Neuropathy and vasculopathy detection method and system based on wearable device

Abstract

The invention belongs to the technical field of lesion detection methods, and particularly relates to a neuropathy and vasculopathy detection method and system based on a wearable device, a pressure sensing unit is calibrated by a standard weight to generate a voltage and pressure corresponding curve, a temperature sensing unit corrects conversion errors in standard environments of 32 DEG C, 34 DEG C and 36 DEG C, and the temperature sensing unit detects the neuropathy and vasculopathy. The vibration sense unit calibrates stimulation parameters through an acceleration sensor, the pressure sensing unit captures plantar dynamic pressure signals, the temperature sensing unit collects the area temperature and the temperature difference between the left foot and the right foot, the vibration sense unit applies stimulation according to a preset scheme and collects sensing response, and the PPG sensing unit collects the original waveform of microcirculation blood flow.

Description

Technical Field

[0001] This invention belongs to the field of lesion detection technology, and particularly relates to a method and system for detecting neurolesions and vascular lesions based on wearable devices. Background Technology

[0002] Currently, clinical detection methods for neuropathy and vascular lesions largely rely on specialized medical equipment and operations by healthcare professionals, such as nerve conduction velocity testing and ankle-brachial index measurement. While these methods provide accurate results, they have significant limitations: the testing scenarios are concentrated in hospitals, making it impossible to achieve daily dynamic monitoring and capture real-time physiological signals such as changes in plantar pressure and microcirculatory blood flow fluctuations, easily missing the insidious characteristics of early lesions; at the same time, traditional tests are mostly single-point and single-modal, unable to integrate multi-dimensional data such as pressure, temperature, vibration sensation, and blood flow, making it difficult to comprehensively reflect the synergistic pathological mechanisms of nerves and blood vessels, resulting in a low detection rate of early lesions. Patients often miss the best intervention opportunity because they only seek medical attention when symptoms are obvious.

[0003] The application of existing wearable devices in lesion detection still has shortcomings: most devices only monitor single physiological indicators, lacking the ability to simultaneously collect and fuse multimodal data, and thus failing to establish correlations between different indicators and pathological states; moreover, data processing is mostly limited to basic feature extraction, failing to fully explore the synergistic abnormal patterns of features in the spatiotemporal dimensions, resulting in insufficient accuracy in lesion grading and risk scoring. In addition, the early warning mechanisms of existing devices are mostly based on fixed thresholds, failing to consider individual baseline differences and lesion development trends, easily leading to false or missed warnings, and lacking targeted intervention suggestion push functions, making it difficult to form a closed-loop management of detection-assessment-early warning-intervention, and failing to meet the dual needs of patients for home monitoring and precise clinical diagnosis and treatment. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned technical problems by providing a method and system for detecting neuropathic and vascular lesions based on wearable devices.

[0005] In view of this, the present invention provides a method for detecting neuropathic and vascular lesions based on a wearable device, comprising the following steps: Step S110: The pressure sensing unit is calibrated using standard weights to generate voltage and pressure curves; the temperature sensing unit is corrected for conversion errors in standard environments of 32℃, 34℃, and 36℃; and the vibration sensing unit is calibrated for stimulation parameters using an accelerometer. Step S120: The pressure sensing unit captures dynamic pressure signals on the sole of the foot, the temperature sensing unit collects the temperature of the area and the temperature difference between the left and right feet, the vibration sensing unit applies stimulation according to the preset scheme and collects the sensory response, and the PPG sensing unit collects the original waveform of microcirculation blood flow. Step S130: Low-pass filtering and median filtering are used to denoise the pressure and blood flow signals, and mean filtering and environmental compensation are used for the temperature signal. Multimodal data are synchronized by timestamps, and misaligned and abnormal data are removed. Step S140: Extract peak value, duration and contact area features from pressure data; extract temperature difference and heat zone distribution features from temperature data; extract sensing pass rate, average reaction time and sensing threshold level from vibration response data; extract pulse wave period, amplitude and relative blood flow from blood flow data. Step S150: Construct a spatiotemporally unified multimodal data matrix, enhance features through convolution enhancement, temperature gradient calculation, perception, stimulus fitting and waveform entropy analysis, and establish a modality and pathology association network to generate cross-modal fusion features; Step S160: Using a dual-branch network combined with a multi-task learning model, the neuropathic grading and vascular lesion risk score are output simultaneously; Step S170: Based on the assessment results and trend analysis, trigger a Level 1, Level 2, or Level 3 early warning and push corresponding intervention suggestions.

[0006] Preferably, the pressure sensing unit generates a voltage-pressure correspondence curve through standard weight calibration, the temperature sensing unit corrects conversion errors in standard environments of 32℃, 34℃, and 36℃, and the vibration sensing unit calibrates stimulation parameters through an accelerometer, including: Standard weights of 10g, 20g, and 30g were used to determine the locations of 10 nerve lesion detection points on the sole of the wearable device. According to the order of the detection points, each standard weight was placed steadily on the pressure sensing unit of each detection point. The pressure state of each weight was kept stable for 3-5 seconds, and the voltage output value of each detection point under different weight pressures was recorded simultaneously. After all detection points have completed the application of different weights and data recording, the voltage-pressure correspondence curve of each detection point was generated based on the collected voltage and pressure data through linear fitting or interpolation algorithms to complete the calibration of the pressure sensing unit. Set up standard constant temperature environments of 32℃, 34℃, and 36℃, and place the temperature sensing unit entirely in these environments. After the temperature sensing unit stabilizes in each standard constant temperature environment for 5-10 minutes to ensure that its detection value tends to stabilize, record the actual resistance output value of the sensing unit at each standard temperature. Based on the deviation between the standard temperature value and the corresponding measured resistance value, adjust the coefficient in the temperature-resistance conversion formula to correct the conversion error, so that the resistance output of the sensing unit at 32℃, 34℃, and 36℃ can be accurately mapped to the corresponding standard temperature. At the same time, collect the current ambient temperature in the calibration environment and record it as the baseline value for temperature data compensation. The calibration procedure of the vibration sensor unit is initiated, activating the built-in accelerometer. The vibration motor is controlled to run according to preset parameters. The actual vibration parameters of the motor are collected in real time through the accelerometer. The collected actual vibration parameters are compared with the preset parameters, and the deviation value is calculated. If the deviation value exceeds ±5%, the drive current of the motor is adjusted through the PWM signal. After each adjustment, the actual vibration parameters are collected and compared with the preset parameters again until the deviation value between the actual vibration parameters and the preset parameters is ≤ ±5%. The adjustment is then stopped, and the calibration of the vibration sensor unit stimulation parameters is completed.

[0007] Preferably, the pressure sensing unit captures dynamic pressure signals from the sole of the foot, the temperature sensing unit collects the regional temperature and the temperature difference between the left and right feet, the vibration sensing unit applies stimulation according to a preset scheme and collects the sensory response, and the PPG sensing unit collects the raw waveform of microcirculatory blood flow, including: The flexible array pressure sensing unit is activated, and the basic sampling frequency is set to 50Hz. The pressure value of each sensing unit is recorded every 10ms to form a time and pressure matrix. During the acquisition process, the foot contact area at each sampling moment is marked in real time, and the time of pressure peak occurrence, duration of continuous force, and contact area are recorded synchronously. If the pressure value of a certain area suddenly exceeds 200kPa, the sampling frequency of the pressure sensing unit in that area is automatically increased to 100Hz until the pressure value falls back below the threshold and the basic frequency is restored, continuously capturing changes in dynamic pressure signals of the foot. Activate the contact-type NTC thermistor temperature sensing unit, set the basic sampling interval to 10 seconds / time, maintain a signal stabilization time of 0.5 seconds during each acquisition to avoid interference from instantaneous temperature fluctuations, and synchronously record the temperature values ​​of the corresponding detection points of the left and right feet. After the acquisition is completed, calculate the temperature difference between the symmetrical detection points of the left and right feet to form a dataset of detection points and temperature differences. If the temperature of a certain area exceeds the normal range of 32-36℃, the sampling interval of the temperature sensing unit in that area is automatically shortened to 5 seconds / time, and the area is marked as a hot zone and the distribution of the hot zone is recorded. The area temperature and the temperature difference between the left and right feet are continuously collected. The system calls upon the built-in miniature columnar vibration motor of the vibration sensor unit, loads a preset stimulation program, and activates the vibration motors at each detection point in the order of toes, forefoot, and heel, applying standardized vibration stimulation. After stimulation is initiated, the system receives active feedback from the patient through the accompanying mobile app, while simultaneously capturing passive feedback from the patient's foot through a miniature accelerometer inside the shoe. The system records whether each detection point is perceived and the perception delay time. If the same detection point is perceived negatively for three consecutive rounds of stimulation, the system automatically switches to a higher intensity stimulation program and re-executes the stimulation and perception response acquisition at that detection point. The near-infrared photoplethysmography (PPG) sensor unit integrated into the upper inner side of the shoe upper is activated. The initial LED drive current is set to 50mA, the sampling frequency to 100Hz, and the ADC conversion resolution to 16-bit. The near-infrared LED light source continuously emits a stable light signal. After penetrating the skin tissue of the foot, the photodiode captures the changes in the intensity of the reflected light in real time and converts the light signal into a weak current signal. After the signal is amplified by the preamplifier circuit, it is initially suppressed by the 10Hz low-pass filter circuit to suppress ambient light and motion interference. Then, the analog signal is converted into a digital signal by the ADC module to form the original waveform data of microcirculation blood flow. During the acquisition process, the signal amplitude is monitored in real time. If the amplitude is lower than 500 count units, the LED drive current is automatically increased. If there is still no effective signal, a poor fit prompt is pushed. At the same time, the filtering algorithm further suppresses motion interference to ensure the stability of the blood flow signal. The original waveform data is bound and stored with a timestamp, and the original waveform of microcirculation blood flow is continuously acquired.

