Diabetic foot super-early risk warning method based on sweat gland nerve timing characteristics

Abstract

The present application relates to the technical field of wearable artificial intelligence monitoring, and particularly relates to a diabetic foot super-early risk warning method based on sweat gland nerve timing characteristics, which comprises the following steps: step 1, adaptive collection of sweat gland timing data in multiple zones of the sole; step 2, hierarchical extraction and noise filtering of timing characteristics; step 3, dynamic incremental modeling of individual sweat gland physiological baseline; step 4, lightweight LSTM timing neural network risk grading reasoning; and step 5, local storage and directional pushing of graded intervention information. The method realizes medical target innovation and architecture collaborative innovation, greatly prolongs the clinical intervention window period, has the effect of high precision and low false alarm, can perform edge local reasoning, guarantees data privacy compliance, has a four-level grading system, can form a complete chronic disease management closed loop, and has the characteristics of light weight and low power consumption, is convenient for direct integration into an existing intelligent monitoring shoe product system, and has good industrial transformation prospects.

Description

Technical Field

[0001] This invention relates to the field of wearable artificial intelligence monitoring technology, and in particular to a method for early warning of diabetic foot risk based on the temporal characteristics of sweat secretion nerves. Background Technology

[0003] The "Expert Consensus on the Diagnosis and Treatment of Diabetic Neuropathy (2021 Edition)" and modern endocrine pathology have clearly stated that the earliest pathological change in diabetic foot is not mechanical damage caused by external pressure on the sole of the foot, but rather unmyelinated type C small fiber autonomic neuropathy. These nerves control the sweat gland secretion function of the foot, and after nerve damage, characteristic abnormalities will appear in the sweating rhythm and the temporal fluctuations of temperature and humidity in the foot. This physiological signal abnormality appears 3 to 12 months earlier than visible damage such as redness, swelling, blisters, and ulcers of the foot, and is currently the only known early biomarker with a reversible intervention window.

[0004] Current mainstream global technologies for monitoring diabetic foot have the following inherent limitations: The first category consists of plantar pressure monitoring devices (represented by commercial products such as Orpyx and Podimetrics). These devices only collect plantar mechanical pressure signals, can only identify existing tissue damage, have a severely delayed warning system, lack the ability to identify early neuropathy, and cannot achieve proactive prevention.

[0005] The second category is ordinary temperature and humidity monitoring devices. These devices only collect instantaneous temperature and humidity values, use fixed thresholds to judge abnormalities, and do not establish an individualized physiological baseline for the user. They cannot effectively distinguish interference signals caused by damp shoes and socks, walking friction, and changes in ambient temperature, resulting in an extremely high false alarm rate in home scenarios. At the same time, their analysis logic does not incorporate the pathological mechanism of sweat neuropathy, and cannot capture the slow and hidden multi-day neurodegenerative trend. Summary of the Invention

[0006] The main objective of this invention is to overcome the shortcomings of existing technologies, such as delayed early warning, high false alarm rate, lack of clinical and pathological evidence, and missing grading standards, and to provide a method for early warning of diabetic foot risk based on the temporal characteristics of sweat secretion nerves.

[0007] The technical solution adopted by this invention to achieve its technical objective is: a method for early risk warning of diabetic foot based on the temporal characteristics of sweat secretion nerves, comprising: During the user's daily waking hours, a flexible temperature and humidity sensor array integrated into the integrated smart monitoring shoe collects long-term time-series data on sweat secretion covering multiple zones of the sole, including the heel, first metatarsal bone, fifth metatarsal bone, toes, and arch area. The time-series data includes time-series information on humidity, temperature, regional gradient, and symmetry difference between the left and right feet for each zone. An individual sweating physiological baseline is constructed based on continuous historical valid time-series data of the user, and the baseline is iteratively updated according to the newly added valid data each day to adapt to individual physiological drift. Mahalanobis distance was used to calculate the deviation residual sequence between real-time temporal features and the individual's sweating physiological baseline; The deviated residual sequence is input into a lightweight long short-term memory network deployed at the edge. The network captures the autonomic degeneration trend based on long-term temporal associations on a multi-month scale and outputs a four-level risk classification that matches the clinical sweating function assessment standards.

