actual activation position.

[0034] In one specific implementation, the current detection timestamp, accurate to the millisecond level, is obtained, along with the unique identifier of the detected entity. The timestamp and the detected entity identifier are concatenated to form an initial string. For example, when the detection time is "2023-05-10 15:30:45.678" and the detected entity identifier is "P12345", the initial string "2023-05-10 15:30:45.678_P12345" is generated.

[0035] The initial string is hashed using SHA-256 to generate a 256-bit hash value, which is then converted into a 64-bit integer to serve as the sequence seed. This hashing operation ensures that even small changes in the input parameters will result in significantly different outputs, enhancing the randomness and unpredictability of the stimulus sequence.

[0036] Use a sequence seed to initialize the pseudo-random number generator, such as the Mersenne Twister algorithm, to ensure the repeatability of the random sequence. The same sequence seed will always generate the same random number sequence, which facilitates subsequent verification and result analysis. For example, when the sequence seed value is "1234567890", calling the pseudo-random number generator will yield a series of random floating-point numbers in the range [0, 1).

[0037] Extract a set of random floating-point numbers from the random number stream, the number of which is equal to the number of detection positions. Assuming the foot sensor array has 16 detection positions, extract 16 random floating-point numbers, such as [0.23, 0.87, 0.56, 0.12, 0.65, 0.34, 0.78, 0.45, 0.91, 0.39, 0.67, 0.21, 0.54, 0.88, 0.32, 0.76].

[0038] The extracted random floating-point number sequence is normalized to ensure uniform numerical distribution. The normalized floating-point numbers are then mapped to detection position indices, ranging from [0, number of detection positions - 1]. For example, the above sequence, after normalization and mapping, yields the following position indices: [3, 14, 9, 1, 10, 5, 12, 7, 15, 6, 11, 2, 8, 13, 4, 0].

[0039] The mapping results are sorted using a sorting algorithm such as quicksort to generate the final stimulus sequence. The stimulus sequence determines the order in which tactile stimuli are applied at different detection locations.

[0040] Another set of random numbers is extracted from the random number stream to generate the time offset sequence. These random numbers are then processed through the exponential distribution function λe. -λt The transformation, where λ is a parameter (e.g., λ=2), generates a non-uniform time offset. This ensures that the stimulus intervals have both short pauses and longer intervals, making it difficult for the subject to predict the timing of the next stimulus. For example, the generated time offset sequence might be [0.15s, 0.72s, 0.43s, 0.31s, 0.86s, 0.25s, 0.52s, 0.67s, 0.39s, 0.45s, 0.28s, 0.93s, 0.61s, 0.37s, 0.54s, 0.76s].

[0041] A hidden stimulus matrix is ​​constructed by combining the stimulus sequence with the time offset sequence. Each row of the matrix contains the detection location index and the corresponding trigger time. For example, the first row [3, 0.15s] indicates that a stimulus is applied at location index 3 0.15 seconds after the reference time.

[0042] Based on the spatiotemporal coordinate information of the hidden stimulus matrix, the tactile sensor array is driven to apply tactile stimuli in a predetermined order and at predetermined time points. The duration of each stimulus is fixed (e.g., 0.2 seconds), and the stimulus intensity can be preset based on detection requirements. For example, when the detection time reaches the reference time + 0.15 seconds, the tactile sensor at position index 3 is activated; when the detection time reaches the reference time + 0.15s + 0.72s = 0.87s, the tactile sensor at position index 14 is activated, and so on.

[0043] Throughout the testing process, the interactive interface only displays a foot outline, not the actual activated location, to avoid visual cues affecting the subject's tactile perception judgment. The subject marks the locations where they perceive tactile stimuli using a touchscreen or a dedicated input device.

[0044] After the test is completed, the marked location of the subject is compared with the actual stimulus location, and indicators such as accuracy and reaction time are calculated to assess the subject's tactile perception ability. The system can also record complete test process data, including stimulus sequence, timestamps, and response results, for subsequent analysis or comparison.

[0045] In this embodiment, hashing and pseudo-random number generation mechanisms ensure that the stimulus sequence for each detection is unique and unpredictable. Furthermore, the use of the same seed generation algorithm ensures reproducibility of the detection results, facilitating verification and analysis by researchers. Additionally, the non-uniform time offset design increases the difficulty and realism of the detection, more closely resembling actual tactile perception scenarios.

[0046] In one optional implementation, extracting a second set of random numbers from the random number stream and generating a non-uniform time offset sequence by exponential distribution transformation includes: Extract a second set of uniformly distributed random numbers from the random number stream, perform a negative logarithmic transformation on each random value in the second set of random numbers to obtain the original time offset of the exponential distribution, and form the original offset sequence. The expected reaction time window of the subject is calculated based on the nerve conduction velocity parameter in the diabetes-related parameters. The ratio of the expected reaction time window to the preset health baseline reaction time is used as the time scale scaling factor to proportionally adjust each original time offset in the original offset sequence. Set upper and lower boundary thresholds for the adjusted offset sequence, and truncate the offsets that exceed the upper and lower boundary thresholds to the boundary values ​​to generate a non-uniform time offset sequence that meets the physiological perception range.

[0047] In one specific implementation, a second set of uniformly distributed random numbers is extracted from the random number stream. The random number stream can be obtained using a pseudo-random number generator, for example, by generating a sequence of random values ​​uniformly distributed in the interval [0, 1] using the linear congruence method or the Mason tween rotation algorithm. The extracted second set of random numbers can be represented as {r1, r2, ..., r...} n}, where each r i All are uniformly distributed random numbers in the interval (0, 1), and n is the preset number of test stimuli, usually set to 30 to 50 times.

[0048] Perform a negative logarithmic transformation on each random value in the second set of random numbers to obtain the original time offsets of the exponential distribution, forming the original offset sequence. For each random number r... i Through formula t i = -ln(r i The transformation yields the original time offset t, which conforms to an exponential distribution. i This transformation is based on the principle of probability integral transformation in probability theory, converting a uniform distribution into an exponential distribution with parameter λ=1. The resulting original offset sequence {t1, t2, ..., t...} is... n It has an average value of 1, and the unit is seconds.

[0049] To accommodate the varying physiological states of different test subjects, the expected reaction time window needs to be calculated based on the nerve conduction velocity parameter, a key parameter related to diabetes. The nerve conduction velocity parameter can be obtained through prior nerve conduction velocity measurement or estimated based on clinical indicators such as the test subject's history of diabetes, age, and HbA1c level. For example, a patient with mild diabetes may have a nerve conduction velocity of 95% of normal, while a patient with severe diabetes may only have 70% of normal. Let the baseline nerve conduction velocity for healthy individuals be v. h (Typically 55-60 m / s), the nerve conduction velocity of the subject is v p The expected reaction time window can then be approximated as the healthy baseline reaction time window multiplied by v. h / v p .