[0008] Preferably, the step of using low-pass filtering and median filtering to denoise the pressure and blood flow signals, and mean filtering and environmental compensation to denoise the temperature signal, and synchronizing multimodal data through timestamps to remove misaligned and abnormal data, includes: A low-pass filter with a cutoff frequency of 50Hz is used to filter the raw plantar pressure signal to remove high-frequency motion noise generated by insole friction during walking. Then, a median filter with a window size of 5 sampling points is selected to further process the low-pass filtered pressure signal to remove pulse noise caused by stones pressing on the feet. If there are missing data in the processed signal due to sensor failure or signal loss, and the missing rate is less than 5%, the missing sampling points are supplemented by linear interpolation to ensure the integrity of the pressure signal. An adaptive filtering algorithm was employed to extract the motion signal synchronously acquired by the plantar pressure sensor as a reference signal. An adaptive filter was constructed to separate motion artifacts in the blood flow signal. Then, a low-pass filter with a cutoff frequency of 10Hz was applied to remove respiratory interference and baseline drift in the blood flow signal. Subsequently, the blood flow signal was decomposed into 5 layers using the db4 wavelet basis. Wavelet transform was used to remove electronic noise and high-frequency interference from sudden changes in ambient light, retaining the effective pulse wave frequency band of 0.5-10Hz. Finally, for the missing data in the processed signal, peak detection and interpolation were used to repair the missing data, resulting in a denoised blood flow signal. First, a moving average filter with a window size of 3 sampling values ​​is used to perform mean filtering on the acquired raw temperature signal to smooth out the interference caused by instantaneous temperature fluctuations. Then, the ambient temperature baseline value acquired during the system initialization phase is retrieved, and the deviation between the acquired temperature at each temperature detection point and the ambient temperature baseline value is calculated. According to the formula: actual temperature = acquired temperature - ambient temperature deviation, environmental compensation is performed on the mean-filtered temperature signal to eliminate the influence of ambient temperature changes on the measurement of the foot body temperature, and the corrected temperature signal is obtained. The clock synchronization unit of the in-shoe data acquisition module is activated to read the timestamps bound during the acquisition of four modal data: pressure, temperature, vibration response, and blood flow. Using the timestamp of the pressure signal as a reference, the timestamps of the temperature, vibration response, and blood flow signals are compared. The missing sampling points in each modal data are filled in by linear interpolation, so that the time axes of the four modal data are completely aligned, forming a spatiotemporally unified multimodal data matrix. By comparing the timestamps of each modality's data with the data acquisition order after synchronization, if the timestamp of a certain modality's data deviates from the timestamps of other modal data by more than 1ms, it is determined to be misaligned data caused by transmission delay, and the misaligned data is directly discarded. For pressure signals, extreme values ​​exceeding 300kPa are discarded; for temperature signals, abnormal values ​​deviating from the normal temperature range after environmental compensation by more than 5℃ are discarded; for blood flow signals, noise values ​​with pulse wave period fluctuations exceeding 20% ​​are discarded. If a sensor experiences a signal interruption and the data from adjacent sensor units cannot be supplemented by interpolation, or if the proportion of abnormal data in a certain modality exceeds 10%, the valid data already acquired for that modality is retained, and the abnormal data area is marked to avoid abnormal data affecting subsequent analysis.

[0009] Preferably, the step of extracting peak value, duration, and contact area features from pressure data, extracting temperature difference and thermal zone distribution features from temperature data, extracting sensing pass rate, average reaction time, and sensing threshold level from vibration response data, and extracting pulse wave period, amplitude, and relative blood flow value from blood flow data includes: A 1-minute time period for feature extraction is determined. All data collected by pressure sensing units in each detection area of ​​the sole within this period are iterated. For each detection area, the maximum value among all pressure values ​​is selected and recorded as the pressure peak value of the area. All sampling times in the area where the pressure value exceeds 50 kPa are counted, and the cumulative duration of these times is calculated and recorded as the pressure duration. The number of sensing units in the area where the pressure value is greater than 0 kPa is counted. Combined with the preset coverage area of ​​each sensing unit, the total coverage area of ​​all sensing units that meet the conditions is calculated and recorded as the contact area. The above operations are repeated to complete the feature extraction of pressure peak value, duration, and contact area for all detection areas of the sole. The temperature data collected from symmetrical detection points on both feet is iterated. For each pair of symmetrical detection points, the absolute value of the temperature difference within a 1-minute acquisition cycle is calculated to generate a dataset of detection points and temperature differences. The normal temperature range is set to 32-36℃. The temperature data of all temperature detection points are iterated, and the areas where the temperature exceeds this range are marked to form a thermal zone distribution record. At the same time, the temperature difference between different detection points on the same foot is calculated to analyze the temperature distribution differences within the same foot and supplement and improve the temperature feature extraction results. The vibration stimulation and feedback results of all detection points on the sole of the foot were statistically analyzed to determine the number of positive detection points and the total number of detection points. The perception pass rate was calculated using the formula (number of positive detection points / total number of detection points) × 100%. For all positive detection points, the perception delay time was extracted, and the arithmetic mean of these delay times was calculated to obtain the average reaction time. Based on the reaction time, the perception threshold levels were divided as follows: ≤0.5 seconds was Level 1 (normal), 0.5-1 second was Level 2 (mild reduction), 1-2 seconds was Level 3 (moderate reduction), 2-3 seconds was Level 4 (severe reduction), and >3 seconds or negative perception was Level 5 (extremely severe reduction). This completed the perception threshold level classification for each detection point. The preprocessed microcirculatory blood flow PPG raw waveform data were analyzed. The peak detection algorithm was used to identify the systolic peak, diastolic peak, and dicrotic wave in the waveform. The time interval between two adjacent S waves was calculated and recorded as the pulse wave period. The difference between the peak value of the S wave and the waveform baseline was measured and recorded as the pulse wave amplitude. Based on the preset tissue optical parameters and combined with the Lambert-Beer law, the area integral of the PPG waveform was calculated to derive the relative value of blood flow. At the same time, the amplitude ratio of the S wave to the D wave and the amplitude variation coefficient obtained by the sliding window method were calculated to supplement the dimensions of blood flow feature extraction.

[0010] Preferably, the construction of a spatiotemporally unified multimodal data matrix, through convolution enhancement, temperature gradient calculation, perception, stimulus fitting, and waveform entropy analysis to enhance features, and establishing a modality and pathological association network to generate cross-modal fusion features, includes: Based on the timestamps bound to each modal data, the missing sampling points are filled in using linear interpolation to ensure that the time axes of pressure, temperature, vibration response, and blood flow data are completely synchronized. Then, the sensor coordinate mapping table preset by the insole and upper is retrieved to establish the spatial correspondence between the plantar pressure sensing points, temperature detection points, vibration stimulation points and the microcirculation monitoring area of ​​the dorsum of the foot. The synchronized multimodal data are integrated according to the dimensions of time frame × spatial point × modal feature to form a spatiotemporally unified multimodal data matrix. For the pressure features in the data matrix, a 1×3 temporal convolution kernel is used to traverse the data of each time frame to extract the temporal trend of the pressure peak changing with time; then a 3×3 spatial convolution kernel is used to cover each spatial point to highlight the features of the abnormal pressure area that has been subjected to continuous force for more than 60 seconds; finally, an attention mechanism is introduced to give higher weight to the features of long-term pressure points to strengthen the expression of abnormal pressure features. Based on the temperature characteristics in the data matrix, the rate of change of temperature difference between the left and right foot symmetrical detection points over time is first calculated to reflect the development speed of temperature anomalies. Then, a heat map diffusion algorithm is used to analyze the spatial spread trend of the heat zone with the temperature exceeding 32-36℃ as the center, generate heat zone diffusion characteristics, and supplement dynamic anomaly information in addition to static temperature difference. The reaction time and corresponding stimulus intensity parameters of the vibration response data in the data matrix are extracted. With stimulus intensity as the independent variable and reaction time as the dependent variable, a quadratic function curve is fitted. The curve slope and inflection point are extracted by curve analysis to form the vibration perception enhancement feature. For PPG waveform data in the data matrix, the waveform is divided into segments according to a preset time window, and the information entropy of the PPG waveform in each window is calculated. By the difference in information entropy values, the disordered regions of the waveform are identified, and the information entropy value is used as a blood flow enhancement feature to supplement the deficiencies of traditional pulse wave amplitude and period features. Calculate the mutual information value between the enhancement features of each modality, and screen the highly correlated feature pairs with mutual information values ​​> 0.6; perform fusion operation on the highly correlated feature pairs: according to the feature product weighting and difference penalty strategy, that is, the correlated feature value = (feature A × feature B × adaptive correlation weight) - (|duration of feature A - stability duration of feature B| × penalty coefficient), where the weight and penalty coefficient are determined by training with clinical data; Using spatial point-time period-pathological correlation as a three-dimensional dimension, the fused feature pairs are classified according to spatial point and time period. The degree of multimodal feature synergy anomaly of each spatial point in different time periods is calculated. A three-dimensional matrix of modality-pathological correlation is constructed to quantify the level of synergy anomaly of cross-modal features and finally generate a cross-modal fused feature set.

[0011] Preferably, the method of employing a dual-branch network combined with a multi-task learning model to simultaneously output neuropathic grading and vascular lesion risk scores includes: A dual-branch network with an immediate feature branch and a temporal trend branch was constructed. The immediate feature branch adopted a CNN network, which input the cross-modal correlation features of the current acquisition cycle to extract the immediate pathological abnormalities such as the instantaneous pressure peak exceeding the standard and the sudden drop in blood flow. The temporal trend branch adopted a BiLSTM network, which input the historical feature sequences of the last 5 acquisition cycles to capture the disease development trend of the gradual increase in temperature difference and the continuous prolongation of vibration perception reaction time. Modal attention automatically assigns weights to each modality. When assessing neuropathology, it increases the weights of vibration and temperature modalities, and when assessing vascular lesions, it increases the weights of microcirculation and pressure modalities. Pathological attention is based on the clinical pathological association rule that the correlation between neuropathology and vibration and temperature is higher than that between neuropathology and pressure, thus strengthening the contribution of key pathological features. By fusing the outputs of the two branches through a fully connected layer, a preliminary multimodal fusion feature vector with a dimension of 1×128 is generated.

[0012] Preferably, the following criteria are used to construct a rule constraint matrix: a temperature difference of >2℃ between the left and right feet indicates inflammation or perfusion abnormality; a vibration sensation reaction time of >2 seconds indicates moderate neuropathy; and a blood flow of <1.5ml・min⁻¹・cm⁻² indicates insufficient blood supply. If the evaluation result corresponding to the initial multimodal fusion feature vector conflicts with the constraint rules, the feature reweighting mechanism is activated to increase the weight of the conflicting modal features and recalculate the fusion result. The normal state data collected during the patient's first wearing of the device is incorporated as individual baseline data. The current fusion features are compared with the individual baseline to generate an individual relative abnormality index. The neuropathy grading task uses the Softmax classifier to map the calibrated fusion features to 5 levels: level 0, level 1, level 2, level 3, and level 4. The classification boundary is optimized through training with 50 clinical samples. The vascular lesion risk scoring task uses a regression model to output a risk value of 0-100. The scoring dimensions cover blood flow adequacy, vascular elasticity, blood flow stability, and cross-modal association abnormality. The consistency between the neuropathic grading and the vascular lesion risk score is verified. If the correspondence between the two does not conform to clinical patterns, the model is recalculated.

[0013] The beneficial effects of this invention are: Through multi-dimensional technological innovation, breakthroughs have been achieved in the precision and comprehensiveness of neuropathic and vascular lesion detection. In the sensor calibration stage, standard weights, a constant temperature environment, and accelerometers are used to precisely calibrate the pressure, temperature, and vibration sensory units, ensuring the accuracy of raw data acquisition. Multimodal data is simultaneously acquired, including pressure, temperature, vibration response, and microcirculation blood flow signals. Combined with targeted denoising algorithms and a timestamp synchronization mechanism, interfering data is effectively eliminated, ensuring data quality. Deep pathological correlations are uncovered through feature enhancement techniques such as convolution enhancement and temperature gradient calculation. A dual-branch network and multi-task learning model are constructed, incorporating clinical rule constraints and individual baseline data, enabling the simultaneous output of neuropathic grading and vascular lesion risk scores. This solves the problems of missed and false positives caused by traditional single-modal detection and fixed threshold assessment.