[0008] Preferably, before constructing the individual sweating physiological baseline, multi-dimensional noise filtering is performed on the long-term data, and the isolated forest algorithm is used to remove outliers caused by instantaneous friction of gait, sudden changes in environmental temperature and humidity, and dry and wet staining of shoes and socks, so as to retain abnormal fluctuations in neurogenic sweating. The outliers do not participate in baseline updates and risk inference.

[0009] Preferably, the collection of long-term time-series data on sweat secretion in multiple zones of the sole includes: adaptively switching the sampling frequency according to the human movement state, using low-frequency sampling of 0.1 to 0.5 Hz for walking, and high-frequency sampling of 1 to 5 Hz for static or sedentary states; automatically detecting the state of being away from the foot during the user's sleep and bathing periods at night, pausing the collection and marking it as invalid segments.

[0010] Preferably, the construction of an individual sweating physiological baseline includes: collecting foot time-series data covering the waking hours of each day within a continuous monitoring period of at least 7 days to generate an initial baseline, and performing incremental self-calibration iterations with newly added valid time-series data each day.

[0011] Preferably, the lightweight long short-term memory network is pruned and quantized for compression to adapt to local deployment on low-power embedded processors in wearable devices, and the power consumption for completing full-time feature inference in a single day is less than 5mWh.

[0012] Preferably, the original foot physiological time-series data is encrypted and stored in the encrypted storage unit of the integrated smart monitoring shoe and is not transmitted to the outside; only the four-level risk classification and corresponding intervention instructions are uploaded to the outside.

[0013] Preferably, the risk classification is uploaded via a Bluetooth and narrowband IoT dual-mode communication module, and targeted intervention push is performed based on the classification results: a home care plan is pushed for mild warnings, a community follow-up reminder is pushed for moderate to high risk, and a specialist referral prompt is pushed for severe emergencies.

[0014] Preferably, the threshold for the four-level risk grading is determined by fitting the subject's baseline deviation distribution to the Youden index of the subject's operating characteristic curve based on the principle of maximizing the deviation.

[0015] The working principle of this early warning method for diabetic foot based on the temporal characteristics of sweat secretion nerves is as follows: by constructing a three-level serial architecture consisting of front-end physical perception, mid-level individualized benchmark alignment, and back-end long-term trend inference, the method can accurately identify the early risk of diabetic foot.

[0016] First, a flexible temperature and humidity sensor array integrated into the smart monitoring shoe continuously collects long-term data on foot sweat secretion in five key areas of the sole, covering the daily waking hours, in order to capture the characteristic abnormal sweat gland secretion rhythm caused by unmyelinated C-type small fiber autonomic neuropathy.

[0017] Secondly, the system establishes a user-specific healthy sweating physiological baseline through a cold start phase of at least 7 days, and uses daily new data for dynamic incremental iterative calibration to adapt to individual physiological drift; at the same time, the isolated forest algorithm is used to remove pseudo-abnormal noise such as motion friction and environmental changes to ensure data purity.

[0018] Ultimately, the system inputs the Mahalanobis distance deviation residual sequence between real-time temporal features and individual baselines into a lightweight long short-term memory network that has been pruned and quantized. This network captures slow and hidden neurodegenerative trends on a multi-month scale and outputs a four-level standardized risk rating that matches clinical treatment guidelines, thereby achieving an intelligent closed loop from simple data collection to pathological trend determination.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This method for early warning of diabetic foot risk based on the temporal characteristics of sweat secretion nerves achieves medical target innovation. It uses abnormal sweat secretion timing of unmyelinated C-type small fiber autonomic nerves as the warning target, which can identify reversible lesions 3 to 12 months before the skin of the foot breaks down, and significantly prolong the clinical intervention window.

