A method and system for dynamic monitoring and early warning of oral microenvironment

Abstract

The application discloses a kind of oral cavity microenvironment dynamic monitoring and early warning method and system, it is related to oral health monitoring technical field, the method includes: through sensor array real-time acquisition oral cavity microenvironment signal;Adaptive Kalman filtering and the fusion algorithm of dynamic sliding window median filtering are used to estimate pH signal, output steady-state pH estimated value;The steady-state pH estimated value of continuous time is constituted time series, input long short-term memory network model, and model training is carried out using compound loss function, and output predicted pH value;Based on individual baseline pH value, pH standard deviation and acidic cumulative duration, construct dynamic early warning threshold function, calculate comprehensive oral health risk index and generate graded early warning information;Every accumulation preset day's monitoring data, calculate predicted mean square error, and utilize gradient descent method to adjust long short-term memory network parameter.The application can realize continuous robust monitoring and individualized graded early warning of oral cavity pH, effectively reduce caries risk.

Description

Technical Field

[0001] This invention relates to the field of oral health monitoring technology, specifically to a method and system for dynamic monitoring and early warning of the oral microenvironment. Background Technology

[0002] The oral microenvironment, especially the dynamic changes in saliva pH, is closely related to the occurrence and development of oral diseases such as dental caries and dental erosion. Real-time monitoring of oral pH and early warning of acid attacks are of great significance for the prevention of dental caries in children.

[0003] However, existing technologies have the following shortcomings: traditional pH measurement devices cannot achieve continuous and imperceptible measurement in the dynamic oral environment and are easily affected by pulse interference such as food debris and sudden changes in saliva flow rate, resulting in low measurement accuracy; most existing oral pH monitoring products use fixed threshold warning methods and cannot dynamically adjust the warning threshold according to individual differences, leading to high false alarm or false alarm rates; most systems only passively alarm when the pH value is below the threshold and cannot predict acid peaks in the future, missing the opportunity for early intervention; the training loss function of existing warning models does not fully consider the special laws of oral pH changes, such as the consistency of trend direction and the degree of harm of acid peaks, causing the model prediction results to deviate from clinical needs.

[0004] Therefore, there is an urgent need for a method and system that can achieve continuous and robust monitoring of the oral microenvironment, personalized dynamic early warning, intelligent trend prediction, and closed-loop adaptive optimization capabilities. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for dynamic monitoring and early warning of the oral microenvironment, which addresses the shortcomings of the prior art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for dynamic monitoring and early warning of the oral microenvironment includes the following steps:

[0008] Step S1: Real-time acquisition of oral microenvironment signals through a sensor array integrated into the oral wearable device. The oral microenvironment signals include at least the original pH voltage signal, oral temperature signal, and oral activity state signal. The original pH voltage signal is preprocessed by low-pass filtering.

[0009] Step S2: The pH measurement value obtained by converting the preprocessed pH voltage signal is estimated by using a fusion algorithm of adaptive Kalman filtering and dynamic sliding window midpoint filtering, and the steady-state pH estimate is output.

[0010] Step S3: Construct a time series of steady-state pH estimates at consecutive time points, input it into a long short-term memory network model, and train the model using a composite loss function consisting of mean squared error loss, trend consistency loss, and acid peak penalty loss, and output the predicted pH value for future time points.

[0011] Step S4: Construct a dynamic early warning threshold function based on individual baseline pH value, pH standard deviation and acid accumulation duration, compare the predicted pH value with the dynamic early warning threshold, calculate the comprehensive oral health risk index, and generate graded early warning information according to the interval of the comprehensive oral health risk index.

[0012] Step S5: Accumulate monitoring data for a preset number of days, calculate the prediction mean square error, and adjust the parameters of the long short-term memory network using the gradient descent method to achieve closed-loop adaptive optimization.

[0013] Furthermore, in step S1, the cutoff frequency of the low-pass filter is 5 Hz.

[0014] Furthermore, step S2 specifically includes the following steps:

[0015] Step S2.1: Convert the preprocessed pH voltage signal into the original pH measurement value through linear calibration. The calibration formula is: pH measurement value equals the product of sensitivity calibration coefficient and voltage value plus bias calibration coefficient.

[0016] Step S2.2: The original pH measurement value is dynamically estimated using an adaptive Kalman filter algorithm to obtain the first estimated value; then, the first estimated value is smoothed using a sliding window midpoint filter with a length of 5 to obtain the steady-state pH estimate.