[0014] Meanwhile, this invention constructs a closed-loop management system of "detection-assessment-early warning-intervention," significantly improving the practicality and accessibility of detection. The wearable device supports dual modes of daily home monitoring and clinical monitoring. The collected parameters can be dynamically adjusted according to abnormal data. The low-power design and local caching function ensure long-term stable use without relying on professional medical scenarios. The three-level early warning mechanism combines trend change, cross-modal, and predictive triggering logic with differentiated intervention suggestions, achieving accurate reminders through two-way push notifications between doctors and patients. The closed-loop optimization mechanism continuously updates the model through incremental learning, adapting to individual differences and complex clinical scenarios. This not only improves patient monitoring compliance but also provides early warning of progressive lesions, allowing sufficient time for early intervention, effectively reducing the risk of serious complications and alleviating the medical burden. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method for detecting neuropathic and vascular lesions based on wearable devices according to the present invention; Figure 2 This is a schematic diagram of the location of the method and system for detecting neuropathic and vascular lesions based on wearable devices according to the present invention; Figure 3 This is a schematic diagram of a temperature sensor-temperature bar for the method and system for detecting neuropathic and vascular lesions based on wearable devices according to the present invention; Figure 4 This is a schematic diagram of the needle prick sensation of the method and system for detecting neuropathic and vascular lesions based on wearable devices according to the present invention; Figure 5 This is a schematic diagram of the two-point position discrimination method and system for detecting neuropathic and vascular lesions based on wearable devices according to the present invention; Figure 6 This is a schematic diagram of the nerve conduction velocity detection method and system for detecting neuropathic and vascular lesions based on wearable devices according to the present invention. Figure 7 This is a schematic diagram of the peripheral artery detection of the main trunk of the foot based on the method and system for detecting neuropathy and vascular lesions using a wearable device, as described in this invention. Figure 8 The diagram shows the vibration sensation test of a tuning fork based on the method and system for detecting neuropathic and vascular lesions using a wearable device, as described in this invention. Figure 9 This is an analysis diagram of the ability of the dorsum of the foot to sense vibration by testing with a 128Hz tuning fork, which is based on the method and system for detecting neuropathic and vascular lesions using wearable devices according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0017] It should be noted that all directional and positional terms used in this invention, such as "up," "down," "left," "right," "front," "back," "vertical," "horizontal," "inner," "outer," "top," "lower," "lateral," "longitudinal," and "center," are used only to explain the relative positional relationships and connections between components in a specific state (as shown in the accompanying drawings). They are merely for the convenience of describing the invention and do not require the invention to be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on the invention. Furthermore, the use of "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0018] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0019] In the description of this specification, the references to "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0020] The shoe adopts the appearance and structure of a smart diabetic foot detection shoe. This structure features a transparent design and can be disassembled for easy on and off. The sole contains measurement holes corresponding to 9 peripheral neuropathy detection points on the sole of the foot and 1 peripheral neuropathy detection point on the dorsum of the foot (see details for 10 positioning points). Figure 2 After the foot position is fixed, ten 10g nylon threads will extend from the expansion holes in the sole of the shoe to detect the corresponding positions on the sole and instep.

[0021] like Figure 2 As shown, place the nylon filament perpendicular to the skin surface of the test site, and slowly apply pressure until the nylon filament bends about 1 cm. Hold for 1-2 seconds and then remove. Spacing between each test point is 2-3 seconds to avoid continuous stimulation.

[0022] The standard test points include the base of the big toe of both feet, the first, third, and fifth metatarsal bones, and the heel, for a total of 10 points.

[0023] If a patient perceives ≤8 points, it is considered abnormal (i.e., no sensation at more than 2 points).

[0024] like Figure 3 As shown, a temperature-sensing device—a temperature bar—is used. Two test bars of different materials are installed in the outer shell of the dorsum and sole of the foot. The non-metallic end and the metallic end are used to touch any part of the skin on the dorsum or sole, respectively. If the cold metallic end touches the patient's skin, the sensation of coldness is detected.

[0025] like Figure 4 As shown, touch the skin of your foot with the blunt end of a sewing needle to see if you feel any pain. If you feel little pain, it means you have reduced pain sensation.

[0026] like Figure 5 As shown, the shortest distance between two needle tips that a toe can distinguish is called two-point discrimination. When the distance between the two needle tips is very short, a normal person cannot distinguish them and will only feel the pain of one needle tip. However, when the distance is wide enough, a normal person can distinguish two needle tips. Patients with diabetic peripheral neuropathy may need a wider distance in their fingers and toes to feel two needle tips. Increased two-point discrimination is a reliable objective physical sign.

[0027] like Figure 6 The nerve conduction velocity test shown is a test of nerve electrophysiology, which is the gold standard for diagnosing diabetic peripheral neuropathy. By measuring the motor conduction velocity and sensory conduction velocity of nerves, it is determined whether diabetic peripheral neuropathy is present. The median nerve, ulnar nerve, common peroneal nerve, sural nerve, superficial peroneal nerve, and posterior tibial nerve are usually tested.

[0028] like Figure 7 As shown, the ankle-brachial index (ABI), also known as ABI, is a non-invasive examination method for assessing the peripheral arteries of the foot's main trunk. It is measured using a Doppler stethoscope, comparing the systolic blood pressure of the ankle arteries (anterior and posterior tibial arteries) to that of the brachial artery. It is an easy-to-perform, repeatable, and objective method for determining the severity of ischemia.

[0029] like Figure 8 and Figure 9 As shown, the tuning fork vibration perception test uses a 128Hz tuning fork to test the instep's ability to perceive vibrations.

[0030] The system determines that the patient has worn the shoe stably based on the initial contact signal from the pressure sensor module inside the shoe. It automatically activates the preset acquisition mode (the default is the clinical monitoring mode, which users can switch to at home daily mode on the App). The system also configures the acquisition parameters of each sensor simultaneously. The foot pressure sensor module is set to a sampling frequency of 50Hz, the temperature sensor module is set to a sampling interval of 10 seconds / time, the vibration sensor module activates the stimulation parameters according to the preset detection sequence, and the microcirculation blood flow module is set to a sampling frequency of 100Hz. All parameters are adapted and adjusted based on the calibration data during the initialization phase.

[0031] An array of pressure sensors embedded in the sole of the shoe (covering 10 key areas including the big toe, first / third / fifth metatarsal bones, and heel) continuously captures force signals from different parts of the sole, records the pressure value changes of each sensing unit in real time, forms a continuous pressure distribution data stream, and synchronously marks the acquisition timestamps of each pressure signal to ensure complete capture of the dynamic force process. It focuses on recording the time when the pressure peak occurs, the duration of continuous force, and the change in contact area.

[0032] Temperature sensing units (8-12 evenly distributed detection points) deployed on the inner layer of the insole surface synchronously collect temperature data of various areas of the sole, and extract the temperature values ​​of the corresponding detection points of the left and right feet separately to calculate the real-time temperature difference; areas with temperatures exceeding the normal range (32-36℃) are marked to form a thermal zone distribution record, and ambient temperature interference is automatically filtered during the collection process to ensure the accuracy of the sole body temperature data.

[0033] The vibration sensory module starts working according to a preset sequence. First, a standardized vibration stimulus (128Hz frequency, 0.5mm amplitude) is applied to the toes. The stimulus is maintained for 1-2 seconds and then stopped. After an interval of 2-3 seconds, it switches to other preset detection points in sequence (following the order of 10 nerve lesion detection points on the sole of the foot). The module captures the patient's sensory feedback through the built-in response sensor (the patient can confirm the perception by clicking on the accompanying App, or the system can automatically recognize the reaction of small foot movements). The sensory results of each detection point, as well as the perception delay time (the time from the start of the stimulus to the confirmation of the feedback), are recorded simultaneously to form a vibration sensory function dataset.

[0034] The near-infrared photoplethysmography (PPG) sensor unit, integrated in the upper part of the shoe upper, emits 850nm near-infrared light, which penetrates the skin tissue of the foot to capture blood flow fluctuation signals. It continuously collects raw pulse wave data and extracts raw signals related to key features such as blood flow peak, pulse wave amplitude, and waveform period. During the acquisition process, a filtering algorithm is used to initially suppress motion interference to ensure the stability of blood flow signals. Simultaneously, the raw waveform data is bound to a timestamp for storage.

[0035] Pressure, temperature, vibration response, and pulse wave data collected by each sensor are transmitted in real time to the multi-channel data acquisition module via flexible cabling. The module performs analog-to-digital conversion on the received analog signals (16-bit resolution) and simultaneously performs synchronization alignment of multimodal data based on timestamps, eliminating misaligned data caused by transmission delays. It also performs preliminary noise reduction processing (low-pass filtering for pressure and blood flow signals, and mean filtering for temperature signals) and temporarily stores the processed structured data in the shoe's local cache unit (capacity supports 8 hours of continuous data acquisition and storage).

[0036] The wireless communication module (Bluetooth 5.0 or WiFi) uploads cached data to the wearable central processing unit (mobile app or independent terminal) in batches at a frequency of 5 seconds per upload. If a network interruption occurs during the upload process, the system automatically switches to local caching mode and prioritizes re-uploading the unuploaded data after the network is restored. During the data acquisition process, the system monitors the working status of each sensor in real time. If a sensor shows abnormal data (such as signal interruption or value exceeding the reasonable range), the system pushes a notification to the user through the app while retaining the valid data already collected, without interrupting the overall data acquisition process.

[0037] The default data collection time is 3 minutes / time (clinical monitoring mode) or 1 minute / time (home daily mode). After a single data collection is completed, it automatically enters a low-power standby state and restarts the data collection process at preset intervals (30 minutes / time in clinical mode and 2 hours / time in home mode). Doctors can customize the data collection interval and single data collection time on the remote platform to meet individualized monitoring needs. If abnormal data is detected during the data collection process (such as excessive pressure peak, temperature difference >2℃, abnormal reduction in blood flow amplitude), the system will automatically shorten the next data collection interval and increase the data collection density.

[0038] To further explain, the plantar pressure sensor employs a flexible array design (using piezoresistive or capacitive flexible sensing units with an accuracy of 0.1 kPa), covering 10 key detection areas according to the plantar anatomy (bottom of the big toe, first / third / fifth metatarsal bones, and heel). Each area has 2-3 sensing units forming a local monitoring sub-array. The temperature sensor uses a contact-type NTC thermistor (accuracy ±0.2℃), synchronously deployed next to the corresponding area of ​​the pressure sensor. Additionally, two redundant temperature sensing units are added symmetrically on both feet to ensure data backup. After system startup, the calibration process is automatically executed. The pressure sensor applies pressure to each detection point sequentially using standard weights (10g, 20g, 30g), recording the voltage-pressure correspondence and generating a calibration curve. The temperature sensor is placed in a standard constant temperature environment of 32℃, 34℃, and 36℃ to correct for temperature-resistance conversion errors. Simultaneously, the current ambient temperature is collected as a baseline value for subsequent data compensation.