[0020] This method for early risk warning of diabetic foot based on the temporal characteristics of sweat secretion nerves achieves synergistic architectural innovation. By deeply coupling individual dynamic baseline calibration with LSTM, it forms a three-level cascaded architecture of "individualized dynamic incremental baseline calibration, Mahalanobis distance deviation residual sequence, and LSTM long-term trend analysis," solving the problem of distinguishing between individual physiological drift and pathological degeneration trends in home monitoring. Furthermore, this architecture improves the recognition sensitivity to 92.7% and reduces the false alarm rate to 5.7%, achieving high accuracy and low false alarm performance, with a synergistic effect far exceeding the sum of the individual applications of each module.

[0021] This method for early risk warning of diabetic foot based on the temporal characteristics of sweat secretion nerves enables local inference at the edge, ensuring data privacy compliance. By completing all AI inference operations locally at the edge of the wearable device, the raw physiological time-series data is encrypted and stored locally throughout the process, with only risk classification labels uploaded, fully meeting the requirements of medical and health data privacy protection regulations.

[0022] This method for early risk warning of diabetic foot based on the temporal characteristics of sweat secretion nerves has a four-level classification system, which can form a complete closed loop for chronic disease management. The four-level risk classification matches the sweat secretion function assessment standard in the "Expert Consensus on the Diagnosis and Treatment of Diabetic Neuropathy (2021 Edition)", forming a complete closed loop from continuous home monitoring to community follow-up and then to referral to tertiary specialist hospitals. It is suitable for multiple application scenarios such as home care, elderly care institutions, and primary healthcare consortia.

[0023] This method for early risk warning of diabetic foot based on the temporal characteristics of sweat secretion nerves is lightweight and low-power, showing promising prospects for industrial application. The LSTM model, after pruning and quantization compression, consumes less than 5mWh of power per day for full-time inference, requiring no additional computing hardware and facilitating direct integration into existing smart monitoring shoe product systems. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall steps of the method of the present invention.

[0025] Figure 2 This is a schematic diagram showing the layout of the five major sensor zones on the sole of the foot.

[0026] Figure 3 This is a schematic diagram of the layer structure of a lightweight LSTM temporal neural network.

[0027] Figure 4 A logical flowchart for the intervention push of the four-level risk classification.

[0028] Figure 5 This is a bar chart comparing the sensitivity, false alarm rate, and warning duration of the present invention with those of existing technologies. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0030] In the description of this invention, it should be noted that when an element is referred to as being "fixed to" or "set on" another element, it can be directly on or indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to or indirectly connected to the other element.

[0031] In the description of this invention, it should be noted that the terms "center," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. "Several" means one or more, unless otherwise explicitly specified.

[0032] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" 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 or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. Example 1:

[0033] Please see Figures 1-4 A method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves is proposed, which is applied to the screening of diabetic foot risk for non-therapeutic purposes. The method includes the following steps: Step 1: Adaptive acquisition of temporal data on sweat secretion in multiple zones of the sole of the foot: 1.1 A flexible temperature and humidity sensor array is deployed in five high-risk disease areas: the sole and heel, the first metatarsal bone, the fifth metatarsal bone, the toes, and the arch of the foot, and integrated inside the integrated smart monitoring shoe; the flexible temperature and humidity sensor array is an integrally sealed and waterproof structure, integrated inside the integrated smart monitoring shoe.

[0034] 1.2 The device adaptively switches the sampling frequency according to the human body's movement state: low frequency sampling is used when the user is walking to reduce hardware power consumption, and the sampling frequency is automatically increased when the user is stationary or sitting for a long time; in the adaptive switching of sampling frequency, the sampling frequency for walking is 0.1 to 0.5 Hz, and the sampling frequency for stationary and sitting for a long time is 1 to 5 Hz.

[0035] 1.3 Continuously collect long-term data streams covering the complete diurnal rhythm, with the collection period covering the user's daily waking activity phase (average effective collection time ≥ 16 hours per day). Each data stream contains time-series information on real-time humidity, temperature, temperature and humidity gradient differences in each zone, and left and right foot symmetry differences.

[0036] 1.4 During the user's sleep and bathing periods at night, the device automatically detects the foot-off state, pauses data collection and enters low-power standby. The corresponding time period is marked as an invalid segment in the data stream and does not participate in subsequent baseline construction and model inference.