[0017] Furthermore, step S3 specifically includes the following steps:

[0018] Step S3.1: Construct the input feature vector from the steady-state pH estimates at consecutive time points;

[0019] Step S3.2: Input the input feature vector into the Long Short-Term Memory Network Model. The Long Short-Term Memory Network Model contains a Long Short-Term Memory layer with 64 units, a dropout layer with a dropout rate of 0.2, and a fully connected output layer, which outputs the predicted pH value at future time.

[0020] Step S3.3: Train the Long Short-Term Memory network model using a composite loss function, which is a weighted sum of mean squared error loss, trend consistency loss, and acid peak penalty loss.

[0021] Further, in step S3.3, the mean square error loss is equal to the reciprocal of the number of prediction steps multiplied by the sum of squares of the differences between the predicted pH value and the steady-state pH value at each prediction step; the trend consistency loss is equal to the reciprocal of the number of prediction steps minus one multiplied by the sum of squares of the differences in the signs of the trend changes between each adjacent prediction step, wherein the sign of the trend change is calculated by a sign function; the acidity peak penalty loss is equal to the reciprocal of the number of prediction steps multiplied by the sum of squares of the differences in the predicted pH values ​​at each prediction step that are lower than the acidity threshold of 5.5.

[0022] Furthermore, step S4 specifically includes the following steps:

[0023] Step S4.1: Construct a dynamic early warning threshold function. The dynamic early warning threshold function is equal to the individual baseline pH value minus the fluctuation coefficient multiplied by the standard deviation of the pH value in the past 24 hours, minus the cumulative acid attack coefficient multiplied by - minus an exponential function with the natural constant as the base, wherein the exponent of the exponential function is a negative decay rate constant multiplied by the cumulative duration of the pH value being below 5.5 in the past 6 hours.

[0024] Step S4.2: Calculate the comprehensive oral health risk index, which is equal to the positive part of the ratio of the difference between the dynamic warning threshold and the predicted pH value at the next moment to the dynamic warning threshold, multiplied by the first weight, plus the second weight multiplied by a subtracted S-shaped function, where the independent variable of the S-shaped function is the steepness coefficient multiplied by the difference between the total duration of pH value below 5.5 in the past 24 hours and the acid load tolerance baseline.

[0025] Step S4.3: Based on the numerical range of the comprehensive oral health risk index, a graded warning is issued. No warning is issued when the comprehensive oral health risk index is between 0 and 0.3. A first-level warning is issued when the comprehensive oral health risk index is between 0.3 and 0.7. A second-level warning is issued when the comprehensive oral health risk index is between 0.7 and 1.0.

[0026] Furthermore, step S5 specifically includes the following steps:

[0027] Step S5.1: Accumulate monitoring data for 7 days, and calculate the average squared difference between the predicted pH value and the steady-state pH value at all predicted times as the overall loss function;

[0028] Step S5.2: Use gradient descent to adjust all trainable parameters of the Long Short-Term Memory network to minimize the overall loss function.

[0029] A dynamic monitoring and early warning system for the oral microenvironment, used to implement any one of the methods for dynamic monitoring and early warning of the oral microenvironment, comprising:

[0030] The data acquisition module is used to acquire oral microenvironment signals in real time through a sensor array integrated into the oral wearable device and to perform low-pass filtering preprocessing on the raw pH voltage signal;

[0031] The signal processing module is used to extract steady-state pH estimates based on a fusion algorithm of adaptive Kalman filtering and dynamic sliding window midpoint filtering.

[0032] The time-series prediction module is used to input the steady-state pH estimation sequence into the long short-term memory network model, and train it using a composite loss function to output the predicted pH values ​​at multiple future time points.

[0033] The early warning decision module is used to construct a dynamic early warning threshold function, calculate a comprehensive oral health risk index, and generate graded early warning information.

[0034] The push and feedback module is used to push early warning content based on the graded early warning information and fine-tune the network parameters based on the gradient descent method.

[0035] Furthermore, the signal processing module internally includes:

[0036] The Kalman filter unit is used to perform adaptive Kalman filter estimation;

[0037] The median filtering unit is used to perform median filtering in a sliding window of length 5.