[0050] If the health baseline reaction time window T h The expected reaction time window T for the subject is set to 0.2 seconds. p = T h × (v h / v pIf the time scale scaling factor s = T, then... p / T h = v h / v p Applying this scaling factor to each of the original time offsets in the original offset sequence yields the adjusted offset t'. i = t i × s. This scaling adjustment ensures that the time offset matches the neurophysiological state of the subject being tested.

[0051] The adjusted offset sequence contains extreme values ​​that exceed the reasonable range of human perception. To ensure the validity of the test, upper and lower boundary thresholds need to be set for the adjusted offset sequence. The upper boundary threshold T... max The time limit is typically set to 5 seconds to avoid prolonged waiting time that could lead to distraction in test subjects; the lower bound threshold T min The time offset is typically set to 0.5 seconds to ensure the subject has sufficient time to prepare for the next stimulus. For any adjusted time offset t'ᵢ, if t'ᵢ > T max Then truncate it to T max If t'ᵢ < T min Then truncate it to T min After boundary processing, the final non-uniform time offset sequence {t"1, t"2, ..., t"} is obtained. n}

[0052] In practice, a sequence containing a large number of random numbers (e.g., 100) is first generated, and then the required number of time offsets are selected from it according to the actual test needs. This pre-generation strategy can improve the flexibility and efficiency of the system.

[0053] Furthermore, statistical analysis is performed on the final generated non-uniform time offset sequence to calculate its statistical characteristics such as mean, standard deviation, maximum, and minimum values, ensuring that its distribution characteristics meet the test requirements. For example, for standard tests, the average time offset should be close to s seconds after adjustment, and the standard deviation should be controlled within a reasonable range to ensure the consistency and comparability of the tests.

[0054] In this embodiment, the generated non-uniform time offset sequence takes into account the individualized neurophysiological characteristics of the test subjects and ensures the unpredictability of the test stimuli, thereby improving the scientificity and accuracy of the visual evoked reaction time test for diabetes.

[0055] In one alternative implementation, a trajectory point sequence is constructed from the spatial deviation sequence of the original dataset, spatial memory attractors are labeled, and memory contamination data is identified, including: Spatial deviation sequence is extracted from the original dataset. The deviation values ​​at consecutive time points in the spatial deviation sequence are combined to construct a multidimensional vector. Each multidimensional vector is used as a trajectory point in the phase space to form a trajectory point sequence. Density clustering is performed on the trajectory point sequence to obtain cluster centers. The number of times each trajectory point returns to the preset neighborhood of the cluster center is calculated. The ratio of the number of returns to the total number of trajectory points is calculated to determine the trajectory return rate. When the trajectory return rate exceeds a preset return threshold, the cluster center is marked as a spatial memory attractor. Calculate the distance from each trajectory point to the spatial memory attractor, count the density peak points of the distance distribution, delineate the boundary of the attraction domain using the distance values ​​corresponding to the density peak points, and mark the data in the original dataset whose corresponding trajectory points fall within the boundary of the attraction domain as memory contamination data.

[0056] In one specific implementation, a spatial skew sequence is extracted from the original dataset. The spatial skew sequence refers to the spatial location difference between adjacent data points in a time series. Specifically, if the spatial location at time i in the original dataset is P... i The spatial deviation can then be expressed as ΔP i = P i - P i-1 The extracted spatial bias sequence is {ΔP1, ΔP2, ..., ΔP}. n}, where n is the number of data points in the original dataset.

[0057] To construct a sequence of trajectory points, the spatial deviations at consecutive time points need to be combined into a multidimensional vector. The spatial deviation values ​​at m consecutive time points are selected and combined to form an m-dimensional vector, which serves as a trajectory point in the phase space. For example, when m=3, the j-th trajectory point can be represented as V. j = [ΔP j ΔP j+1 ΔP j+2 The sliding window method can be used to obtain the trajectory point sequence {V1, V2, ..., V...}. n-m+1 These trajectory points form the trajectory of the system's dynamic evolution in the m-dimensional phase space.

[0058] Density-based clustering (DBSCAN) is performed on the trajectory point sequence to obtain cluster centers. DBSCAN, a density-based spatial clustering algorithm, is used to cluster the trajectory points. This algorithm can discover clusters of arbitrary shapes and automatically determine the number of clusters. The DBSCAN algorithm requires two parameters: the neighborhood radius ε and the minimum number of points MinPts. For each trajectory point, the number of points in its ε-neighborhood is calculated. If the number of points is not less than MinPts, the point is marked as a core point. Core points that are density-reachable from each other are grouped into the same cluster. The cluster center is defined as the mean vector of all points within the cluster.

[0059] In practice, for diabetic foot tactile screening data, ε=0.15 (normalized unit) and MinPts=5 were selected as density clustering parameters. In a sample containing 200 trajectory points, assuming the spatial deviation range is between [-0.5, 0.5], after density clustering, three clusters were obtained, containing 78, 65, and 42 trajectory points respectively, with the remaining 15 points identified as noise points. The calculated three cluster centers are C1=[0.12, 0.08, -0.03], C2=[-0.15, -0.08, 0.04], and C3=[0.02, -0.21, 0.18].

[0060] For each obtained cluster center, the trajectory return rate is calculated to determine whether it is a spatial memory attractor. The preset neighborhood of a cluster center is defined as a hypersphere region with radius r centered at the cluster center. For the trajectory point sequence {V1, V2, ..., V_(n-m+1)}, the number of times each trajectory point enters the preset neighborhood is calculated, denoted as N_return. The trajectory return rate equals N_return divided by the total number of trajectory points (n-m+1). When the trajectory return rate exceeds a preset return threshold θ (e.g., θ=0.15), the cluster center is marked as a spatial memory attractor.

[0061] In the above samples, for cluster centers C1, C2, and C3, the number of times the trajectory points returned to the preset neighborhood was calculated to be 35, 28, and 18, respectively, with corresponding trajectory return rates of 0.175, 0.14, and 0.09. Since the trajectory return rate of C1 (0.175) is greater than the preset return threshold of 0.15, C1 is marked as a spatial memory attractor.

[0062] To identify memory contamination data, the Euclidean distance from each trajectory point to the spatial memory attractor is calculated. All distance values ​​are plotted as a histogram to observe the distance distribution. Density peaks are identified within this distribution, and the distance value corresponding to these peaks is defined as the attraction domain boundary distance, D_boundary. Specifically, a smooth curve of the distance distribution is obtained using kernel density estimation, and then local maxima are identified as density peaks.