[0037] Step 2, Layered Extraction of Temporal Features and Noise Filtering: Perform multi-layered feature extraction on the acquired long-term temporal data stream, simultaneously removing pseudo-anomaly noise caused by the environment and motion. Specifically, this includes: 2.1 Steady-state baseline feature extraction: Extract the long-term average sweat secretion and diurnal secretion steady-state baseline values ​​of a single user to construct individual basic physiological characteristics.

[0038] 2.2 Dynamic temporal feature extraction: Extract humidity fluctuation cycle, fluctuation decay rate, diurnal sweat secretion difference, and regional humidity gradient temporal offset.

[0039] 2.3 Extraction of pathological features: Identify neurogenic sweating abnormalities such as prolonged local low humidity, disruption of sweating rhythm, and persistent regional temperature differences.

[0040] 2.4 Multi-dimensional noise filtering: The isolated forest algorithm is used to identify outliers caused by instantaneous friction during gait, sudden changes in environmental temperature and humidity, and dry and wet stains of shoes and socks. It distinguishes interference signals and retains only the real sweating time-series fluctuations caused by autonomic neuropathy. Outliers do not participate in baseline updates and risk inference.

[0041] Step 3: Dynamic incremental modeling of individual sweating physiological baseline: 3.1 Cold Start Baseline Construction: Collect foot time-series data covering the awake time of each day during a continuous 7-day monitoring period (average effective collection time ≥16 hours per day, and effective collection time no less than 14 hours per day) to generate a user-specific initial healthy sweating physiological baseline.

[0042] 3.2 Incremental self-calibration iteration: The baseline model is iteratively updated with newly added valid time-series data every day to adapt to changes in users' seasonal changes, weight changes, and daily exercise habits.

[0043] 3.3 Quantitative Calculation of Damage Deviation: Mahalanobis distance is used to calculate the degree of deviation between real-time temporal features and individual health baseline, quantifying the severity of unmyelinated Class C small fiber nerve damage; the sensitivity for identifying early small fiber autonomic neuropathy is not less than 90%, and the overall false alarm rate in home scenarios after noise filtering is less than 6%. The sensitivity and false alarm rate values ​​are derived from experimental verification, see Example 3 and Table 1 for details.

[0044] The Mahalanobis distance ≤ 2 is considered normal, 2 < Mahalanobis distance ≤ 5 is considered mild nerve damage, 5 < Mahalanobis distance ≤ 10 is considered moderate sweating dysfunction, and Mahalanobis distance > 10 is considered a precursor to severe diabetic foot ulcers. The Mahalanobis distance grading thresholds ≤ 2, 2~5, 5~10, and > 10 are determined by fitting the baseline deviation distribution of the subjects to the maximum Youden index of the ROC curve.

[0045] Step 4: Lightweight LSTM (Long Short-Term Memory) Temporal Neural Network Risk Grading Inference: The power consumption of the lightweight LSTM temporal neural network to complete the full-time-series feature inference in a single day is less than 5mWh. The power consumption value comes from experimental verification. See Example 3 and Table 1 for details.

[0046] 4.1 Model Training Dataset Construction: The training set was constructed using time-series samples of desensitized diabetic foot subjects from a tertiary hospital. The samples were divided into four categories: normal individuals without nerve damage, individuals with mild small fiber nerve damage, individuals with moderate sweating dysfunction, and individuals with severe diabetic foot ulcer precursors. The dataset was divided into training and test sets in a 7:3 ratio.

[0047] 4.2 Lightweight Model Processing: Pruning and quantization compression are performed on the complete LSTM network to reduce the number of model parameters and computational overhead, making it suitable for local deployment on low-power embedded processors in wearable devices.

[0048] 4.3. Temporal Feature Input Inference: The Mahalanobis distance deviation residual temporal sequence obtained in step 3 is used as the input feature of the LSTM network, which captures the slow and hidden autonomic degeneration trend on a multi-month scale (e.g., 3 to 12 months).