[0039] After activating the continuous acquisition mode, the pressure sensor captures dynamic force signals on the sole of the foot at a sampling frequency of 50Hz, recording the pressure values ​​of each sensing unit every 10ms to form a time-pressure matrix and synchronously marking the sole contact area at each sampling moment. The temperature sensor collects temperature data at each detection point at an interval of 10 seconds, maintaining a signal stabilization time of 0.5 seconds during each acquisition to avoid interference from instantaneous temperature fluctuations. Simultaneously, it records the temperature acquisition timestamps for the corresponding detection points on the left and right feet to ensure time synchronization. During the acquisition process, the time tags for pressure and temperature data are bound to the clock synchronization unit (1ms accuracy) of the in-shoe data acquisition module to avoid timing misalignment caused by transmission delays. An interrupt trigger mechanism is also employed: when the pressure value suddenly exceeds 200kPa (the preset maximum pressure threshold for normal walking) or the temperature value exceeds the 32-36℃ range, the sampling frequency of the corresponding sensor area is automatically increased (pressure sampling frequency increases to 100Hz, temperature sampling interval shortens to 5 seconds), focusing on capturing abnormal data.

[0040] High-frequency motion noise (such as interference signals generated by shoe insole friction during walking) is filtered out using a 50Hz low-pass filter. Then, median filtering (window size of 5 sampling points) is used to remove pulse noise (such as instantaneous abnormal pressure caused by pebbles pressing against the foot). Finally, linear interpolation is used to supplement missing data caused by sensor failure or signal loss (when the missing rate is less than 5%). Temperature data is smoothed using a moving average filter (window size of 3 sampling values) to smooth temperature fluctuations. Simultaneously, based on the ambient temperature baseline from the initialization phase, environmental compensation is performed on the data at each temperature detection point (compensation formula: actual temperature = collected temperature - ambient temperature deviation value) to eliminate the influence of ambient temperature changes on plantar temperature measurement. After preprocessing, pressure data is converted into structured data of region-time-pressure value, and temperature data is converted into comparative data of left and right feet-detection points-temperature values.

[0041] Calculate the pressure peak value (maximum pressure value of all sensing units in the area), pressure duration (cumulative duration of pressure value exceeding 50 kPa), and contact area (area covered by sensing units with pressure value > 0 kPa) within a 1-minute acquisition cycle to form a regional pressure feature set. Set an abnormality judgment threshold: Referring to clinical normal plantar pressure distribution data, areas with pressure peak values ​​exceeding 300 kPa, durations exceeding 120 seconds / minute, and contact areas accounting for more than 60% of the total area of ​​the region are marked as "potential abnormal pressure points". Further, through spatial correlation analysis, if a potential abnormal pressure point meets the abnormal conditions for 3 consecutive acquisition cycles (3 minutes in total), and the pressure difference between adjacent sensing units in the area exceeds 100 kPa (indicating force concentration), it is finally judged as an "abnormal pressure point". At the same time, the specific anatomical location of the point on the sole of the foot (e.g., at the first metatarsal bone of the right foot) is determined by coordinate mapping (based on the regional coordinates preset on the insole).

[0042] First, the stimulation response results of the vibration sensory module in the corresponding foot detection area are retrieved (e.g., after vibration stimulation of a certain area, if the patient does not perceive it through the App within 5 seconds, or the system does not detect any minor foot movements), and this area is marked as a vibration-unresponsive area. Then, the pressure data characteristics of this area are analyzed: if the pressure duration of the vibration-unresponsive area exceeds 60 seconds / minute, and the pressure value is within the range of 50-200 kPa (within the normal force range), but the patient does not produce any perception-related feedback (e.g., does not adjust walking posture), it is initially determined to be an unresponsive area. Further dynamic verification is performed: during subsequent collection cycles, gradient pressure stimulation is applied to this area (gradually increasing from 50 kPa to 200 kPa, increasing by 50 kPa each time, and holding for 10 seconds). If the patient still does not perceive any feedback, and the pressure sensor records no significant change in the pressure distribution of this area (excluding sensor malfunction), then the "unresponsive area" is finally confirmed, and the nerve innervation information of this area is associated (e.g., "right foot third metatarsal region, corresponding to the superficial peroneal nerve innervation area").

[0043] For each symmetrical detection point (e.g., the base of the left big toe and the base of the right big toe), the absolute value of the temperature difference within a 1-minute acquisition cycle is calculated to form a "detection point-temperature difference" dataset. A threshold for temperature asymmetry is set: if the temperature difference of a symmetrical detection point exceeds 2℃ for three consecutive acquisition cycles, or the temperature difference between different detection points on the same foot exceeds 3℃ (e.g., the toe temperature is 3℃ lower than the heel temperature), it is marked as a "temperature asymmetry area". Further analysis of the clinical correlation of temperature asymmetry is conducted: if the temperature asymmetry area shows "unilateral low temperature" (e.g., the temperature of a certain area on the left foot is 2.5℃ lower than the corresponding area on the right foot), combined with the microcirculation blood flow data of that area (if the PPG amplitude of the synchronously acquired data is less than 15% of the normal range), it suggests insufficient microvascular blood supply in that area; if the temperature asymmetry area shows "unilateral high temperature" (e.g., the temperature of a certain area on the right foot is 2.3℃ higher than the corresponding area on the left foot), and the pressure data of that area shows no obvious abnormal compression, it suggests possible local inflammatory reaction (sensory inflammation caused by neuropathy). Meanwhile, the time trend of temperature asymmetry areas is tracked. If the temperature difference continues to widen within 10 consecutive minutes (e.g., from 2°C to 3°C), a level 2 warning is triggered, indicating an increased risk of nerve or microvascular abnormalities.

[0044] If an area is identified as an abnormal pressure point, its temperature data must be checked simultaneously. If the temperature is normal (32-36℃) and the vibration sensation response is normal, the pressure threshold needs to be reassessed (the threshold may be inappropriate due to changes in patient weight, requiring dynamic adjustment). If an area is identified as a temperature asymmetry area, but the pressure data shows no force on that area (contact area is 0), then "temperature measurement deviation due to lack of contact" (e.g., the area did not touch the ground when the patient walked) needs to be ruled out, and the temperature data when the area touched the ground needs to be collected again. If an area is identified as an unsensible area, but the temperature data shows that the temperature in that area is normal and the pressure data is normal, then a secondary stimulation test of the vibration sensory module needs to be initiated (adjusting the vibration frequency to 256Hz and the amplitude to 0.8mm) to rule out misjudgments caused by vibration sensory module malfunction or delayed initial patient feedback. Through multi-dimensional data cross-validation, the misjudgment rate is controlled below 5%, ensuring the accuracy of the identification results.

[0045] The vibration module uses a miniature columnar vibration motor (5mm in diameter, 10mm in length, with an adjustable vibration frequency range of 60-200Hz). Based on the distribution characteristics of nerves in the foot, it is precisely deployed at 10 key detection sites, including the toes (the pads of the 1st to 5th toes), the forefoot (the heads of the first, third, and fifth metatarsal heads), and the heel. The vibration motor at each site is triangularly distributed with the corresponding pressure sensing unit and temperature sensing unit (8-10mm apart) to ensure precise matching between the stimulation area and the sensor monitoring area. At the same time, the motor is wrapped with a flexible silicone sleeve to avoid direct contact with the skin and cause pressure discomfort.

[0046] The system incorporates three clinically validated standardized stimulation protocols. The default protocol is Protocol 1 (suitable for diabetic neuropathy screening): vibration frequency 128Hz (commonly used clinical nerve detection frequency), amplitude 0.5mm (compliant with the Semimes-Weinstein monofilament test efficacy stimulation intensity), stimulation duration 1-2 seconds, and stimulation interval 2-3 seconds (to avoid sensory fatigue caused by continuous stimulation). Protocol 2 (suitable for early neuropathy monitoring): frequency 60Hz, amplitude 0.3mm. Protocol 3 (suitable for severe neuropathy assessment): frequency 200Hz, amplitude 0.8mm. Doctors can switch between these protocols remotely via a platform or users via an app, depending on their testing needs. Upon system startup, the vibration module automatically performs calibration: it detects the actual vibration parameters of the motor using a built-in accelerometer and compares them with preset parameters. If the deviation exceeds ±5%, the motor drive current is adjusted via a PWM signal to ensure the consistency and accuracy of the stimulation parameters.

[0047] The system determines the patient's condition based on plantar pressure data. When the pressure sensing unit detects stable pressure in all areas of the sole (pressure fluctuation ≤10% within 5 seconds) and no obvious walking movement (pressure distribution change rate <20% / second), it automatically initiates a vibration stimulation sequence. The stimulation sequence activates the vibration motors of each area in the order of toes, forefoot, and heel. Each area is stimulated once according to a preset plan. After the entire sequence is completed, it is repeated after a 30-second interval. A total of 3 rounds of stimulation are performed within a single detection cycle (to ensure repeatability of results). If the patient walks during the detection process (pressure distribution change rate ≥20% / second), the system immediately pauses the stimulation and resumes stimulation from the paused area after the pressure stabilizes again.

[0048] Two feedback methods are provided for patients to choose from, with the default being a dual-mode of active feedback + passive monitoring: Active feedback is achieved through a companion mobile app. After the patient senses the vibration stimulus, they need to tap the "Confirm Sensation" button on the phone screen. The system records the time difference between the tap and the stimulus activation as the reaction time. Passive feedback is achieved through a miniature accelerometer (sampling frequency 100Hz) built into the shoe. When the patient experiences a slight foot movement after sensing the vibration (such as toe curling or slight foot movement), the sensor captures the movement signal and transmits it to the data acquisition module. The system automatically determines that the vibration has been sensed and records the time difference between the movement signal occurrence and the stimulus activation. The two feedback methods can work independently. If active feedback is not triggered but passive feedback detects movement, it is still considered as sensed, avoiding misjudgments due to patient operation delays.

[0049] After stimulation is initiated, the system simultaneously records three types of data: stimulation initiation timestamp (T0), feedback trigger timestamp (T1, the moment of active feedback click or passive feedback action detection), and stimulation end timestamp (T2, automatically terminated according to the preset duration). If a feedback signal is detected within the stimulation duration (T0 to T2), it is determined as "positive perception," and the reaction time = T1 - T0. If no feedback signal is detected within 3 seconds after the stimulation ends, it is determined as "negative perception," and the reaction time is recorded as ">T2 - T0 + 3 seconds." If three consecutive rounds of stimulation to the same area result in "negative perception," the system automatically switches to a higher intensity stimulation scheme (e.g., switching from scheme 1 to scheme 3) and re-executes the stimulation detection for that area to eliminate false negatives caused by insufficient stimulation intensity.

[0050] Based on the collected perception status and reaction time data, three core evaluation indicators were extracted: ① Perception pass rate = (number of positive detection sites / total number of detection sites) × 100%, reflecting the overall vibration perception range; ② Average reaction time = the average reaction time of all positive detection sites, reflecting the vibration perception speed; ③ Perception threshold level, divided into 5 levels according to reaction time (Level 1: ≤0.5 seconds, normal; Level 2: 0.5-1 seconds, mild reduction; Level 3: 1-2 seconds, moderate reduction; Level 4: 2-3 seconds, severe reduction; Level 5: >3 seconds or negative perception, extremely severe reduction). The system weights and fuses three indicators (with weights of 0.3, 0.4, and 0.3 respectively) to generate a comprehensive vibration sensitivity score (0-100 points, ≥80 points for normal, 60-79 points for mild abnormality, 40-59 points for moderate abnormality, and <40 points for severe abnormality), and associates it with the corresponding nerve innervation areas (e.g., the toes correspond to the superficial peroneal nerve, and the heel corresponds to the posterior tibial nerve), providing a basis for the localization of nerve lesions.