[0063] In practical processing, a Gaussian kernel function is selected for kernel density estimation, with a bandwidth of 0.05. For the C1 spatial memory attractor, the calculated distance distribution exhibits a bimodal characteristic, with the first density peak appearing at a distance of 0.08 and the second density peak appearing at a distance of 0.25. The first peak represents data points strongly influenced by spatial memory, while the second peak represents data points less affected. The first density peak of 0.08 is taken as the boundary distance D_boundary of the attraction domain.

[0064] Data points in the original dataset whose distance to the spatial memory attractor is less than D_boundary are labeled as memory-contaminated data. In the above sample, 52 trajectory points have a distance less than 0.08, and these data points are labeled as memory-contaminated data. These data points appear as dense regions forming around the spatial memory attractor in phase space, reflecting the potential positional memory effect that subjects may have during tactile screening, i.e., subjects report tactile locations based on memory rather than actual perception.

[0065] In practical applications, spatial memory contamination also manifests as a gradual increase in the accuracy of feedback from the subject when repeated tactile stimuli occur at similar locations. For example, in a diabetic foot screening, after three consecutive triggers of stimulation in the second toe bulb region on the sole of the foot, even with reduced stimulation intensity, the subject could still quickly identify the stimulation location, while exhibiting significant sensory dullness in other areas. This phenomenon can be effectively identified using the algorithm described above: in the multidimensional phase space, the trajectory points surrounding this region form a significant cluster, and the calculated trajectory return rate is 0.22, far exceeding the preset threshold of 0.15, indicating a strong spatial memory effect.

[0066] For the identified memory contamination data, this data was removed from the original dataset, retaining only the data that truly reflects the tactile sensitivity of the foot. In the above sample, after removing 52 memory contamination data points, the remaining 148 data points constituted the purified dataset. These data points exhibited a more dispersed and random spatial distribution, and could more accurately reflect the differences in tactile sensitivity in different areas of the foot.

[0067] The effectiveness of the memory contamination identification algorithm was verified by comparing the variance of the datasets before and after screening. In a typical test, the spatial bias variance of the original dataset was 0.023. After removing the memory contamination data, the spatial bias variance of the cleaned dataset increased to 0.038, indicating enhanced data dispersion and a better reflection of the actual differences in nerve perception in different areas of the foot. The spatial bias distribution of the tactile screening data processed by this method is closer to the actual neuropathy state, avoiding data distortion caused by the memory effect, and providing a more reliable data foundation for subsequent assessment of diabetic foot neuropathy.

[0068] In one alternative implementation, the response delay sequence of the original dataset is symbolically encoded to identify suspected contamination data, including: Response delay sequences are extracted from the original dataset. Multiple interval thresholds are set according to the numerical range of response delay values. Each response delay value in the response delay sequence is compared with the interval thresholds, and each response delay value is mapped to the corresponding symbol category to form a symbol sequence. The transition relationships between adjacent symbols in the symbol sequence are statistically analyzed, the number of transitions between each symbol pair is recorded, a symbol transition matrix is ​​constructed, the conditional probability distribution of symbol transitions is calculated based on the number of transitions in the symbol transition matrix, the conditional entropy is calculated based on the conditional probability distribution, the joint entropy is calculated based on the joint occurrence probability of symbols, and the response mode complexity is determined based on the ratio of the conditional entropy to the joint entropy. When the complexity of the response pattern is lower than a preset complexity threshold, the original data within the corresponding time period is marked as speculative contaminated data.

[0069] In one specific implementation, extracting the response delay sequence from the raw dataset is the first step in speculative contamination data identification. Response delay refers to the time interval, measured in milliseconds, between the application of a tactile stimulus and the subject's response. For diabetic foot tactile screening data, response delays typically range from 200 to 2000 milliseconds, while in the normal population, response delays are concentrated between 400 and 800 milliseconds. In a set of tactile screening samples containing 150 data points, the extracted response delay sequence was {645, 723, 512, 698, 843, 912, 756, 632, 578, ..., 825} milliseconds, reflecting the subject's reaction speed to tactile stimuli at different foot locations.

[0070] The purpose of setting interval thresholds based on the numerical range of response delay values ​​is to discretize continuous response delay values ​​into a finite number of symbol categories. In the application of tactile screening for diabetic foot, the set interval thresholds are [400, 600, 800, 1000, 1200] milliseconds, mapping response delay values ​​to 5 symbol categories, denoted as A, B, C, D, and E respectively. The specific mapping rules are as follows: response delays less than 400 milliseconds are mapped to symbol A; those between 400 and 600 milliseconds are mapped to symbol B; those between 600 and 800 milliseconds are mapped to symbol C; those between 800 and 1000 milliseconds are mapped to symbol D; those between 1000 and 1200 milliseconds are mapped to symbol E; and those greater than 1200 milliseconds are mapped to symbol F. After symbolic encoding, the above response delay sequence forms the symbol sequence {C, C, B, C, D, D, C, C, B, ..., D}.

[0071] This study analyzes the transition relationships between adjacent symbols in a sequence of symbols, recording the number of transitions between each symbol pair to construct a symbol transition matrix. The symbol transition matrix is ​​a two-dimensional table where rows represent the current symbol and columns represent the next symbol. The values ​​in the table represent the number of times a symbol transitions from a row to a column. In the above symbol sequence, the statistically obtained symbol transition relationships include: C to C 18 times, C to B 7 times, C to D 12 times; B to C 15 times, B to B 3 times, B to D 2 times; D to C 9 times, D to B 5 times, D to D 6 times, and so on. The resulting 6×6 symbol transition matrix contains the number of transitions for each corresponding symbol pair in each cell.

[0072] Based on the number of transitions in the symbol transition matrix, calculate the conditional probability distribution of the symbol transitions. Conditional probability represents the probability of the next symbol appearing given that the current symbol is known. For example, the conditional probability of symbol C appearing after symbol C is 18 / (18+7+12) = 0.486, the conditional probability of symbol B appearing after symbol C is 7 / (18+7+12) = 0.189, and the conditional probability of symbol D appearing after symbol C is 12 / (18+7+12) = 0.324. Similarly, calculate the conditional probabilities of all symbol pairs to form a conditional probability matrix.

[0073] Conditional entropy is calculated based on the conditional probability distribution, and it measures the degree of uncertainty in a symbol sequence. The method for calculating conditional entropy is to calculate the entropy value of the next symbol for each current symbol, and then take a weighted average according to the probability of the current symbol appearing. A higher entropy value indicates greater uncertainty in the sequence; a lower entropy value indicates stronger regularity in the sequence. For the above symbol sequence, the calculated conditional entropy value is 1.42 bits.