[0049] 4.4 Standardized Four-Level Risk Classification Output: The network outputs a four-level standardized risk rating, namely normal, mild nerve damage warning, moderate sweating disability high risk, and severe diabetic foot ulcer precursor emergency. The classification standard matches the sweating function assessment standard in the "Expert Consensus on Diagnosis and Treatment of Diabetic Neuropathy (2021 Edition)".

[0050] Step 5: Local storage and targeted delivery of tiered intervention information: 5.1 The original foot physiological time series data is encrypted and stored in the integrated intelligent monitoring shoe's encrypted storage unit throughout the entire process, and is not transmitted to the outside.

[0051] 5.2 Only the Level 4 risk rating and corresponding intervention instructions are uploaded to the cloud and user terminal via the wireless communication module; the cloud is equipped with a chronic disease management platform, which supports medical institutions to view the risk rating data of patients under their jurisdiction in batches, and the data interface is standardized to connect with the county-level medical consortium's electronic health record system.

[0052] 5.3 The wireless communication module adopts Bluetooth and narrowband IoT dual-mode communication (BLE-NB-IoT dual-mode communication). It connects to the user's mini-program via Bluetooth at close range and automatically switches to cellular network to upload early warning and classification information at long distances.

[0053] The logic of tiered intervention is as follows: Mild alerts will be sent out with home foot moisturizing and foot massage care plans, and continuous monitoring and reminders will be provided. For those at moderate to high risk, a reminder to schedule a follow-up visit with a community family doctor will be sent, and the message will also be sent to the primary care chronic disease management platform. In cases of severe emergency, referral requests are sent to the endocrinology and wound repair departments of tertiary hospitals, and high-risk warning records are simultaneously pushed to the back-end systems of medical institutions.

[0054] Specifically, in home monitoring scenarios, when users wear the integrated smart monitoring shoes during their daily daytime activities, the system will automatically start and continuously collect time-series data on foot temperature and humidity. It will also adaptively switch the sampling frequency based on walking (low frequency) or still / sitting (high frequency) status, with an average effective collection time of no less than 16 hours per day. When sleeping or bathing at night, the device will automatically detect when the foot is removed and enter low-power standby mode.

[0055] Every day at dawn, the device's built-in embedded processor automatically performs localized AI inference: first, it filters out noise from the new valid data added the previous day, compares it with the user's existing health baseline, calculates the Mahalanobis distance deviation residual sequence, and then inputs the sequence into the deployed lightweight LSTM network for hierarchical inference.

[0056] Ultimately, the raw physiological data is encrypted and stored locally throughout the entire process. Only the output of the four-level risk rating (normal / mild / moderate-high risk / severe emergency) and the corresponding standardized intervention instructions are uploaded to the cloud, user terminal or medical institution backend through the BLE-NB-IoT dual-mode communication module. This triggers home care reminders, community family doctor follow-up appointments or tertiary hospital specialist referral prompts according to the risk level, forming a continuous management process of "daily monitoring - edge reasoning - graded early warning - targeted intervention". Example 2:

[0057] Please see Figures 1-4Based on the above embodiments, the method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves further includes an integrated intelligent monitoring shoe with a flexible temperature and humidity sensor array arranged in five areas: the sole, heel, first metatarsal bone, fifth metatarsal bone, toe, and arch. The shoe also has an embedded processor, an encrypted storage unit, and a wireless communication module.

[0058] When the embedded processor executes a computer program, it performs the following steps: Step 1: Adaptive Acquisition of Multi-Zone Sweat Secretion Data from the Sole of the Foot: The device adaptively switches the sampling frequency according to the user's movement state. Low-frequency sampling is used when the user is walking to reduce hardware power consumption, while the sampling frequency is automatically increased when the user is stationary or sitting for a long time. The device continuously collects long-term data streams covering the complete circadian rhythm, with the collection period covering the user's daily waking activity phase, and the average effective collection time per day is ≥16 hours. During the user's sleep and bathing periods at night, the device automatically detects the foot-off state, pauses the collection, and enters low-power standby. The corresponding time period is marked as invalid segments in the data stream.