[0051] During the data acquisition process, the system monitors the operating status of the vibration motor in real time. If the motor current is abnormal (exceeding the rated current ±20%) or the acceleration sensor does not detect a vibration signal, it is determined to be a "module failure". The system immediately suspends the detection of that area, pushes a fault prompt to the App (such as "vibration module of the second toe of the right foot is abnormal"), and skips that area to continue the detection of other areas. If the difference in the perception pass rate of two consecutive tests for the same patient exceeds 30%, or the difference in the average reaction time exceeds 1 second, the system automatically starts a third retest and takes the median of the three tests as the final result to avoid evaluation bias caused by accidental factors. All detection data and evaluation results are uploaded to the cloud platform simultaneously, allowing doctors to trace the original data and stimulation parameters for easy clinical review.

[0052] Employing a near-infrared photoplethysmography (PPG) sensing unit, the core components include an 850nm wavelength near-infrared LED light source, a high-sensitivity photodiode (PD), and a signal amplification circuit. The module is packaged as a flexible patch structure (15mm × 8mm × 2mm) and integrated into the upper inner side of the shoe upper (corresponding to the dorsalis pedis artery branch area), ensuring that the distance between the light source and the skin is controlled at 3-5mm to avoid light scattering interference. After system startup, the sensing parameters are automatically configured: the LED drive current is set to 50mA (ensuring a penetration depth of 3-5mm, covering subcutaneous microcirculation vessels), the signal sampling frequency is 100Hz, the ADC conversion resolution is 16-bit, the acquisition cycle is continuous for 30 seconds / time, with repeated acquisition at 5-minute intervals, and supports dynamic adjustment based on foot activity status feedback from the pressure sensor (shortening the acquisition interval to 2 minutes during walking and extending it to 10 minutes when stationary).

[0053] Near-infrared LED light sources continuously emit stable light signals. After penetrating the skin tissue of the foot, the hemoglobin (oxygenated and deoxygenated hemoglobin) in the subcutaneous blood absorbs and scatters the light. Photodiodes capture changes in the intensity of the reflected light in real time and convert them into a weak current signal. After the signal is amplified by a preamplifier circuit (1000x magnification), it is initially suppressed by a low-noise filter circuit (10Hz cutoff frequency) to suppress ambient light and motion interference. Then, the analog signal is converted into a digital signal by an ADC module to form the raw PPG waveform data. The timestamp of each sampling point is recorded synchronously to ensure the integrity of the data timing. During the acquisition process, if the detected signal amplitude is lower than the threshold (<500 count units), the system automatically increases the LED drive current (maximum not exceeding 80mA). If there is still no effective signal, a "poor fit" prompt is pushed to guide the patient to adjust the shoe upper position.

[0054] First, an adaptive filtering algorithm is used to separate motion artifacts in the PPG signal. By extracting the synchronous motion signal from the plantar pressure sensor as a reference, an adaptive filter is constructed to remove low-frequency interference caused by walking and foot micro-movements. Then, a high-pass filter (cutoff frequency 0.5Hz) is performed to remove breathing interference and baseline drift. Wavelet transform (db4 wavelet basis, decomposed into 5 layers) is used to remove high-frequency noise (such as electronic noise and ambient light abrupt changes), retaining the effective pulse wave frequency band of 0.5-10Hz. Finally, peak detection and interpolation are used to repair local data loss caused by signal interruption (when the missing length is ≤5 sampling points), forming a smooth and continuous standardized PPG waveform.

[0055] Based on the preprocessed PPG waveform, two core feature indicators are extracted: First, temporal features, which use a peak detection algorithm to identify the systolic peak (S wave), diastolic peak (D wave), and dicrotic wave (P wave) of the pulse wave, calculating the pulse wave period (the time interval between two adjacent S waves, reflecting heart rate), systolic rise time (the duration from the waveform baseline to the S wave peak), and diastolic fall time (the duration from the S wave peak to the D wave trough). Second, amplitude features, which calculate the difference between the S wave peak and the baseline as the pulse wave amplitude (reflecting blood perfusion intensity), calculate the S / D wave amplitude ratio (S / D ratio, reflecting vascular elasticity), and calculate the amplitude variation coefficient (reflecting blood flow stability) using a sliding window method (window size of 5 pulse cycles). Simultaneously, based on the pulse wave area integral and preset tissue optical parameters (such as hemoglobin absorption coefficient and skin scattering coefficient), the relative blood flow value (unit: ml・min⁻¹・cm⁻²) is derived using the Lambert-Beer law, achieving a quantitative assessment of blood flow.

[0056] The extracted blood flow parameters were correlated with clinical reference standards. The pulse wave cycle was compared with the simultaneously acquired heart rate data (obtained via a mobile app heart rate sensor), and the deviation was required to be ≤5%. Correlation analysis was performed on the pulse wave amplitude and the data measured simultaneously by a near-infrared flowmeter (clinical standard equipment), ensuring R² ≥ 0.85. If any parameter exceeded the normal reference range (e.g., pulse wave amplitude < 0.2V, relative blood flow < 1.5 ml・min⁻¹・cm⁻², amplitude coefficient of variation > 20%), it was marked as a potential abnormal parameter, and the corresponding acquisition time, foot position, and patient status (walking / resting) were recorded. Simultaneously, persistent abnormalities were screened through trend analysis: if the same parameter exceeded the normal range in three consecutive acquisitions, and the trend showed a continuous deterioration (e.g., continuously decreasing amplitude, continuously increasing coefficient of variation), it was judged as "significantly abnormal," triggering the subsequent risk assessment process.

[0057] The extracted blood flow, pulse wave amplitude, pulse wave period, S / D ratio, and amplitude variation coefficient indicators are structured and packaged in the format of collection timestamp, indicator name, value, and validity label. They are synchronized with other modal data such as pressure, temperature, and vibration sensation (based on unified timestamp alignment) and uploaded in batches to the wearable central processing unit and cloud platform via Bluetooth / WiFi module. The original data and feature indicators of the most recent 7 days are retained locally, supporting offline viewing and retrospective analysis. The cloud platform stores complete data to provide data support for subsequent longitudinal trend analysis and disease progression prediction.

[0058] The central processing unit receives structured data packaged and uploaded by the acquisition module via Bluetooth / WiFi. The data package contains raw data, preprocessed features, and unified timestamps for four modalities: pressure (P), temperature (T), vibration sensation (V), and microcirculation (PPG). First, standardization processing is performed: missing sampling points for each modality are filled in using linear interpolation based on the timestamp (missing rate ≤5%) to ensure complete synchronization of the timeline. Spatially, based on a pre-defined sensor coordinate mapping table between the insole and upper, a spatial correspondence is established between pressure sensing points, temperature detection points, vibration stimulation points, and microcirculation monitoring areas, forming a spatiotemporally unified multimodal data matrix (dimension: time frame × spatial point × modal feature).

[0059] For pressure data, spatiotemporal attention convolution enhancement is employed: a 1×3 temporal convolution kernel is used to extract the temporal trend of pressure peaks, and a 3×3 spatial convolution kernel is used to highlight the regional characteristics of continuous abnormal pressure. Then, the feature contribution of long-term pressure points (duration > 60 seconds) is weighted using an attention mechanism. For temperature data, temperature gradient + heat zone diffusion features are constructed: in addition to the basic left and right foot temperature difference, the rate of change of temperature difference over time (gradient) is calculated, and the spread trend of potential heat zones (such as inflammation spread) is identified through a heat map diffusion algorithm. For vibration sensation data, a dynamic fitting of perception response and stimulus intensity is used: based on the reaction time under different stimulus intensities, a quadratic function curve is fitted, and the curve slope (reflecting the rate of change of perception sensitivity) and inflection point (critical perception intensity) are extracted as enhancement features. For microcirculation data, pulse wave morphological entropy features are added: by calculating the information entropy of PPG waveforms, waveform irregularity (an important indicator of abnormal vascular elasticity) is quantified, supplementing the deficiencies of traditional amplitude and period features.

[0060] A modality-feature-pathology association network is constructed to uncover potential pathological associations between different modalities, forming creative association features. First, the mutual information value between features of each modality is calculated, and highly associated feature pairs (mutual information value > 0.6) are selected, such as the duration of abnormal pressure areas and the temperature difference of the corresponding areas, the peak pressure and microcirculatory blood flow, and the vibration perception threshold and the temperature of the corresponding nerve innervation area. Fusion operations are performed on highly associated feature pairs: a feature product weighting + difference penalty strategy is adopted, such as pressure-blood flow association feature = peak pressure × blood flow × association weight - |pressure duration - blood flow stabilization time| × penalty coefficient. The weights and penalty coefficients are adaptively adjusted through clinical data training. Further, a three-dimensional association matrix is ​​constructed: using spatial points (10 detection areas on the sole of the foot) - time periods (1 minute units) - pathological association degree as dimensions, the degree of multimodal feature synergy abnormality of each spatial point in different time periods is quantified, forming an intermediate feature set for cross-modal fusion.

[0061] A dual-branch network consisting of an immediate feature branch and a temporal trend branch is designed, combined with an attention mechanism to achieve multi-dimensional fusion. The immediate feature branch uses a CNN network, inputting cross-modal correlation features from the current acquisition cycle to extract immediate pathological abnormalities (such as exceeding the instantaneous pressure peak and a sudden drop in blood flow). The temporal trend branch uses a BiLSTM network, inputting historical feature sequences from the last 5 acquisition cycles (a total of 15 minutes) to capture the trend of lesion development (such as gradually widening temperature difference and continuously prolonged vibration response time). A modality-pathology dual attention mechanism is introduced into the dual-branch output layer: modality attention automatically allocates the weights of each modality (such as increasing the weights of vibration and temperature modalities when assessing neuropathology, and increasing the weights of microcirculation and pressure modalities when assessing vascular lesions), and pathology attention strengthens the contribution of key pathological features based on clinical pathological association rules (such as the correlation between neuropathology and vibration and temperature being higher than that of pressure). Finally, the dual-branch outputs are fused through a fully connected layer to generate a preliminary multimodal fusion feature vector (dimension: 1×128).

[0062] Clinical diagnostic criteria are introduced as constraints for the fusion model to improve the clinical credibility and accuracy of the results. Clinically recognized pathological judgment rules (e.g., a temperature difference >2°C between the left and right feet indicates inflammation or perfusion abnormalities, a vibration sensation reaction time >2 seconds indicates moderate neuropathy, and blood flow <1.5 ml・min⁻¹・cm⁻² indicates insufficient vascular supply) are converted into numerical constraint thresholds to construct a "rule constraint matrix." The fusion feature vector output by the dual-branch network is calibrated: if the evaluation result corresponding to a certain feature vector conflicts with the constraint rule (e.g., the model predicts mild neuropathy, but the vibration sensation reaction time is >3 seconds), a feature reweighting mechanism is activated to increase the weight of the conflicting modality feature and recalculate the fusion result. Simultaneously, patient baseline data (normal state data collected during the first wear) is incorporated, and the current fusion feature is compared with the individual baseline to generate an individual relative abnormality index, avoiding misjudgments due to individual differences (e.g., obese patients have higher baseline plantar pressure, requiring personalized threshold adjustment).