[0074] Joint entropy is calculated based on the joint occurrence probability of symbols. Joint entropy measures the degree of uncertainty regarding the co-occurrence of two adjacent symbols. The method for calculating joint entropy is to take the weighted sum of the negative logarithms of the joint probabilities of all possible symbol pairs. In the above symbol sequence, the joint probability of symbol pair (C, C) is 18 / 149 = 0.121, the joint probability of symbol pair (C, B) is 7 / 149 = 0.047, and so on. The calculated joint entropy value is 2.26 bits.

[0075] The response pattern complexity is determined by the ratio of conditional entropy to joint entropy. A complexity value closer to 1 indicates a more random response pattern, while a complexity value closer to 0 indicates a more regular response pattern. The response pattern complexity of the above symbol sequence is 1.42 / 2.26 = 0.628.

[0076] In practical applications, response pattern complexity varies over time. For example, in a 30-minute tactile screening process for diabetic foot, the screening was divided into three time periods: pre-screening, mid-screening, and post-screening, each containing approximately 50 data points. The calculated response pattern complexities for each time period were 0.731, 0.682, and 0.458, respectively. It is evident that the response pattern complexity gradually decreases as the screening time increases, indicating that the subjects' response patterns become increasingly regular.

[0077] When the response pattern complexity is lower than a preset complexity threshold, the original data within the corresponding time period is marked as speculative contaminated data. If the preset complexity threshold is set to 0.5, then the data within the third time period of the above screening process is marked as speculative contaminated data because the response pattern complexity of 0.458 for that time period is lower than the preset threshold of 0.5.

[0078] Further analysis of the characteristics of the guessed contamination data revealed a clear periodic pattern in the response delay sequences of these data points. For example, within the time period marked as guessed contamination, the response delay sequences frequently exhibited repetitive patterns such as {B, C, D, B, C, D, ...} or {C, C, D, C, C, D, ...}, indicating that the subjects may have formed an expectation of the stimulus sequence and responded through guessing rather than actual perception.

[0079] The distribution of guessed contamination data across different areas of the foot also exhibits specific patterns. In a typical diabetic foot tactile screening, data points marked as guessed contamination accounted for as much as 65% in the dorsum of the foot in the later time period, while only 23% were in the heel. This uneven distribution suggests that the subjects developed a strong expected pattern for the stimulus sequence in the dorsum of the foot, while their perception of the heel remained relatively realistic.

[0080] The identified speculative contamination data was removed from the original dataset, resulting in a purified dataset. In the above sample, after removing the speculative contamination data, approximately 112 data points remained from the original 150 data points. The response pattern complexity of these data points increased to 0.715, which is closer to a random response pattern and can more accurately reflect the true tactile perception ability of the tested subject.

[0081] A method using response pattern complexity analysis to identify speculative contamination data improves the reliability of tactile screening data for diabetic foot. The dataset purified by this method removes spurious responses generated by the subjects' expectations or guesses, retaining data that truly reflects neurosensory function. This provides more accurate data support for the assessment of diabetic foot neuropathy, thereby improving the accuracy of early warning of diabetic foot ulcers and reducing the risk of misdiagnosis and missed diagnosis.

[0082] In one alternative implementation, spatial biases in the purified dataset are mapped to a foot neural atlas to construct a weighted graph. Lesion paths connecting aberrant nodes are extracted using a graph cut algorithm and depth-first traversal to generate a lesion path graph, including: The location coordinates and spatial deviation values ​​of each tactile stimulus are extracted from the cleaned dataset. Each location coordinate is mapped to a neural node in a preset foot neural atlas. The corresponding spatial deviation value is used as the node attribute value. The degree of deviation of each node attribute value from the preset normal threshold is calculated. Neural nodes whose deviation exceeds the preset abnormal judgment threshold are marked as abnormal nodes. The neural connection weights between adjacent nerve nodes in the foot neural atlas are calculated. The weighted combination of the spatial deviation difference of each nerve node and the neural anatomical distance is used as the edge weight. A weighted graph containing all nerve nodes and edge weights is constructed. A node affinity matrix is ​​constructed and a graph cut algorithm is performed on the weighted graph to segment the abnormal node subgraph and the normal node subgraph. The boundary nodes of the abnormal node subgraph are extracted. A depth-first traversal is performed based on the boundary nodes to search for the set of paths connecting each abnormal node. The joint score of the cumulative edge weight of each path and the degree of node deviation is calculated. The path with the highest joint score is selected as the lesion path. The neural node sequence and edge weight sequence in the lesion path are visualized and mapped to generate a lesion path graph.

[0083] In one specific implementation, extracting the position coordinates and spatial deviation values ​​of each tactile stimulus from the purified dataset is the initial step in constructing the pathogenesis map. Position coordinates are represented as (x, y) coordinate pairs on a two-dimensional plane, in millimeters, with the origin set at the center of the heel of the foot model. Spatial deviation values ​​represent the Euclidean distance between the actual perceived position and the stimulus position, also in millimeters. In a purified dataset containing 85 data points, the extracted position coordinates and corresponding spatial deviation values ​​are as follows: spatial deviation of 12.3 mm for the big toe ball region (78, 162), 8.7 mm for the third toe ball region (65, 148), 15.6 mm for the medial arch (42, 95), 6.2 mm for the lateral heel (18, 15), and 19.8 mm for the mid-dorsum of the foot (58, 120).

[0084] The location coordinates are mapped to nerve nodes in a pre-defined foot neural atlas. This atlas is a topological network constructed based on the anatomy of the foot's nerves, containing 32 nerve nodes, each corresponding to a key point in the nerve distribution of different regions of the foot. Each nerve node has a specific anatomical location, represented as a coordinate point on a standard foot model. The mapping process uses the nearest neighbor rule, assigning the actual stimulus location to the nearest nerve node. For example, location coordinates (78, 162) are mapped to node N1 (the big toe ball node, coordinates (80, 160)) in the neural atlas; location coordinates (65, 148) are mapped to node N4 (the third toe ball node, coordinates (63, 150)); and location coordinates (42, 95) are mapped to node N13 (the medial arch node, coordinates (45, 98)).