[0059] Step 2: Layered Extraction and Noise Filtering of Temporal Features: Multi-layer feature extraction and removal of pseudo-anomalies are performed on the collected long-term temporal data stream. Specifically, this includes: extracting the long-term average sweat secretion of a single user and the diurnal secretion steady-state baseline value to construct the individual's basic physiological characteristics; extracting humidity fluctuation period, fluctuation decay rate, diurnal sweat secretion difference, and temporal offset of regional humidity gradient; identifying neurogenic sweat secretion abnormalities such as local long-term low humidity, break in sweat secretion rhythm, and continuous regional temperature difference; and using the isolated forest algorithm to remove outliers caused by instantaneous friction of gait, sudden changes in environmental temperature and humidity, and dry and wet staining of shoes and socks. Outliers are not included in subsequent baseline updates and risk inference.

[0060] Step 3: Dynamic Incremental Modeling of Individual Sweat Secretion Physiological Baseline: Collect foot time-series data covering the waking hours of each day within a continuous 7-day monitoring period (average effective collection time ≥16 hours per day, and effective collection time no less than 14 hours per day) to generate an initial healthy sweat secretion physiological baseline for each user; iteratively update the baseline model with newly added effective time-series data each day to adapt to changes in the user's seasons, weight, and daily exercise habits; use Mahalanobis distance to calculate the degree of deviation between real-time time-series characteristics and the individual's healthy baseline to quantify the severity of unmyelinated type C small fiber nerve damage.

[0061] Step 4: Lightweight LSTM Temporal Neural Network Risk Grading Inference: A training set was constructed using temporal samples from desensitized diabetic foot subjects in a tertiary hospital. The samples were divided into four categories: normal individuals without nerve damage, individuals with mild small fiber nerve damage, individuals with moderate sweating dysfunction, and individuals with severe diabetic foot ulcer precursors. The complete LSTM temporal network was pruned and quantized for local deployment on a low-power embedded processor for wearable devices. The Mahalanobis distance deviation residual temporal sequence calculated in Step 3 was used as the input feature of the LSTM network, which captured the slow and hidden autonomic nerve degeneration trend over a multi-month scale. The network output a four-level standardized risk rating: normal, mild nerve damage warning, moderate sweating dysfunction high risk, and severe diabetic foot ulcer precursor emergency. The grading standard matches the sweating function assessment specifications in the "Expert Consensus on the Diagnosis and Treatment of Diabetic Neuropathy (2021 Edition)".

[0062] Step 5: Local Storage and Targeted Push of Tiered Intervention Information: The original foot physiological time-series data is encrypted and stored in the encrypted storage unit of the integrated smart monitoring shoe and is not transmitted externally; only the four-level risk rating and corresponding intervention instructions are uploaded to the cloud and user terminal through the wireless communication module; the tiered intervention logic is as follows: mild warning pushes home foot moisturizing and foot massage care plan; moderate to high risk pushes community family doctor follow-up appointment reminder; severe emergency pushes referral prompts to the endocrinology department and wound repair specialist department of tertiary hospital and high-risk warning ledger.

[0063] The solution in this embodiment can be selectively combined with solutions in other embodiments. Example 3:

[0064] Please see Figures 1-5 Based on the above embodiments, the specific implementation method for early risk warning of diabetic foot based on the temporal characteristics of sweat secretion nerves is as follows: S1. Home-based monitoring process for high-risk groups of diabetes: (1) When the user wears the integrated smart monitoring shoes during daily activities, the device starts a 7-day cold start data collection process, covering all waking hours from when the user puts the shoes on in the morning until when the user takes them off at night (the average effective collection time per day is ≥16 hours, and the effective collection time per day is not less than 14 hours), and obtains the temperature and humidity time series data of the five zones of the sole of the foot. During nighttime sleep and bathing periods, the device automatically detects foot-free status, pauses data collection, and enters low-power standby. The corresponding time period is marked as an invalid segment in the data stream and does not participate in baseline construction and model inference.