[0063] Based on the calibrated fusion feature vectors, a multi-task learning model is used to simultaneously output neuropathy grading and vascular lesion risk scores. The neuropathy grading task uses a Softmax classifier, mapping the fusion features to five levels (Level 0: normal, Level 1: mild impairment, Level 2: moderate impairment, Level 3: severe impairment, Level 4: profound impairment), with the classification boundaries optimized through training on 50 clinical samples. The vascular lesion risk scoring task uses a regression model, outputting risk values ​​from 0 to 100. The scoring dimensions include blood flow adequacy (30% weight), vascular elasticity (25% weight), blood flow stability (20% weight), and cross-modal association anomaly (25% weight). Finally, consistency verification is performed: if the correspondence between neuropathy grading and vascular risk scores does not conform to clinical patterns (e.g., profound neuropathy but vascular risk score < 40), the model is recalculated to ensure the clinical rationality of the output results.

[0064] After the fusion processing is completed, the central processing unit pushes the evaluation results to the user's app and cloud platform. Simultaneously, it dynamically adjusts subsequent acquisition parameters based on the evaluation results, forming a closed-loop optimization. If a high-risk area is detected (such as a foot area exhibiting abnormal pressure, temperature asymmetry, decreased blood flow, and no vibration response), the pressure sampling frequency for that area is automatically increased to 100Hz, the temperature sampling interval is shortened to 5 seconds, the vibration stimulation frequency is increased by 50%, and the microcirculation acquisition cycle is shortened to 2 minutes, focusing on capturing the progression of lesions in that area. If the evaluation result is normal or slightly abnormal, the default acquisition parameters are maintained to save power. Simultaneously, the modal weight allocation and rule constraint triggering during this fusion processing are recorded to provide data support for subsequent algorithm iterations.

[0065] Multimodal data were collected from 500 diabetic patients (including healthy controls, early-stage, moderate-stage, and severe-stage patients), covering the entire lifecycle of data acquisition for pressure, temperature, vibration sensation, and microcirculation. Clinical diagnostic labels (neuropathy grade, degree of vascular impairment) and individual baseline information (age, disease duration, weight) were also recorded. A dual-labeling method was used to improve data quality: first, a gold standard label based on clinical test results from physicians; second, auxiliary labels automatically generated through modal feature-pathology mapping rules (e.g., abnormal pressure + temperature asymmetry + decreased blood flow corresponding to combined neurovascular lesions). Samples with consistent labels were included in the core training set (80%), while inconsistent samples were supplemented with labels after physician review. To address the scarcity of early-stage lesion samples, cross-modal feature transfer enhancement technology was employed: the modal features of severe-stage lesion samples were attenuated according to pathological progression patterns (e.g., reducing vibration reaction time and blood flow amplitude) to generate synthetic samples simulating early-stage lesions, improving the balance of the training set samples by 40%.

[0066] A parallel model combining static features, a random forest path, and a dynamic temporal-LSTM path is constructed to creatively integrate static pathological features with dynamic trends. In the static path, the random forest model inputs single-modal deep enhancement features and cross-modal correlation features (such as pressure-temperature correlation values ​​and vibration-blood flow correlation values), and outputs static lesion assessment results (emphasizing the current state) through voting by 100 decision trees. In the dynamic path, the LSTM network adopts a two-layer bidirectional structure. The first LSTM layer extracts the short-term temporal trends of each modality feature (data from the last 10 minutes), and the second LSTM layer captures long-term evolution patterns (data from the last 24 hours), outputting the temporal lesion progression trend value. A "modal attention fusion gate" is introduced in the middle layer of the model to dynamically allocate the weights of the two paths based on the modal quality of the current data (such as the signal-to-noise ratio of microcirculation signals and the integrity of vibration feedback). For example, when the signal quality is high, the static path weight accounts for 60%, and when temporal fluctuations are significant, the dynamic path weight increases to 70%.

[0067] An iterative training model combining hierarchical training and difficult case mining is employed. The first round of training is based on the core training set, using the Adam optimizer to minimize a hybrid loss function (classification loss: FocalLoss addresses class imbalance in neuropathic grading; regression loss: HuberLoss reduces the impact of outliers in vascular risk scores). The initial learning rate is set to 0.001, decaying by 10% every 5 rounds. After training, difficult case samples (samples with prediction errors > a threshold, such as early lesions misclassified as healthy or severe lesions missed) are selected. Key pathological features of these difficult cases are supplemented through "feature enhancement annotation" (e.g., annotating subtle variations in microcirculatory waveforms and slight prolongations in vibration perception reaction time in early lesions), constructing a dedicated difficult case training set. The second round of training uses a transfer learning approach: the model parameters from the first round of training are used as initial weights, optimizing the model's decision boundaries only for the difficult case set. Clinical prior rules are introduced as regularization terms (e.g., "when the temperature difference between the left and right feet is >2℃, the neuropathic grading is at least mild") to constrain the model's prediction results to conform to clinical logic, thereby increasing the sensitivity of early lesion detection to 95%.

[0068] After receiving real-time multimodal fusion feature vectors, the model calculates the probability distribution of each level for the static path using a random forest (e.g., normal 30%, mild 50%, moderate 20%), and outputs the level change trend for the dynamic path using an LSTM (e.g., the slope of the level probability change in the last 3 acquisitions). The final level is output using a "probability-trend joint determination method": if the probability of a certain level in the static path is >60% and the dynamic trend shows no reverse change (e.g., the mild probability continues to rise), that level is directly determined; if no single level probability is >60%, then the level with the highest probability and rising trend is selected based on the dynamic trend value (e.g., mild probability 45%, moderate 40%, but the mild trend slope is positive and the moderate trend slope is negative, so it is determined as mild). Simultaneously, the level confidence score (a weighted sum of the consistency scores of static probability and dynamic trend) is output. When the confidence score is <70%, supplementary detection is triggered (e.g., increasing vibration stimulation intensity, or increasing microcirculation acquisition) to ensure level accuracy.

[0069] The model outputs time-series predictions of blood flow-related features (blood flow and pulse wave amplitude changes over the next hour) using a dynamic path LSTM, combined with current blood flow features from a static path, to construct a two-dimensional scoring system of "immediate risk + trend risk". Immediate risk is based on the relative value of blood flow and pulse wave morphological entropy features derived from Lambert-Beer's law, mapped to a score of 0-50 through linear regression. Trend risk is calculated based on the difference between the time-series prediction and the individual's baseline; if the predicted blood flow continues to decrease, points are added (maximum 50 points), while points are deducted if it stabilizes or increases. An "individual difference calibration factor" is incorporated into the scoring process: the scoring threshold is adjusted based on the patient's disease duration and baseline weight information. For example, for patients with a disease duration >10 years, the risk score threshold is reduced by 10% to avoid misjudgments due to differences in individual baseline conditions. The final risk score is the sum of immediate risk and trend risk, categorized into levels of 0-30 (low risk), 31-60 (medium risk), and 61-100 (high risk).

[0070] After system deployment, a closed-loop mechanism of real-time monitoring, clinical feedback, and model updates is established. The central processing unit records the difference between the model's prediction and the doctor's final diagnosis for each patient. When the difference rate exceeds 10%, the patient's multimodal data is automatically included in the incremental training set. Model fine-tuning is performed monthly, using an incremental learning algorithm (Finetune) to update the decision tree weights of the random forest and the network parameters of the LSTM, eliminating the need to retrain the entire dataset and saving computational resources. Simultaneously, a doctor's experience embedding module is introduced, allowing doctors to remotely annotate key features of special cases (such as modal manifestations of rare diseases) through a remote platform. This annotation information is transformed into additional constraint rules for the model, enabling it to gradually adapt to complex clinical scenarios. After long-term use, the accuracy rate has improved to over 93%.

[0071] After receiving structured data uploaded by the central processing unit, the cloud platform constructs a three-level time-series database for hierarchical storage. The first-level database stores raw multimodal data (retained for 90 days), the second-level database stores extracted feature indicators (retained for 1 year), and the third-level database stores neuropathic grading and vascular risk scoring assessment results (long-term storage). All data is indexed by "patient ID-timestamp-modal type" and supports millisecond-level backtracking queries. Simultaneously, individual baseline modeling is initiated: a 30-day sliding window algorithm is used, with stable data from the first 3-7 days of patient wear as the initial baseline. This baseline is updated every 7 days thereafter, and the baseline range is dynamically adjusted through a weighted average (0.7 weight for recent data, 0.3 weight for older data), generating an "individual dynamic baseline interval" (e.g., a patient's normal pulse wave amplitude baseline is 0.3-0.5V, slowly adjusted as the disease progresses), avoiding misjudgments caused by a fixed baseline.

[0072] A dual-track trend model was designed to integrate individual longitudinal evolution with group lateral reference to uncover early disease signals. In longitudinal trend analysis, three types of trend parameters were calculated for each core indicator (duration of peak pressure, temperature difference between left and right feet, vibration reaction time, and relative blood flow): first, the linear trend slope (reflecting the rate of change of the indicator, such as a monthly increase of 0.3℃ in temperature difference); second, the curve curvature (reflecting the acceleration of change, such as an accelerated rate of decrease in blood flow); and third, the fluctuation entropy (reflecting the stability of the indicator, such as an increase in the amplitude variation coefficient). In lateral trend analysis, the trend of patient indicators was compared with the trend curve of a reference group of the same age group, disease duration, and weight to calculate the "trend deviation" (e.g., the rate of increase in vibration reaction time of the patient is 2.5 times the group mean). Further "multimodal collaborative trend mining" was performed: the linkage relationship between cross-modal trends was captured through association rule algorithms, such as the probability of the synergistic occurrence of "continuous expansion of abnormal pressure area" and "continuous decrease in blood flow in the corresponding area" and "gradual increase in temperature difference," generating a "synergistic trend strength" index. If the strength > 0.8, it is marked as a high-risk trend.

[0073] A progressive model of abnormal indicators, abnormal trends, and accumulated risks is constructed to avoid false alarms triggered by single abnormalities. First, initial warning thresholds are set for each indicator (based on clinical standards and population data). When a single indicator exceeds the individual's dynamic baseline range, it is recorded as a "single abnormality" and accumulates 1 risk point. If the longitudinal trend slope of the indicator exceeds the safe range (e.g., blood flow decrease slope < -0.05 ml・min⁻¹・cm⁻² / day), an additional 2 risk points are accumulated. If cross-modal synergistic abnormalities occur simultaneously (e.g., abnormal pressure + decreased blood flow), the risk value doubles. Risk values ​​are updated using an exponential decay mechanism (decreasing by 20% daily when there are no abnormalities). When the accumulated risk value > 10 points, dynamic threshold calibration is initiated: a Bayesian algorithm is used to fuse recent patient data, clinical history (e.g., history of foot ulcers), and population risk data to adjust the patient's personalized warning threshold (e.g., for patients with a history of ulcers, the threshold is reduced by 30%), improving the targeting of warnings.