[0038] Furthermore, the time-series prediction module internally includes:

[0039] Long Short-Term Memory (LSTM) units are used to extract long-term dependence features of pH time series.

[0040] The composite loss calculation unit is used to calculate the weighted sum of mean squared error loss, trend consistency loss and acid peak penalty loss as a composite loss function during the initial training phase, and to guide the initial training of the model with this composite loss function.

[0041] The parameter fine-tuning unit is used to fine-tune and update the parameters of the long short-term memory network by using the prediction mean square error as the loss function and the gradient descent method after accumulating monitoring data for every 7 days.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. This invention employs a fusion algorithm of adaptive Kalman filtering and dynamic sliding window midpoint filtering to robustly estimate pH signals. This algorithm effectively suppresses impulse noise and Gaussian noise caused by food residue impact, sudden changes in saliva flow rate, etc., and significantly improves the accuracy and stability of pH measurement in a dynamic oral environment.

[0044] 2. This invention constructs a composite loss function consisting of mean squared error loss, trend consistency loss, and acid peak penalty loss to train an LSTM prediction model. The trend consistency loss penalizes cases where the predicted trend does not match the actual trend direction, while the acid peak penalty loss enhances the model's sensitivity to low pH danger zones, making the prediction results more consistent with the clinical characteristics of oral pH changes.

[0045] 3. This invention integrates a dynamic early warning threshold function based on individual baseline pH value, pH fluctuation and duration of acid accumulation, and calculates a comprehensive oral health risk index based on predicted pH value, thereby achieving personalized and forward-looking graded early warning and effectively reducing false alarm rate and missed alarm rate.

[0046] 4. This invention uses the prediction mean square error as the loss function to fine-tune the LSTM network parameters with gradient descent every 7 days of accumulated monitoring data. This allows the model to continuously adapt to individual user characteristics as usage time increases, thereby continuously improving prediction accuracy. There is no need to adjust the threshold coefficient online, which simplifies the system complexity and ensures long-term reliability. Attached Figure Description

[0047] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0048] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the system architecture according to an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] like Figure 1 As shown, a method for dynamic monitoring and early warning of the oral microenvironment includes the following steps:

[0052] Step S1: Real-time acquisition of oral microenvironment signals through a sensor array integrated into the oral wearable device. The oral microenvironment signals include at least the original pH voltage signal, oral temperature signal, and oral activity state signal. The original pH voltage signal is preprocessed by low-pass filtering.

[0053] Step S2: The pH measurement value obtained by converting the preprocessed pH voltage signal is estimated by using a fusion algorithm of adaptive Kalman filtering and dynamic sliding window midpoint filtering, and the steady-state pH estimate is output.

[0054] Step S3: Construct a time series of steady-state pH estimates at consecutive time points, input it into a long short-term memory network model, and train the model using a composite loss function consisting of mean squared error loss, trend consistency loss, and acid peak penalty loss, and output the predicted pH value for future time points.

[0055] Step S4: Construct a dynamic early warning threshold function based on individual baseline pH value, pH standard deviation and acid accumulation duration, compare the predicted pH value with the dynamic early warning threshold, calculate the comprehensive oral health risk index, and generate graded early warning information according to the interval of the comprehensive oral health risk index.

[0056] Step S5: Accumulate monitoring data for a preset number of days, calculate the prediction mean square error, and adjust the parameters of the long short-term memory network using the gradient descent method to achieve closed-loop adaptive optimization.

[0057] In step S1, the cutoff frequency of the low-pass filter is 5 Hz.

[0058] A sensor array integrated into a smart toothbrush or oral wearable device collects oral microenvironment signals in real time at a preset sampling frequency, such as 1Hz. The sensor array includes:

[0059] An ISFET solid-state pH sensor is used to acquire the raw pH voltage signal of oral saliva;

[0060] Temperature sensor used to collect oral temperature;

[0061] An accelerometer is used to collect data on the user's oral cavity activity.

[0062] The original pH voltage signal was low-pass filtered to remove high-frequency noise. A first-order Butterworth low-pass filter with a cutoff frequency of 5Hz was used to obtain the preprocessed pH voltage signal.

[0063] Step S2 specifically includes the following steps:

[0064] Step S2.1: Convert the preprocessed pH voltage signal into the original pH measurement value through linear calibration. The calibration formula is: pH measurement value equals the product of sensitivity calibration coefficient and voltage value plus bias calibration coefficient.