[0065] (2) The device adaptively switches the sampling frequency: low frequency sampling (0.2Hz) is used in the walking state to reduce power consumption, and high frequency sampling (2Hz) is used in the sitting state to collect the sweating timing signal; noise filtering is performed simultaneously to remove false abnormal signals caused by damp shoes and socks and walking friction.

[0066] (3) Start local lightweight LSTM time series inference at a fixed time in the early morning of each day, compare the newly added effective time series data of the previous day with the healthy baseline to calculate the Mahalanobis distance deviation residual sequence, input the residual sequence into the LSTM network to complete trend inference, and iteratively update the individual baseline.

[0067] (4) The model outputs a four-level risk label, and only the risk level and corresponding nursing / referral instructions are uploaded to the user's family member's mini-program; the original foot time series data is encrypted and stored in the encrypted storage unit inside the shoe and is not transmitted to the outside.

[0068] (5) If the injury is determined to be mild nerve damage, the mini-program will push a home foot care plan; if the injury is determined to be a severe ulcer precursor, a referral reminder to the endocrinology department of a tertiary hospital will be pushed simultaneously.

[0069] S2, Application of batch chronic disease management in county-level medical consortia: (1) All diabetic patients in the jurisdiction are uniformly equipped with the integrated intelligent monitoring shoes described in this invention. The device continuously collects sweating time sequence and completes AI risk inference locally.

[0070] (2) The cloud-based chronic disease management platform receives the four-level risk classification data of patients in the jurisdiction in batches every day and automatically generates a high-risk patient ledger.

[0071] (3) Community family doctors arrange home visits for people at medium and high risk based on the records, and directly send specialist referral forms to patients at high risk.

[0072] (4) All risk data are standardized and connected to the county residents' electronic health record system to complete the normalized digital management and control of diabetic complications.

[0073] S3, a baseline-controlled validation experiment with 200 subjects: (1) Grouping of experimental equipment: Experimental group: Five-zone flexible temperature and humidity integrated smart monitoring shoe that integrates the sweat secretion neural time sequence LSTMAI model of this invention; Control group 1: Commercial plantar pressure diabetic foot monitoring device (Orpyx); Control group 2: Traditional fixed threshold temperature and humidity monitoring shoes, without dynamic baseline and without temporal deep learning network.

[0074] (2) Subject sample: Two hundred diabetic subjects were selected, with an average age of 52 ± 7 years. There were 50 subjects in each of the four groups: normal subjects without neuropathy, subjects with mild small fiber nerve damage, subjects with moderate sweating dysfunction, and subjects with severe prodromal diabetic foot ulcers. All subjects completed the clinical Sudoscan sweating nerve test at the same time as the basis for judging the actual nerve damage.

[0075] (3) Experimental conditions: During a continuous 7-day monitoring period, subjects wore the integrated smart monitoring shoes described in this invention throughout their daily activities, covering the entire waking period each day (average effective wearing time ≥ 16 hours per day, and continuous effective data collection for at least 14 hours per day). Subjects removed the devices during sleep and bathing at night; the system automatically detected the removal of the foot and paused data collection, marking the corresponding removal time segment in the data stream as invalid and not included in subsequent time-series analysis. All subjects maintained a uniform indoor temperature and humidity (25°C ± 2°C, relative humidity 50% ± 10%), a uniform daily activity level (8000 ± 2000 steps per day), and a uniform shoe and sock replacement frequency (cotton socks changed once daily). Uniform evaluation indicators included: sensitivity for early autonomic neuropathy recognition, overall false alarm rate in home scenarios, ulcer risk warning lead time, and daily model inference power consumption.

[0076] (4) Basis for determining the Mahalanobis distance threshold: The Mahalanobis distance grading thresholds (≤2, 2-5, 5-10, >10) were determined by fitting the ROC curve Youden exponent maximum principle based on the baseline deviation distribution of 200 subjects. The sensitivity and specificity corresponding to each threshold point met the clinically acceptable standard.

[0077] (5) The quantitative experimental results are shown in Table 1 below:

[0078] Table 1 shows the experimental results generated during the testing of this invention.