[0085] Using the corresponding spatial deviation value as the node attribute value, the deviation of each node attribute value from the preset normal threshold is calculated. The preset normal threshold is determined based on tactile perception data of healthy individuals. The normal threshold varies for different areas of the foot: 8 mm for the big toe ball area, 10 mm for the toe ball area, 12 mm for the arch area, 9 mm for the heel area, and 11 mm for the dorsum of the foot area. The deviation is calculated as the difference between the spatial deviation value and the normal threshold, divided by the normal threshold, and then multiplied by 100%, expressed as a percentage. The deviations for the above sample data are: N1 node 53.8%, N4 node -13.0%, N13 node 30.0%, N25 node -31.1%, and N18 node 80.0%. A positive deviation indicates that the spatial deviation exceeds the normal threshold, resulting in decreased perception accuracy; a negative deviation indicates that the spatial deviation is below the normal threshold, resulting in normal or even better perception accuracy.

[0086] Neural nodes whose deviation exceeds a preset anomaly threshold are marked as anomalous nodes. The preset anomaly threshold is set to 25%, meaning that when the spatial deviation exceeds 25% of the normal threshold, the node is considered to have a neural perceptual abnormality. In the example above, the deviations of nodes N1, N13, and N18 are 53.8%, 30.0%, and 80.0%, respectively, all exceeding the 25% anomaly threshold; therefore, these three nodes are marked as anomalous nodes. The deviations of nodes N4 and N25 are -13.0% and -31.1%, respectively, which do not exceed the anomaly threshold, and they remain marked as normal nodes.

[0087] The neural connection weights between adjacent neural nodes in the foot neural atlas are calculated. The neural connection weights consider two factors: spatial deviation difference and neuroanatomical distance. The spatial deviation difference represents the degree of difference in perceptual accuracy between adjacent nodes and is calculated as the absolute difference in spatial deviation values ​​between the two nodes; the neuroanatomical distance represents the length of the neuroanatomical connection between the two nodes, in millimeters. The weighted combination of these two factors uses a linear weighting method, with weighting coefficients of 0.7 and 0.3, respectively. For example, if the spatial deviation difference between node N1 and its adjacent node N2 is |12.3-9.5|=2.8 mm, and the neuroanatomical distance is 25 mm, then the neural connection weight is calculated as 2.8×0.7+25×0.3=9.46.

[0088] Construct a weighted graph containing all neural nodes and edge weights, and then construct a node affinity matrix. The node affinity matrix is ​​a 32×32 square matrix representing the affinity between any two neural nodes. For adjacent nodes, the affinity is set to the reciprocal of the neural connection weight; for non-adjacent nodes, the affinity is set to zero. For example, the affinity between nodes N1 and N2 is 1 / 9.46 = 0.106; nodes N1 and N18 are not adjacent, so their affinity is 0. The diagonal elements of the affinity matrix are set to the negative of the row sum to ensure that the row sum of the matrix is ​​zero, satisfying the property of a Laplace matrix.

[0089] The graph cut algorithm is applied to the weighted graph to separate the abnormal node subgraph from the normal node subgraph. Based on the minimum cut principle, the graph cut algorithm connects abnormal nodes to the source node (s) and normal nodes to the sink node (t) by introducing a source node (s) and a sink node (t). The edge weights are set to the degree of deviation of the nodes. The maximum flow minimum cut algorithm is then executed to find a cut set that divides the node set into two parts: one part connected to the source node (s) and containing abnormal nodes; the other part connected to the sink node (t) and containing normal nodes. In the example above, after the graph cut algorithm, the 32 neural nodes are divided into two subsets: the abnormal node subset contains 8 nodes {N1, N3, N7, N13, N14, N18, N19, N22}, and the normal node subset contains 24 nodes.

[0090] Extract the boundary nodes of the abnormal node subgraph. A boundary node is defined as an abnormal node that is adjacent to at least one normal node. In the above subset of abnormal nodes, nodes N1, N7, N13, and N22 are boundary nodes because they are adjacent to some nodes in the subset of normal nodes.

[0091] A depth-first search is performed on the boundary nodes to search for a set of paths connecting all the abnormal nodes. Starting from each boundary node, the subgraph of abnormal nodes is traversed using a depth-first search algorithm to find a path connecting all abnormal nodes. If a path covers all abnormal nodes, it is added to the path set. In complex cases, a path is allowed to contain a small number of normal nodes as relay points, but the proportion of abnormal nodes on the path must be no less than 80%. In the example above, starting from the boundary node N13, a path {N13→N14→N18→N19→N3→N1→N7→N22} is found, which covers all 8 abnormal nodes.

[0092] Calculate the joint score of the cumulative edge weight and node deviation for each path. The cumulative edge weight is the sum of the weights of all edges on the path; the node deviation is the weighted sum of the deviations of all nodes on the path, with the weights related to the node's position in the path, and nodes closer to the path's starting point having higher weights. The joint score is the weighted sum of the cumulative edge weight and the node deviation score, with weight coefficients of 0.4 and 0.6, respectively. In the example above, the cumulative edge weight of the path {N13→N14→N18→N19→N3→N1→N7→N22} is 72.3, the node deviation score is 42.8, and the joint score is 72.3×0.4+42.8×0.6=54.6.

[0093] The path with the highest joint score is selected as the lesion path. The neural node sequence and edge weight sequence in the lesion path are visualized and mapped to generate a lesion path diagram. The lesion path diagram is a visual representation that intuitively shows the spatial distribution characteristics of neuropathies, including the lesion origin, propagation path, and severity. In the lesion path diagram, each node is represented by a circle, and the size of the circle is proportional to the spatial deviation value of the node; the node color changes from green to red, indicating that the degree of deviation increases from low to high; the edges connecting the nodes are represented by line segments, and the thickness of the line segments is inversely proportional to the neural connection weight. For the above example, the generated lesion path diagram clearly shows the propagation path of the neuropathy starting from the medial arch of the foot (N13), passing through the middle of the dorsum of the foot (N18), and finally reaching the big toe ball region (N1) and the lateral heel (N22), intuitively showing the spatial distribution characteristics and severity of the neuropathy.

[0094] By using pathogenesis maps, doctors can intuitively understand the distribution patterns of foot neuropathy in diabetic patients, identify the starting points of lesions and high-risk areas, and develop targeted foot protection and treatment plans to prevent the occurrence of diabetic foot ulcers. Patients can also intuitively understand their own foot neuropathy status through pathogenesis maps, improving their self-protection awareness and compliance, and reducing the risk of diabetic foot complications.

[0095] like Figure 2 As shown, a flowchart of lesion path analysis is presented.