[0065] Step S2.2: The original pH measurement value is dynamically estimated using an adaptive Kalman filter algorithm to obtain the first estimated value; then, the first estimated value is smoothed using a sliding window midpoint filter with a length of 5 to obtain the steady-state pH estimate.

[0066] The preprocessed pH voltage signal is converted into the original pH measurement value through linear calibration. The calibration formula is: pH measurement value = sensitivity calibration coefficient × voltage value + bias calibration coefficient;

[0067] To address random noise in oral saliva and impulse interference caused by food debris impacts, a fusion algorithm combining adaptive Kalman filtering and dynamic sliding window median filtering is employed. The adaptive Kalman filter dynamically recursively estimates the original pH measurement, adjusting the covariance matrix of process noise and observation noise in real time based on the observation's information sequence. This effectively suppresses Gaussian noise and slowly changing drift, outputting a first estimate. A sliding window median filter of length 5 is then applied to this first estimate, sorting the estimates from the five most recent moments and using the median as the output. This median filter exhibits excellent ability to eliminate isolated large-amplitude impulse interference, such as the instantaneous impact of food debris on the sensor. Through these two stages of cascaded filtering, a highly stable steady-state pH estimate is finally output, providing a reliable data foundation for subsequent time-series predictions.

[0068] Step S3 specifically includes the following steps:

[0069] Step S3.1: Construct the input feature vector from the steady-state pH estimates at consecutive time points;

[0070] Step S3.2: Input the input feature vector into the Long Short-Term Memory Network Model. The Long Short-Term Memory Network Model contains a Long Short-Term Memory layer with 64 units, a dropout layer with a dropout rate of 0.2, and a fully connected output layer, which outputs the predicted pH value at future time.

[0071] Step S3.3: Train the Long Short-Term Memory network model using a composite loss function, which is a weighted sum of mean squared error loss, trend consistency loss, and acid peak penalty loss.

[0072] In step S3.3, the mean square error loss is equal to the reciprocal of the number of prediction steps multiplied by the sum of squares of the differences between the predicted pH value and the steady-state pH value at each prediction step; the trend consistency loss is equal to the reciprocal of the number of prediction steps minus one multiplied by the sum of squares of the differences in the signs of the trend changes between each adjacent prediction step, wherein the sign of the trend change is calculated by a sign function; the acidity peak penalty loss is equal to the reciprocal of the number of prediction steps multiplied by the sum of squares of the differences in the predicted pH values ​​at each prediction step that are lower than the acidity threshold of 5.5.

[0073] The steady-state pH estimates at 24 consecutive time points are used as the input feature vector;

[0074] The vector is input into a Long Short-Term Memory (LSTM) network model, which consists of: an LSTM layer with 64 memory units; a Dropout layer with a dropout rate of 0.2; and a fully connected output layer with an output dimension of 5, which predicts the pH value at the next 5 time points.

[0075] Define the composite loss function, with the following formula:

[0076]

[0077] in, Represents the composite loss function. Indicates the mean square error loss. This indicates a loss of trend consistency. This indicates the acid peak penalty loss. The weighting coefficient representing the trend consistency loss, with a value of 0.1, is used to balance the impact of trend loss on the total loss. The weighting coefficient representing the acid peak penalty loss is 0.05, used to balance the impact of the acid penalty on the total loss;

[0078] The specific formulas for the mean squared error loss, trend consistency loss, and acid peak penalty loss are as follows:

[0079]

[0080]

[0081]

[0082] in, This represents the number of prediction steps, with a value of 5, indicating that the model outputs predictions for the next 5 time points at a time. Indicates the index of the prediction step. The model represents time intervals. The predicted pH value, Indicates the time of output The steady-state pH estimate is used as the true label for training. This represents the change from step j to step (j+1) in the predicted sequence. This represents the change from step j to step (j+1) in the steady-state sequence. The sign function is represented by a value of 1 when the independent variable is greater than 0, -1 when it is less than 0, and 0 when it is equal to 0. 5.5 is the acid threshold. When the pH value is lower than 5.5, the teeth begin to demineralize and enter a high-risk state for caries.

[0083] During training, the Adam optimizer is used to update all trainable parameters of the Long Short-Term Memory network with the goal of minimizing the composite loss function.