[0079] (6) Experimental conclusions: like Figure 5 As shown, this invention uses the temporal characteristics of sweat-secreting nerve small fiber lesions as the identification target, combined with a three-level cascade architecture of individualized dynamic incremental baseline calibration, Mahalanobis distance deviation residual sequence, and LSTM long-term trend analysis. Compared with the two existing mainstream monitoring technologies, it has significant technological progress in terms of early lesion identification accuracy (not less than 90%), early warning window period (3-12 months), home use stability (false alarm rate less than 6%), and low hardware power consumption (less than 5mWh), and can realize ultra-early reversible intervention of diabetic foot.

[0080] The solution in this embodiment can be selectively combined with solutions in other embodiments.

[0081] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of this invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of this invention, or equivalent structural, procedural, or functional transformations made using the description and drawings of this invention, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of protection of this invention.

Claims

1. A method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves, characterized in that, include: During the user's daily waking hours, a flexible temperature and humidity sensor array integrated into the integrated smart monitoring shoe collects long-term time-series data on sweat secretion covering multiple zones of the sole, including the heel, first metatarsal bone, fifth metatarsal bone, toes, and arch area. The time-series data includes time-series information on humidity, temperature, regional gradient, and symmetry difference between the left and right feet for each zone. An individual sweating physiological baseline is constructed based on continuous historical valid time-series data of the user, and the baseline is iteratively updated according to the newly added valid data each day to adapt to individual physiological drift. Mahalanobis distance was used to calculate the deviation residual sequence between real-time temporal features and the individual's sweating physiological baseline; The deviated residual sequence is input into a lightweight long short-term memory network deployed at the edge. The network captures the autonomic degeneration trend based on long-term temporal associations on a multi-month scale and outputs a four-level risk classification that matches the clinical sweating function assessment standards.

2. The method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves according to claim 1, characterized in that, Before constructing the individual sweating physiological baseline, multi-dimensional noise filtering is performed on the long-term data. The isolated forest algorithm is used to remove outliers caused by instantaneous friction of gait, sudden changes in environmental temperature and humidity, and dry and wet staining of shoes and socks, so as to retain abnormal fluctuations in neurogenic sweating. These outliers do not participate in baseline updates and risk inference.

3. The method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves according to claim 1 or 2, characterized in that, The collection of long-term time-series data on sweat secretion in multiple zones of the sole of the foot includes: adaptively switching the sampling frequency according to the human movement state, using low-frequency sampling of 0.1 to 0.5 Hz for walking, and high-frequency sampling of 1 to 5 Hz for static or sedentary states; automatically detecting the state of being away from the foot during the user's sleep and bathing periods at night, pausing the collection and marking it as invalid segments.

4. The method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves according to claim 1, characterized in that, The construction of an individual sweating physiological baseline includes: collecting foot time-series data covering the waking hours of each day for at least 7 consecutive days of monitoring to generate an initial baseline, and performing incremental self-calibration iterations with newly added valid time-series data each day.

5. The method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves according to claim 1, characterized in that, The lightweight long short-term memory network is pruned and quantized for compression to adapt to local deployment on low-power embedded processors in wearable devices, and consumes less than 5mWh to complete full-time feature inference in a single day.

6. The method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves according to claim 1, characterized in that, The original foot physiological time-series data is encrypted and stored in the encrypted storage unit of the integrated smart monitoring shoe and is not transmitted to the outside; only the four-level risk classification and corresponding intervention instructions are uploaded to the outside.

7. The method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves according to claim 6, characterized in that, The risk classification is uploaded via a dual-mode communication module using Bluetooth and narrowband IoT, and targeted interventions are pushed based on the classification results: home care plans are pushed for mild warnings, community follow-up reminders are pushed for moderate to high risk, and specialist referral prompts are pushed for severe emergencies.

8. The method for early risk warning of diabetic foot based on the temporal characteristics of sweating nerves according to claim 1, characterized in that, The thresholds for the four-level risk grading are determined by fitting the subject's baseline deviation distribution to the Youden index of the subject's operating characteristic curve, based on the principle of maximizing the deviation distribution.

Images