[0096] In one alternative implementation, a node affinity matrix is ​​constructed, and a graph cut algorithm is performed on the weighted graph to segment the abnormal node subgraph from the normal node subgraph. Extracting the boundary nodes of the abnormal node subgraph includes: The spatial deviation values ​​of each neural node in the weighted graph are normalized. The attribute similarity between adjacent neural nodes is calculated based on the normalized spatial deviation values. A node affinity matrix is ​​constructed based on the attribute similarity and edge weights. The attribute similarity is used as the connection strength between nodes. In a weighted graph, a source node and a sink node are defined. All abnormal nodes are connected to the source node and assigned connection weights based on the node affinity matrix. All normal nodes are connected to the sink node and assigned connection weights based on the node affinity matrix. Capacity allocation is performed on each edge in the weighted graph. Maximum flow minimum cut calculation is performed to determine the edge cut set that maximizes the total capacity from the source node to the sink node. The edge cut set divides the weighted graph into an abnormal node subgraph containing the source node and a normal node subgraph containing the sink node. Extract the boundary edges connecting the abnormal node subgraph and the normal node subgraph in the edge cut set, and identify the neural nodes in the abnormal node subgraph connected by the boundary edges as boundary nodes.

[0097] In one specific implementation, normalizing the spatial deviation values ​​of each neural node in the weighted graph is a preprocessing step of the graph cut algorithm. Normalization converts spatial deviation values ​​of different dimensions and ranges into values ​​of a uniform scale, facilitating subsequent calculations and comparisons. The normalization method uses maximum-minimum normalization, mapping the spatial deviation values ​​of all neural nodes to between 0 and 1. In a weighted graph containing 32 neural nodes, the original spatial deviation values ​​range from 5.2 mm to 24.7 mm. The minimum deviation value of 5.2 mm appears at node N26 in the central heel region, and the maximum deviation value of 24.7 mm appears at node N19 in the lateral dorsal region. After normalization, the normalized spatial deviation value of node N26 is 0, the normalized spatial deviation value of node N19 is 1, and the normalized values ​​of the remaining nodes are distributed between 0 and 1. For example, the original spatial deviation of node N1 in the big toe region of the foot is 12.3 mm, which is 0.36 after normalization; the original spatial deviation of node N13 in the medial arch region is 15.6 mm, which is 0.53 after normalization.

[0098] The attribute similarity between adjacent neural nodes is calculated based on the normalized spatial deviation values. Attribute similarity measures the degree of similarity between adjacent nodes in terms of spatial deviation features; a higher value indicates greater similarity between the two nodes. Attribute similarity is calculated using a Gaussian similarity function. The input is the square of the difference between the normalized spatial deviation values ​​of the two nodes, and the output is a similarity value between 0 and 1. The scaling parameter of the Gaussian similarity function is set to 0.15, which controls the rate of similarity decrease. In the actual calculation, the normalized spatial deviation values ​​of adjacent nodes N1 and N2 are 0.36 and 0.22, respectively, with a squared difference of 0.0196. Substituting these values ​​into the Gaussian similarity function yields an attribute similarity of 0.878, indicating that these two nodes are relatively similar in terms of spatial deviation characteristics. The normalized spatial deviation values ​​of adjacent nodes N13 and N14 are 0.53 and 0.68, respectively, with a squared difference of 0.0225. Substituting these values ​​into the Gaussian similarity function yields an attribute similarity of 0.861, also indicating a high similarity. However, the normalized spatial deviation values ​​of adjacent nodes N18 and N20 are 0.75 and 0.31, respectively, with a squared difference of 0.1936. Substituting these values ​​into the Gaussian similarity function yields an attribute similarity of only 0.275, indicating that these two nodes differ significantly in terms of spatial deviation characteristics.

[0099] A node affinity matrix is ​​constructed based on attribute similarity and edge weights, with attribute similarity representing the connection strength between nodes. The node affinity matrix is ​​a 32×32 symmetric matrix representing the affinity relationship between any two neural nodes. For adjacent nodes, the affinity is set to the product of attribute similarity and the reciprocal of the edge weight; for non-adjacent nodes, the affinity is set to zero. For example, nodes N1 and N2 have an edge weight of 9.46 and an attribute similarity of 0.878, resulting in an affinity of 0.878 × (1 / 9.46) = 0.0928; nodes N13 and N14 have an edge weight of 11.25 and an attribute similarity of 0.861, resulting in an affinity of 0.861 × (1 / 11.25) = 0.0765; while nodes N1 and N13 are not adjacent, so their affinity is directly set to 0. The diagonal elements of the affinity matrix are set to the negative of the row sum to ensure that the sum of each row is zero, satisfying the property of the Laplace matrix.

[0100] In a weighted graph, source and sink nodes are defined, denoted as 's' and 't' respectively. Source and sink nodes are special nodes in graph cut algorithms used to divide a graph into two disjoint subgraphs. All aberrant nodes are connected to the source node, and connection weights are assigned based on the node affinity matrix. The connection weight is set to the normalized spatial deviation of the aberrant node; a larger value indicates a more aberrant node and a stronger connection to the source node. In the example above, the set of aberrant nodes includes {N1, N3, N7, N13, N14, N18, N19, N22}, with normalized spatial deviations of {0.36, 0.41, 0.48, 0.53, 0.68, 0.75, 1.00, 0.47} respectively. All normal nodes are connected to the sink node, and connection weights are assigned based on the node affinity matrix. The connection weight is set to 1 minus the normalized spatial deviation of the normal node; a larger value indicates a more normal node and a stronger connection to the sink node. In the example above, the normal node set includes the remaining 24 neural nodes, and the connection weights of these nodes are calculated according to the rules described above.

[0101] The edges in the weighted graph are assigned capacity. The capacity of an edge represents the maximum flow that the edge can transmit, and is used to limit the size of the flow in the graph cut algorithm. For the edge connecting the source node s and the abnormal node, the capacity is set to the normalized spatial deviation value of the abnormal node; for the edge connecting the normal node and the sink node t, the capacity is set to 1 minus the normalized spatial deviation value of the normal node; for the edges between neurons, the capacity is set to the corresponding affinity value in the node affinity matrix. For example, the edge capacity between the source node s and the abnormal node N19 is 1.00; the edge capacity between the normal node N26 and the sink node t is 1-0=1.00; and the edge capacity between neurons N1 and N2 is 0.0928.

[0102] Perform maximum flow minimum cut calculation to determine the set of edge cuts that maximizes the total capacity from the source node to the sink node. The maximum flow minimum cut algorithm iteratively calculates augmenting paths, gradually increasing the flow from the source node s to the sink node t until the maximum flow is reached. After reaching the maximum flow, the algorithm automatically finds the minimum cut set, which is the smallest set of edges that can cut all paths from the source node s to the sink node t. In the example above, the Ford-Fulkerson algorithm is used to perform maximum flow minimum cut calculation, and the final maximum flow is 4.35. The corresponding minimum cut set contains 15 edges, which divide the weighted graph into a subgraph containing abnormal nodes of the source node s and a subgraph containing normal nodes of the sink node t.