[0074] The design incorporates a triple-triggered logic to achieve accurate early warning, covering different stages of disease. The first level of "trend mutation trigger": When the change of an indicator exceeds 50% of the individual's baseline range within 24 hours (e.g., a sudden increase in temperature difference from 1℃ to 1.8℃), or a sudden change in the trend slope (e.g., the rate of increase in vibration perception reaction time from 0.1 seconds / day to 0.5 seconds / day), a Level 1 warning (potential acute abnormality) is immediately triggered. The second level of "cross-modal trigger": When at least two modalities simultaneously show a persistent abnormal trend (e.g., persistent abnormal pressure + temperature asymmetry + decreased blood flow, or decreased vibration perception + increased temperature in the corresponding area), and the synergistic trend strength is >0.85, a Level 2 warning (risk of combined neurovascular lesions) is triggered. The third level of "predictive trigger": Based on the LSTM time-series prediction model, the patient's multimodal trend data for the past 7 days is input to predict the trend of indicator changes in the next 14 days. If the prediction results show that the neuropathic grading will improve by 1 level or more, or the vascular risk score will exceed 60 points (the threshold for medium-to-high risk), a Level 3 warning (predictive warning of progressive lesions) is triggered 3-7 days in advance, reserving a time window for intervention.

[0075] The warning system is divided into three levels, with corresponding differentiated push notifications and intervention suggestions. Level 1 Warning (Red): For acute abnormalities or high-risk progression, a pop-up and voice reminder is immediately pushed to the patient via the app, and a warning report (including abnormal indicators, trend charts, and corresponding pathological explanations) is sent to the attending physician simultaneously. It is recommended to seek medical attention within 24 hours. Level 2 Warning (Yellow): For persistent abnormalities, a reminder is pushed once a day, and a weekly trend report is sent to the doctor. It is recommended to adjust the home care plan (such as changing insoles or adjusting walking posture). Level 3 Warning (Blue): For predictable risks, trend warning notifications and personalized intervention guidelines are pushed (such as increasing the frequency of foot massages or adjusting diet). The prediction results are updated every 3 days. The warning report adopts a "clinical-common language" design: the doctor's side displays professional indicators, trend data, and pathological correlation analysis, while the patient's side uses visual charts (such as trend curves and heat maps) and simplified interpretations (such as "Your right first metatarsal area has persistently high pressure and decreased blood flow; you should avoid standing for long periods of time").

[0076] A closed-loop mechanism of early warning, intervention, and feedback is established to continuously optimize the accuracy of early warnings. After receiving an early warning, patients or doctors can upload intervention measures (such as medical examination results and nursing adjustment records) through the App. The cloud platform will correlate the intervention measures with subsequent indicator changes and analyze the effectiveness of the intervention (e.g., a risk value decrease of more than 50% after intervention is considered effective). For ineffective early warnings (e.g., no abnormalities are found after the early warning), the triggering cause is automatically traced, and the patient's dynamic baseline range and early warning threshold are adjusted (e.g., expanding the baseline range and increasing the risk accumulation threshold). The early warning data of all patients are summarized monthly, and the parameters of the trend analysis model and prediction model are updated using incremental learning algorithms to optimize the trend slope calculation, collaborative trend intensity assessment, and prediction accuracy, thereby improving the long-term early warning accuracy rate to over 92% and controlling the false early warning rate to below 8%. At the same time, it supports doctors to manually annotate special cases (such as early warning signals for rare diseases), transforming the annotation information into additional constraint rules for the model, enhancing the model's adaptability to complex clinical scenarios.

[0077] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for detecting neuropathic and vascular lesions based on wearable devices, characterized in that: Includes the following steps: Step S110: The pressure sensing unit is calibrated using standard weights to generate voltage and pressure curves; the temperature sensing unit is corrected for conversion errors in standard environments of 32℃, 34℃, and 36℃; and the vibration sensing unit is calibrated for stimulation parameters using an accelerometer. Step S120: The pressure sensing unit captures dynamic pressure signals on the sole of the foot, the temperature sensing unit collects the temperature of the area and the temperature difference between the left and right feet, the vibration sensing unit applies stimulation according to the preset scheme and collects the sensory response, and the PPG sensing unit collects the original waveform of microcirculation blood flow. Step S130: Low-pass filtering and median filtering are used to denoise the pressure and blood flow signals, and mean filtering and environmental compensation are used for the temperature signal. Multimodal data are synchronized by timestamps, and misaligned and abnormal data are removed. Step S140: Extract peak value, duration and contact area features from pressure data; extract temperature difference and heat zone distribution features from temperature data; extract sensing pass rate, average reaction time and sensing threshold level from vibration response data; extract pulse wave period, amplitude and relative blood flow from blood flow data. Step S150: Construct a spatiotemporally unified multimodal data matrix, enhance features through convolution enhancement, temperature gradient calculation, perception, stimulus fitting and waveform entropy analysis, and establish a modality and pathology association network to generate cross-modal fusion features; Step S160: Using a dual-branch network combined with a multi-task learning model, the neuropathic grading and vascular lesion risk score are output simultaneously; Step S170: Based on the assessment results and trend analysis, trigger a Level 1, Level 2, or Level 3 early warning and push corresponding intervention suggestions.

2. The method for detecting neuropathic and vascular lesions based on a wearable device according to claim 1, characterized in that: The pressure sensing unit generates a voltage-pressure correspondence curve through standard weight calibration; the temperature sensing unit corrects conversion errors in standard environments of 32℃, 34℃, and 36℃; and the vibration sensing unit calibrates stimulation parameters through an accelerometer, including: Standard weights of 10g, 20g, and 30g were used to determine the locations of 10 nerve lesion detection points on the sole of the wearable device. According to the order of the detection points, each standard weight was placed steadily on the pressure sensing unit of each detection point. The pressure state of each weight was kept stable for 3-5 seconds, and the voltage output value of each detection point under different weight pressures was recorded simultaneously. After all detection points have completed the application of different weights and data recording, the voltage-pressure correspondence curve of each detection point was generated based on the collected voltage and pressure data through linear fitting or interpolation algorithms to complete the calibration of the pressure sensing unit. Set up standard constant temperature environments of 32℃, 34℃, and 36℃, and place the temperature sensing unit entirely in these environments. After the temperature sensing unit stabilizes in each standard constant temperature environment for 5-10 minutes to ensure that its detection value tends to stabilize, record the actual resistance output value of the sensing unit at each standard temperature. Based on the deviation between the standard temperature value and the corresponding measured resistance value, adjust the coefficient in the temperature-resistance conversion formula to correct the conversion error, so that the resistance output of the sensing unit at 32℃, 34℃, and 36℃ can be accurately mapped to the corresponding standard temperature. At the same time, collect the current ambient temperature in the calibration environment and record it as the baseline value for temperature data compensation. The calibration procedure of the vibration sensor unit is initiated, activating the built-in accelerometer. The vibration motor is controlled to run according to preset parameters. The actual vibration parameters of the motor are collected in real time through the accelerometer. The collected actual vibration parameters are compared with the preset parameters, and the deviation value is calculated. If the deviation value exceeds ±5%, the drive current of the motor is adjusted through the PWM signal. After each adjustment, the actual vibration parameters are collected and compared with the preset parameters again until the deviation value between the actual vibration parameters and the preset parameters is ≤ ±5%. The adjustment is then stopped, and the calibration of the vibration sensor unit stimulation parameters is completed.

3. The method for detecting neuropathic and vascular lesions based on a wearable device according to claim 1, characterized in that: The pressure sensing unit captures dynamic pressure signals from the sole of the foot, the temperature sensing unit collects the regional temperature and the temperature difference between the left and right feet, the vibration sensing unit applies stimulation according to a preset scheme and collects the sensory response, and the PPG sensing unit collects the raw waveforms of microcirculatory blood flow, including: The flexible array pressure sensing unit is activated, and the basic sampling frequency is set to 50Hz. The pressure value of each sensing unit is recorded every 10ms to form a time and pressure matrix. During the acquisition process, the foot contact area at each sampling moment is marked in real time, and the time of pressure peak occurrence, duration of continuous force, and contact area are recorded synchronously. If the pressure value of a certain area suddenly exceeds 200kPa, the sampling frequency of the pressure sensing unit in that area is automatically increased to 100Hz until the pressure value falls back below the threshold and the basic frequency is restored, continuously capturing changes in dynamic pressure signals of the foot. Activate the contact-type NTC thermistor temperature sensing unit, set the basic sampling interval to 10 seconds / time, maintain a signal stabilization time of 0.5 seconds during each acquisition to avoid interference from instantaneous temperature fluctuations, and synchronously record the temperature values ​​of the corresponding detection points of the left and right feet. After the acquisition is completed, calculate the temperature difference between the symmetrical detection points of the left and right feet to form a dataset of detection points and temperature differences. If the temperature of a certain area exceeds the normal range of 32-36℃, the sampling interval of the temperature sensing unit in that area is automatically shortened to 5 seconds / time, and the area is marked as a hot zone and the distribution of the hot zone is recorded. The area temperature and the temperature difference between the left and right feet are continuously collected. The system calls upon the built-in miniature columnar vibration motor of the vibration sensor unit, loads a preset stimulation program, and activates the vibration motors at each detection point in the order of toes, forefoot, and heel, applying standardized vibration stimulation. After stimulation is initiated, the system receives active feedback from the patient through the accompanying mobile app, while simultaneously capturing passive feedback from the patient's foot through a miniature accelerometer inside the shoe. The system records whether each detection point is perceived and the perception delay time. If the same detection point is perceived negatively for three consecutive rounds of stimulation, the system automatically switches to a higher intensity stimulation program and re-executes the stimulation and perception response acquisition at that detection point. The near-infrared photoplethysmography (PPG) sensor unit integrated into the upper inner side of the shoe upper is activated. The initial LED drive current is set to 50mA, the sampling frequency to 100Hz, and the ADC conversion resolution to 16-bit. The near-infrared LED light source continuously emits a stable light signal. After penetrating the skin tissue of the foot, the photodiode captures the changes in the intensity of the reflected light in real time and converts the light signal into a weak current signal. After the signal is amplified by the preamplifier circuit, it is initially suppressed by the 10Hz low-pass filter circuit to suppress ambient light and motion interference. Then, the analog signal is converted into a digital signal by the ADC module to form the original waveform data of microcirculation blood flow. During the acquisition process, the signal amplitude is monitored in real time. If the amplitude is lower than 500 count units, the LED drive current is automatically increased. If there is still no effective signal, a poor fit prompt is pushed. At the same time, the filtering algorithm further suppresses motion interference to ensure the stability of the blood flow signal. The original waveform data is bound and stored with a timestamp, and the original waveform of microcirculation blood flow is continuously acquired.