[0084] Step S4 specifically includes the following steps:

[0085] Step S4.1: Construct a dynamic early warning threshold function. The dynamic early warning threshold function is equal to the individual baseline pH value minus the fluctuation coefficient multiplied by the standard deviation of the pH value in the past 24 hours, minus the cumulative acid attack coefficient multiplied by - minus an exponential function with the natural constant as the base, wherein the exponent of the exponential function is a negative decay rate constant multiplied by the cumulative duration of the pH value being below 5.5 in the past 6 hours.

[0086] Step S4.2: Calculate the comprehensive oral health risk index, which is equal to the positive part of the ratio of the difference between the dynamic warning threshold and the predicted pH value at the next moment to the dynamic warning threshold, multiplied by the first weight, plus the second weight multiplied by a subtracted S-shaped function, where the independent variable of the S-shaped function is the steepness coefficient multiplied by the difference between the total duration of pH value below 5.5 in the past 24 hours and the acid load tolerance baseline.

[0087] Step S4.3: Based on the numerical range of the comprehensive oral health risk index, a graded warning is issued. No warning is issued when the comprehensive oral health risk index is between 0 and 0.3. A first-level warning is issued when the comprehensive oral health risk index is between 0.3 and 0.7. A second-level warning is issued when the comprehensive oral health risk index is between 0.7 and 1.0.

[0088] Define a dynamic early warning threshold function, the specific formula of which is:

[0089]

[0090] in, Indicates time A dynamic early warning threshold is set, and an early warning is triggered when the predicted pH value is lower than this threshold. This represents the user's baseline pH level, calculated from the median of at least 7 days of monitoring data, reflecting the user's normal oral pH level. This represents the fluctuation coefficient, ranging from 1.5 to 2.5, which controls the penalty strength of the pH standard deviation on the threshold. This represents the standard deviation of steady-state pH over the past 24 hours. This represents the acid attack accumulation coefficient, ranging from 0.3 to 0.7, which controls the penalty intensity of the acid accumulation time on the threshold. This represents the decay rate constant, ranging from 0.01 to 0.05, which controls the rate of exponential decay. This indicates the cumulative duration of pH values ​​below 5.5 over the past 6 hours. The longer the acidity accumulation time, the closer the exponential term approaches 0, and the greater the threshold decreases.

[0091] The comprehensive oral health risk index is defined by the following formula:

[0092]

[0093] in, This represents the comprehensive oral health risk index; the higher the value, the higher the risk. and These represent weighting coefficients, taken as 0.6 and 0.4 respectively, used to balance the risk of prediction bias and the risk of long-term acid load. This represents the predicted pH value at the next time step output by the Long Short-Term Memory (LSTM) network model. This indicates the percentage of time in the past 24 hours during which the pH value was below 5.5. This represents the baseline tolerance to acidic load, set at 10%, and indicates the percentage of time that can be tolerated at low pH levels under normal conditions. The steepness coefficient of the S-shaped function ranges from 0.8 to 1.2, and the control risk index varies with... Sensitivity to change.

[0094] Step S5 specifically includes the following steps:

[0095] Step S5.1: Accumulate monitoring data for 7 days, and calculate the average squared difference between the predicted pH value and the steady-state pH value at all predicted times as the overall loss function;

[0096] Step S5.2: Use gradient descent to adjust all trainable parameters of the Long Short-Term Memory network to minimize the overall loss function.

[0097] Every 7 days of accumulated monitoring data, the mean squared error of prediction for all prediction times within the past 7 days is calculated as the overall loss function. The gradient of the overall loss function with respect to each trainable parameter of the Long Short-Term Memory (LSTM) network is calculated using the backpropagation algorithm. Starting from the output layer, the partial derivatives of the loss function with respect to the weight matrix and bias terms of each layer are calculated layer by layer using the chain rule. After obtaining all gradients, the Adam optimizer is used to update the network parameters, with an initial learning rate set to 0.001. The updated network parameters reduce the mean squared error of prediction, thereby improving the model's ability to predict future pH values. The fluctuation coefficient, acid attack accumulation coefficient, and decay rate constant in the dynamic warning threshold function are pre-set based on population statistics before user use and remain unchanged during use, requiring no online adjustment. Through periodic parameter updates, the system can gradually adapt to the individual user's oral pH variation patterns, continuously improving prediction accuracy and reducing false alarm and false negative rates.