[0103] The edge cut set divides the weighted graph into an anomalous node subgraph containing the source node and a normal node subgraph containing the sink node. In the example above, the anomalous node subgraph contains the source node s and 8 anomalous nodes {N1, N3, N7, N13, N14, N18, N19, N22}; the normal node subgraph contains the sink node t and 24 normal nodes. This segmentation divides the neural nodes in the foot neural atlas into two categories: one category consists of neural nodes with spatial deviations, indicating decreased tactile perception in these areas; the other category consists of neural nodes with normal spatial deviations, indicating normal tactile perception in these areas.

[0104] Extract the boundary edges connecting the abnormal node subgraph and the normal node subgraph from the edge cut set. A boundary edge is an edge whose endpoint is in the abnormal node subgraph and its other endpoint is in the normal node subgraph. In the example above, the boundary edges in the edge cut set include four edges: {(N1, N5), (N7, N8), (N13, N12), and (N22, N25)}. These boundary edges represent the boundary between the neuropathy region and the normal region, and are potential paths for the spread of neuropathy.

[0105] The neural nodes in the aberrant node subgraph connected by the boundary edges are identified as boundary nodes. Boundary nodes are nodes adjacent to normal nodes in the aberrant node subgraph, representing the leading edge of the neuropathy spreading into the normal region. In the example above, there are four boundary nodes: {N1, N7, N13, N22}. These boundary nodes play a crucial role in subsequent lesion path searching because they act as bridges connecting the aberrant and normal regions, allowing the tracking of the neuropathy's spread path.

[0106] Boundary nodes hold particular clinical significance in the assessment of diabetic foot neuropathy. In clinical practice, the foot regions corresponding to these boundary nodes are often the forefront of neuropathy spreading from severely affected areas to less severely affected areas, and are also key locations for early intervention. For example, boundary node N1 is located in the big toe bulb region of the sole; abnormal nerve function in this area may lead to uneven weight-bearing on the forefoot, increasing the risk of ulcers. Boundary node N13 is located in the medial arch region; abnormal nerve function in this area may affect gait stability. Boundary node N22 is located in the lateral heel region; abnormal nerve function in this area may lead to a shift in the heel's weight-bearing point. Targeted protection and early intervention in the foot regions corresponding to these boundary nodes help to block the spread of neuropathy and reduce the risk of diabetic foot ulcers.

[0107] The portable diabetic foot tactile screening data management system of this invention includes: The parameter setting unit is used to acquire diabetes-related parameters of the subject and determine the set of detection locations and standard stimulus intensity. The stimulus execution unit is used to generate a sequence seed based on hash operation, randomly generate a stimulus sequence and a time offset sequence according to the sequence seed, combine them to construct a hidden stimulus matrix, and drive the sensor array to apply tactile stimulation according to the hidden stimulus matrix. The data acquisition unit is used to receive the sensing location and sensing intensity, calculate the spatial deviation, intensity deviation and response delay, and form the raw dataset; The data cleansing unit is used to construct trajectory point sequences from the spatial deviation sequences of the original dataset, label spatial memory attractors, identify memory contamination data, symbolically encode the response delay sequences of the original dataset, identify guess contamination data, remove memory contamination data and guess contamination data, and form a cleaned dataset. The path generation unit is used to map the spatial bias in the purified dataset to the foot neural atlas to construct a weighted graph. Based on the graph cut algorithm and depth-first traversal, the lesion path connecting abnormal nodes is extracted to generate a lesion path graph. An accelerated warning unit is used to store the first detected clean dataset as a baseline dataset, calculate the second derivative of the rate of change of the difference vector magnitude between the subsequent clean dataset and the baseline dataset, and generate an accelerated warning signal when the second derivative is positive. The data upload unit is used to upload lesion path maps and accelerated early warning signals.

[0108] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0109] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0110] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A portable method for managing tactile screening data of diabetic foot, characterized in that, include: Obtain diabetes-related parameters from the subjects and determine the set of detection locations and standard stimulus intensity; A sequence seed is generated based on hash operation. A stimulus sequence and a time offset sequence are randomly generated based on the sequence seed. These are combined to construct a hidden stimulus matrix. The sensor array is then driven to apply tactile stimuli according to the hidden stimulus matrix. Receive the sensing location and sensing intensity, calculate spatial deviation, intensity deviation and response delay to form the raw dataset; A trajectory point sequence is constructed from the spatial deviation sequence of the original dataset, spatial memory attractors are labeled, memory contamination data is identified, the response delay sequence of the original dataset is symbolically encoded, guess contamination data is identified, memory contamination data and guess contamination data are removed to form a clean dataset. Spatial biases in the purified dataset are mapped to foot neural atlases to construct weighted graphs. Based on graph cut algorithm and depth-first traversal, lesion paths connecting abnormal nodes are extracted to generate lesion path graphs. The first detected clean dataset is stored as a baseline dataset. The second derivative of the rate of change of the difference vector magnitude between the subsequent clean datasets and the baseline dataset is calculated. When the second derivative is positive, an acceleration warning signal is generated. Upload the lesion path map and the accelerated early warning signal.

2. The method according to claim 1, characterized in that, A sequence seed is generated based on hash operations. A stimulus sequence and a time offset sequence are randomly generated based on the sequence seed, and these are combined to construct a hidden stimulus matrix. Tactile stimuli are then applied to the sensor array according to the hidden stimulus matrix, including: The current detection timestamp and the detected person's identifier are concatenated to form an initial string. A hash operation is performed on the initial string to generate a sequence seed. A pseudo-random number generator is initialized based on the sequence seed to obtain a random number stream. Extract a random floating-point number sequence equal to the number of detection positions from the random number stream, normalize the random floating-point number sequence, map it to each detection position index, sort the mapping results using a sorting algorithm to generate a stimulus order sequence, and simultaneously extract a second set of random numbers from the random number stream and generate a non-uniform time offset sequence through exponential distribution transformation. A hidden stimulus matrix is ​​constructed by combining the stimulus sequence and the time offset sequence. The sensor array is driven to apply tactile stimulation according to the spatiotemporal coordinates of the hidden stimulus matrix. Only the foot outline is displayed on the interactive interface to mask the actual activation position.