4. The method and system for detecting neuropathic and vascular lesions based on wearable devices according to claim 1, characterized in that: The pressure and blood flow signals are denoised using low-pass and median filtering, and the temperature signal is denoised using mean filtering and environmental compensation. Multimodal data is synchronized via timestamps, and misaligned and abnormal data are removed. This includes: A low-pass filter with a cutoff frequency of 50Hz is used to filter the raw plantar pressure signal to remove high-frequency motion noise generated by insole friction during walking. Then, a median filter with a window size of 5 sampling points is selected to further process the low-pass filtered pressure signal to remove pulse noise caused by stones pressing on the feet. If there are missing data in the processed signal due to sensor failure or signal loss, and the missing rate is less than 5%, the missing sampling points are supplemented by linear interpolation to ensure the integrity of the pressure signal. An adaptive filtering algorithm was employed to extract the motion signal synchronously acquired by the plantar pressure sensor as a reference signal. An adaptive filter was constructed to separate motion artifacts in the blood flow signal. Then, a low-pass filter with a cutoff frequency of 10Hz was applied to remove respiratory interference and baseline drift in the blood flow signal. Subsequently, the blood flow signal was decomposed into 5 layers using the db4 wavelet basis. Wavelet transform was used to remove electronic noise and high-frequency interference from sudden changes in ambient light, retaining the effective pulse wave frequency band of 0.5-10Hz. Finally, for the missing data in the processed signal, peak detection and interpolation were used to repair the missing data, resulting in a denoised blood flow signal. First, a moving average filter with a window size of 3 sampling values ​​is used to perform mean filtering on the acquired raw temperature signal to smooth out the interference caused by instantaneous temperature fluctuations. Then, the ambient temperature baseline value acquired during the system initialization phase is retrieved, and the deviation between the acquired temperature at each temperature detection point and the ambient temperature baseline value is calculated. According to the formula: actual temperature = acquired temperature - ambient temperature deviation, environmental compensation is performed on the mean-filtered temperature signal to eliminate the influence of ambient temperature changes on the measurement of the foot body temperature, and the corrected temperature signal is obtained. The clock synchronization unit of the in-shoe data acquisition module is activated to read the timestamps bound during the acquisition of four modal data: pressure, temperature, vibration response, and blood flow. Using the timestamp of the pressure signal as a reference, the timestamps of the temperature, vibration response, and blood flow signals are compared. The missing sampling points in each modal data are filled in by linear interpolation, so that the time axes of the four modal data are completely aligned, forming a spatiotemporally unified multimodal data matrix. By comparing the timestamps of each modality's data with the data acquisition order after synchronization, if the timestamp of a certain modality's data deviates from the timestamps of other modal data by more than 1ms, it is determined to be misaligned data caused by transmission delay, and the misaligned data is directly discarded. For pressure signals, extreme values ​​exceeding 300kPa are discarded; for temperature signals, abnormal values ​​deviating from the normal temperature range after environmental compensation by more than 5℃ are discarded; for blood flow signals, noise values ​​with pulse wave period fluctuations exceeding 20% ​​are discarded. If a sensor experiences a signal interruption and the data from adjacent sensor units cannot be supplemented by interpolation, or if the proportion of abnormal data in a certain modality exceeds 10%, the valid data already acquired for that modality is retained, and the abnormal data area is marked to avoid abnormal data affecting subsequent analysis.

5. The method for detecting neuropathic and vascular lesions based on a wearable device according to claim 1, characterized in that: The extraction of peak value, duration, and contact area characteristics from pressure data; the extraction of temperature difference and heat zone distribution characteristics from temperature data; the extraction of sensing pass rate, average reaction time, and sensing threshold level from vibration response data; and the extraction of pulse wave period, amplitude, and relative blood flow values ​​from blood flow data include: A 1-minute time period for feature extraction is determined. All data collected by pressure sensing units in each detection area of ​​the sole within this period are iterated. For each detection area, the maximum value among all pressure values ​​is selected and recorded as the pressure peak value of the area. All sampling times in the area where the pressure value exceeds 50 kPa are counted, and the cumulative duration of these times is calculated and recorded as the pressure duration. The number of sensing units in the area where the pressure value is greater than 0 kPa is counted. Combined with the preset coverage area of ​​each sensing unit, the total coverage area of ​​all sensing units that meet the conditions is calculated and recorded as the contact area. The above operations are repeated to complete the feature extraction of pressure peak value, duration, and contact area for all detection areas of the sole. The temperature data collected from symmetrical detection points on both feet is iterated. For each pair of symmetrical detection points, the absolute value of the temperature difference within a 1-minute acquisition cycle is calculated to generate a dataset of detection points and temperature differences. The normal temperature range is set to 32-36℃. The temperature data of all temperature detection points are iterated, and the areas where the temperature exceeds this range are marked to form a thermal zone distribution record. At the same time, the temperature difference between different detection points on the same foot is calculated to analyze the temperature distribution differences within the same foot and supplement and improve the temperature feature extraction results. The vibration stimulation and feedback results of all detection points on the sole of the foot were statistically analyzed to determine the number of positive detection points and the total number of detection points. The perception pass rate was calculated using the formula (number of positive detection points / total number of detection points) × 100%. For all positive detection points, the perception delay time was extracted, and the arithmetic mean of these delay times was calculated to obtain the average reaction time. Based on the reaction time, the perception threshold levels were divided as follows: ≤0.5 seconds was Level 1 (normal), 0.5-1 second was Level 2 (mild reduction), 1-2 seconds was Level 3 (moderate reduction), 2-3 seconds was Level 4 (severe reduction), and >3 seconds or negative perception was Level 5 (extremely severe reduction). This completed the perception threshold level classification for each detection point. The preprocessed microcirculatory blood flow PPG raw waveform data were analyzed. The peak detection algorithm was used to identify the systolic peak, diastolic peak, and dicrotic wave in the waveform. The time interval between two adjacent S waves was calculated and recorded as the pulse wave period. The difference between the peak value of the S wave and the waveform baseline was measured and recorded as the pulse wave amplitude. Based on the preset tissue optical parameters and combined with the Lambert-Beer law, the area integral of the PPG waveform was calculated to derive the relative value of blood flow. At the same time, the amplitude ratio of the S wave to the D wave and the amplitude variation coefficient obtained by the sliding window method were calculated to supplement the dimensions of blood flow feature extraction.

6. The method for detecting neuropathic and vascular lesions based on a wearable device according to claim 1, characterized in that: The construction of a spatiotemporally unified multimodal data matrix enhances features through convolution enhancement, temperature gradient calculation, perception, stimulus fitting, and waveform entropy analysis. It establishes a modality and pathological association network to generate cross-modal fusion features, including: Based on the timestamps bound to each modal data, the missing sampling points are filled in using linear interpolation to ensure that the time axes of pressure, temperature, vibration response, and blood flow data are completely synchronized. Then, the sensor coordinate mapping table preset by the insole and upper is retrieved to establish the spatial correspondence between the plantar pressure sensing points, temperature detection points, vibration stimulation points and the microcirculation monitoring area of ​​the dorsum of the foot. The synchronized multimodal data are integrated according to the dimensions of time frame × spatial point × modal feature to form a spatiotemporally unified multimodal data matrix. For the pressure features in the data matrix, a 1×3 temporal convolution kernel is used to traverse the data of each time frame to extract the temporal trend of the pressure peak changing with time; then a 3×3 spatial convolution kernel is used to cover each spatial point to highlight the features of the abnormal pressure area that has been subjected to continuous force for more than 60 seconds; finally, an attention mechanism is introduced to give higher weight to the features of long-term pressure points to strengthen the expression of abnormal pressure features. Based on the temperature characteristics in the data matrix, the rate of change of temperature difference between the left and right foot symmetrical detection points over time is first calculated to reflect the development speed of temperature anomalies. Then, a heat map diffusion algorithm is used to analyze the spatial spread trend of the heat zone with the temperature exceeding 32-36℃ as the center, generate heat zone diffusion characteristics, and supplement dynamic anomaly information in addition to static temperature difference. The reaction time and corresponding stimulus intensity parameters of the vibration response data in the data matrix are extracted. With stimulus intensity as the independent variable and reaction time as the dependent variable, a quadratic function curve is fitted. The curve slope and inflection point are extracted by curve analysis to form the vibration perception enhancement feature. For PPG waveform data in the data matrix, the waveform is divided into segments according to a preset time window, and the information entropy of the PPG waveform in each window is calculated. By the difference in information entropy values, the disordered regions of the waveform are identified, and the information entropy value is used as a blood flow enhancement feature to supplement the deficiencies of traditional pulse wave amplitude and period features. Calculate the mutual information value between the enhancement features of each modality, and screen the highly correlated feature pairs with mutual information values ​​> 0.6; perform fusion operation on the highly correlated feature pairs: according to the feature product weighting and difference penalty strategy, that is, the correlated feature value = (feature A × feature B × adaptive correlation weight) - (|duration of feature A - stability duration of feature B| × penalty coefficient), where the weight and penalty coefficient are determined by training with clinical data; Using spatial point-time period-pathological correlation as a three-dimensional dimension, the fused feature pairs are classified according to spatial point and time period. The degree of multimodal feature synergy anomaly of each spatial point in different time periods is calculated. A three-dimensional matrix of modality-pathological correlation is constructed to quantify the level of synergy anomaly of cross-modal features and finally generate a cross-modal fused feature set.

7. The method for detecting neuropathic and vascular lesions based on a wearable device according to claim 1, characterized in that: The method employs a dual-branch network combined with a multi-task learning model to simultaneously output neuropathic grading and vascular lesion risk scores, including: A dual-branch network with an immediate feature branch and a temporal trend branch was constructed. The immediate feature branch adopted a CNN network, which input the cross-modal correlation features of the current acquisition cycle to extract the immediate pathological abnormalities such as the instantaneous pressure peak exceeding the standard and the sudden drop in blood flow. The temporal trend branch adopted a BiLSTM network, which input the historical feature sequences of the last 5 acquisition cycles to capture the disease development trend of the gradual increase in temperature difference and the continuous prolongation of vibration perception reaction time. Modal attention automatically assigns weights to each modality. When assessing neuropathology, it increases the weights of vibration and temperature modalities, and when assessing vascular lesions, it increases the weights of microcirculation and pressure modalities. Pathological attention is based on the clinical pathological association rule that the correlation between neuropathology and vibration and temperature is higher than that between neuropathology and pressure, thus strengthening the contribution of key pathological features. By fusing the outputs of the two branches through a fully connected layer, a preliminary multimodal fusion feature vector with a dimension of 1×128 is generated.

8. The method for detecting neuropathic and vascular lesions based on a wearable device according to claim 7, characterized in that: The following criteria were used to construct a rule constraint matrix: a temperature difference of >2℃ between the left and right feet indicates inflammation or perfusion abnormality; a vibration sensation reaction time of >2 seconds indicates moderate neuropathy; and a blood flow of <1.5ml・min⁻¹・cm⁻² indicates insufficient blood supply. These criteria were converted into numerical constraint thresholds. If the evaluation result corresponding to the initial multimodal fusion feature vector conflicts with the constraint rules, the feature reweighting mechanism is activated to increase the weight of the conflicting modal features and recalculate the fusion result. The normal state data collected during the patient's first wearing of the device is incorporated as individual baseline data. The current fusion features are compared with the individual baseline to generate an individual relative abnormality index. The neuropathy grading task uses the Softmax classifier to map the calibrated fusion features to 5 levels: level 0, level 1, level 2, level 3, and level 4. The classification boundary is optimized through training with 50 clinical samples. The vascular lesion risk scoring task uses a regression model to output a risk value of 0-100. The scoring dimensions cover blood flow adequacy, vascular elasticity, blood flow stability, and cross-modal association abnormality. The consistency between the neuropathic grading and the vascular lesion risk score is verified. If the correspondence between the two does not conform to clinical patterns, the model is recalculated.

Images