[0098] A dynamic monitoring and early warning system for the oral microenvironment, used to implement any one of the methods for dynamic monitoring and early warning of the oral microenvironment, comprising:

[0099] The data acquisition module is used to acquire oral microenvironment signals in real time through a sensor array integrated into the oral wearable device and to perform low-pass filtering preprocessing on the raw pH voltage signal;

[0100] The signal processing module is used to extract steady-state pH estimates based on a fusion algorithm of adaptive Kalman filtering and dynamic sliding window midpoint filtering.

[0101] The time-series prediction module is used to input the steady-state pH estimation sequence into the long short-term memory network model, and train it using a composite loss function to output the predicted pH values ​​at multiple future time points.

[0102] The early warning decision module is used to construct a dynamic early warning threshold function, calculate a comprehensive oral health risk index, and generate graded early warning information.

[0103] The push and feedback module is used to push early warning content based on the graded early warning information and fine-tune the network parameters based on the gradient descent method.

[0104] The signal processing module includes:

[0105] The Kalman filter unit is used to perform adaptive Kalman filter estimation;

[0106] The median filtering unit is used to perform median filtering in a sliding window of length 5.

[0107] The time series prediction module includes:

[0108] Long Short-Term Memory (LSTM) units are used to extract long-term dependence features of pH time series.

[0109] The composite loss calculation unit is used to calculate the weighted sum of mean squared error loss, trend consistency loss and acid peak penalty loss as a composite loss function during the initial training phase, and to guide the initial training of the model with this composite loss function.

[0110] The parameter fine-tuning unit is used to fine-tune and update the parameters of the long short-term memory network by using the prediction mean square error as the loss function and the gradient descent method after accumulating monitoring data for every 7 days.

[0111] The system pushes warning information through multiple channels based on tiered warning information, including APP pop-ups, WeChat service notifications, and vibration alerts. Simultaneously, every 7 days of accumulated monitoring data, the system fine-tunes the parameters of the Long Short-Term Memory network using gradient descent based on the prediction mean square error, achieving closed-loop adaptive optimization.

[0112] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0113] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

Claims

1. A method for dynamic monitoring and early warning of the oral microenvironment, characterized in that, Includes the following steps: Step S1: Real-time acquisition of oral microenvironment signals through a sensor array integrated into the oral wearable device. The oral microenvironment signals include at least the original pH voltage signal, oral temperature signal, and oral activity state signal. The original pH voltage signal is preprocessed by low-pass filtering. Step S2: The pH measurement value obtained by converting the preprocessed pH voltage signal is estimated by using a fusion algorithm of adaptive Kalman filtering and dynamic sliding window midpoint filtering, and the steady-state pH estimate is output. Step S3: Construct a time series of steady-state pH estimates at consecutive time points, input it into a long short-term memory network model, and train the model using a composite loss function consisting of mean squared error loss, trend consistency loss, and acid peak penalty loss, and output the predicted pH value for future time points. Step S4: Construct a dynamic early warning threshold function based on individual baseline pH value, pH standard deviation and acid accumulation duration, compare the predicted pH value with the dynamic early warning threshold, calculate the comprehensive oral health risk index, and generate graded early warning information according to the interval of the comprehensive oral health risk index. Step S5: Accumulate monitoring data for a preset number of days, calculate the prediction mean square error, and adjust the parameters of the long short-term memory network using the gradient descent method to achieve closed-loop adaptive optimization.

2. The method according to claim 1, characterized in that, In step S1, the cutoff frequency of the low-pass filter is 5 Hz.

3. The method according to claim 2, characterized in that, Step S2 specifically includes the following steps: Step S2.1: Convert the preprocessed pH voltage signal into the original pH measurement value through linear calibration. The calibration formula is: pH measurement value equals the product of sensitivity calibration coefficient and voltage value plus bias calibration coefficient. Step S2.2: The original pH measurement value is dynamically estimated using an adaptive Kalman filter algorithm to obtain the first estimated value; then, the first estimated value is smoothed using a sliding window midpoint filter with a length of 5 to obtain the steady-state pH estimate.