3. The method according to claim 2, characterized in that, The second set of random numbers is extracted from the random number stream, and a non-uniform time offset sequence is generated by exponential distribution transformation, including: Extract a second set of uniformly distributed random numbers from the random number stream, perform a negative logarithmic transformation on each random value in the second set of random numbers to obtain the original time offset of the exponential distribution, and form the original offset sequence. The expected reaction time window of the subject is calculated based on the nerve conduction velocity parameter in the diabetes-related parameters. The ratio of the expected reaction time window to the preset health baseline reaction time is used as the time scale scaling factor to proportionally adjust each original time offset in the original offset sequence. Set upper and lower boundary thresholds for the adjusted offset sequence, and truncate the offsets that exceed the upper and lower boundary thresholds to the boundary values ​​to generate a non-uniform time offset sequence that meets the physiological perception range.

4. The method according to claim 1, characterized in that, Construct trajectory point sequences from the spatial deviation sequences of the original dataset, label spatial memory attractors, and identify memory-contaminated data, including: Spatial deviation sequence is extracted from the original dataset. The deviation values ​​at consecutive time points in the spatial deviation sequence are combined to construct a multidimensional vector. Each multidimensional vector is used as a trajectory point in the phase space to form a trajectory point sequence. Density clustering is performed on the trajectory point sequence to obtain cluster centers. The number of times each trajectory point returns to the preset neighborhood of the cluster center is calculated. The ratio of the number of returns to the total number of trajectory points is calculated to determine the trajectory return rate. When the trajectory return rate exceeds a preset return threshold, the cluster center is marked as a spatial memory attractor. Calculate the distance from each trajectory point to the spatial memory attractor, count the density peak points of the distance distribution, delineate the boundary of the attraction domain using the distance values ​​corresponding to the density peak points, and mark the data in the original dataset whose corresponding trajectory points fall within the boundary of the attraction domain as memory contamination data.

5. The method according to claim 1, characterized in that, The response delay sequences of the original dataset are symbolically encoded to identify suspected contamination data, including: Response delay sequences are extracted from the original dataset. Multiple interval thresholds are set according to the numerical range of response delay values. Each response delay value in the response delay sequence is compared with the interval thresholds, and each response delay value is mapped to the corresponding symbol category to form a symbol sequence. The transition relationships between adjacent symbols in the symbol sequence are statistically analyzed, the number of transitions between each symbol pair is recorded, a symbol transition matrix is ​​constructed, the conditional probability distribution of symbol transitions is calculated based on the number of transitions in the symbol transition matrix, the conditional entropy is calculated based on the conditional probability distribution, the joint entropy is calculated based on the joint occurrence probability of symbols, and the response mode complexity is determined based on the ratio of the conditional entropy to the joint entropy. When the complexity of the response pattern is lower than a preset complexity threshold, the original data within the corresponding time period is marked as speculative contaminated data.

6. The method according to claim 1, characterized in that, Spatial biases in the purified dataset are mapped to foot neural atlases to construct a weighted graph. Based on graph cut and depth-first traversal algorithms, lesion paths connecting aberrant nodes are extracted, generating a lesion path graph including: The location coordinates and spatial deviation values ​​of each tactile stimulus are extracted from the cleaned dataset. Each location coordinate is mapped to a neural node in a preset foot neural atlas. The corresponding spatial deviation value is used as the node attribute value. The degree of deviation of each node attribute value from the preset normal threshold is calculated. Neural nodes whose deviation exceeds the preset abnormal judgment threshold are marked as abnormal nodes. The neural connection weights between adjacent nerve nodes in the foot neural atlas are calculated. The weighted combination of the spatial deviation difference of each nerve node and the neural anatomical distance is used as the edge weight. A weighted graph containing all nerve nodes and edge weights is constructed. A node affinity matrix is ​​constructed and a graph cut algorithm is performed on the weighted graph to segment the abnormal node subgraph and the normal node subgraph. The boundary nodes of the abnormal node subgraph are extracted. A depth-first traversal is performed based on the boundary nodes to search for the set of paths connecting each abnormal node. The joint score of the cumulative edge weight of each path and the degree of node deviation is calculated. The path with the highest joint score is selected as the lesion path. The neural node sequence and edge weight sequence in the lesion path are visualized and mapped to generate a lesion path graph.

7. The method according to claim 6, characterized in that, Construct a node affinity matrix and perform a graph cut algorithm on the weighted graph to separate the abnormal node subgraph from the normal node subgraph. Extract the boundary nodes of the abnormal node subgraph, including: The spatial deviation values ​​of each neural node in the weighted graph are normalized. The attribute similarity between adjacent neural nodes is calculated based on the normalized spatial deviation values. A node affinity matrix is ​​constructed based on the attribute similarity and edge weights. The attribute similarity is used as the connection strength between nodes. In a weighted graph, a source node and a sink node are defined. All abnormal nodes are connected to the source node and assigned connection weights based on the node affinity matrix. All normal nodes are connected to the sink node and assigned connection weights based on the node affinity matrix. Capacity allocation is performed on each edge in the weighted graph. Maximum flow minimum cut calculation is performed to determine the edge cut set that maximizes the total capacity from the source node to the sink node. The edge cut set divides the weighted graph into an abnormal node subgraph containing the source node and a normal node subgraph containing the sink node. Extract the boundary edges connecting the abnormal node subgraph and the normal node subgraph in the edge cut set, and identify the neural nodes in the abnormal node subgraph connected by the boundary edges as boundary nodes.

8. A portable diabetic foot tactile screening data management system, used to implement the method of any one of claims 1-7, characterized in that, include: The parameter setting unit is used to acquire diabetes-related parameters of the subject and determine the set of detection locations and standard stimulus intensity. The stimulus execution unit is used to generate a sequence seed based on hash operation, randomly generate a stimulus sequence and a time offset sequence according to the sequence seed, combine them to construct a hidden stimulus matrix, and drive the sensor array to apply tactile stimulation according to the hidden stimulus matrix. The data acquisition unit is used to receive the sensing location and sensing intensity, calculate the spatial deviation, intensity deviation and response delay, and form the raw dataset; The data cleansing unit is used to construct trajectory point sequences from the spatial deviation sequences of the original dataset, label spatial memory attractors, identify memory contamination data, symbolically encode the response delay sequences of the original dataset, identify guess contamination data, remove memory contamination data and guess contamination data, and form a cleaned dataset. The path generation unit is used to map the spatial bias in the purified dataset to the foot neural atlas to construct a weighted graph. Based on the graph cut algorithm and depth-first traversal, the lesion path connecting abnormal nodes is extracted to generate a lesion path graph. An accelerated warning unit is used to store the first detected clean dataset as a baseline dataset, calculate the second derivative of the rate of change of the difference vector magnitude between the subsequent clean dataset and the baseline dataset, and generate an accelerated warning signal when the second derivative is positive. The data upload unit is used to upload lesion path maps and accelerated early warning signals.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images