4. The method according to claim 3, characterized in that, Step S3 specifically includes the following steps: Step S3.1: Construct the input feature vector from the steady-state pH estimates at consecutive time points; Step S3.2: Input the input feature vector into the Long Short-Term Memory Network Model. The Long Short-Term Memory Network Model contains a Long Short-Term Memory layer with 64 units, a dropout layer with a dropout rate of 0.2, and a fully connected output layer, which outputs the predicted pH value at future time. Step S3.3: Train the Long Short-Term Memory network model using a composite loss function, which is a weighted sum of mean squared error loss, trend consistency loss, and acid peak penalty loss.

5. The method according to claim 4, characterized in that, In step S3.3, the mean square error loss is equal to the reciprocal of the number of prediction steps multiplied by the sum of squares of the differences between the predicted pH value and the steady-state pH value at each prediction step; the trend consistency loss is equal to the reciprocal of the number of prediction steps minus one multiplied by the sum of squares of the differences in the signs of the trend changes between each adjacent prediction step, wherein the sign of the trend change is calculated by a sign function; the acidity peak penalty loss is equal to the reciprocal of the number of prediction steps multiplied by the sum of squares of the differences in the predicted pH values ​​at each prediction step that are lower than the acidity threshold of 5.

5.

6. The method according to claim 5, characterized in that, Step S4 specifically includes the following steps: Step S4.1: Construct a dynamic early warning threshold function. The dynamic early warning threshold function is equal to the individual baseline pH value minus the fluctuation coefficient multiplied by the standard deviation of the pH value in the past 24 hours, minus the cumulative acid attack coefficient multiplied by - minus an exponential function with the natural constant as the base, wherein the exponent of the exponential function is a negative decay rate constant multiplied by the cumulative duration of the pH value being below 5.5 in the past 6 hours. Step S4.2: Calculate the comprehensive oral health risk index, which is equal to the positive part of the ratio of the difference between the dynamic warning threshold and the predicted pH value at the next moment to the dynamic warning threshold, multiplied by the first weight, plus the second weight multiplied by a subtracted S-shaped function, where the independent variable of the S-shaped function is the steepness coefficient multiplied by the difference between the total duration of pH value below 5.5 in the past 24 hours and the acid load tolerance baseline. Step S4.3: Based on the numerical range of the comprehensive oral health risk index, a graded warning is issued. No warning is issued when the comprehensive oral health risk index is between 0 and 0.

3. A first-level warning is issued when the comprehensive oral health risk index is between 0.3 and 0.

7. A second-level warning is issued when the comprehensive oral health risk index is between 0.7 and 1.

0.

7. The method according to claim 6, characterized in that, Step S5 specifically includes the following steps: Step S5.1: Accumulate monitoring data for 7 days, and calculate the average squared difference between the predicted pH value and the steady-state pH value at all predicted times as the overall loss function; Step S5.2: Use gradient descent to adjust all trainable parameters of the Long Short-Term Memory network to minimize the overall loss function.

8. A dynamic monitoring and early warning system for the oral microenvironment, used to implement the dynamic monitoring and early warning method for the oral microenvironment as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire oral microenvironment signals in real time through a sensor array integrated into the oral wearable device and to perform low-pass filtering preprocessing on the raw pH voltage signal; The signal processing module is used to extract steady-state pH estimates based on a fusion algorithm of adaptive Kalman filtering and dynamic sliding window midpoint filtering. The time-series prediction module is used to input the steady-state pH estimation sequence into the long short-term memory network model, and train it using a composite loss function to output the predicted pH values ​​at multiple future time points. The early warning decision module is used to construct a dynamic early warning threshold function, calculate a comprehensive oral health risk index, and generate graded early warning information. The push and feedback module is used to push early warning content based on the graded early warning information and fine-tune the network parameters based on the gradient descent method.

9. The system according to claim 8, characterized in that, The signal processing module includes: Kalman filter unit, used to perform adaptive Kalman filter estimation; The median filtering unit is used to perform median filtering in a sliding window of length 5.

10. The system according to claim 9, characterized in that, The time series prediction module includes: Long Short-Term Memory (LSTM) units are used to extract long-term dependence features of pH time series. The composite loss calculation unit is used to calculate the weighted sum of mean squared error loss, trend consistency loss and acid peak penalty loss as a composite loss function during the initial training phase, and to guide the initial training of the model with this composite loss function. The parameter fine-tuning unit is used to fine-tune and update the parameters of the long short-term memory network by using the prediction mean square error as the loss function and the gradient descent method after accumulating monitoring data for every 7 days.

Images