Method and system for reducing data error of dynamic glucometer

Abstract

The invention discloses a method and system for reducing data errors of a dynamic glucometer, and the method comprises the steps: collecting a current signal of fluid between subcutaneous tissues in real time through an implantable sensor, and synchronously recording the working environment parameters of the sensor; carrying out denoising and normalization processing on the current signal, and carrying out temperature compensation and drift correction on the signal according to a working environment parameter; extracting time sequence features and frequency domain features from the signals processed in the step S2; training a machine learning model based on the time sequence feature data and the frequency domain feature data, wherein the machine learning model is used for predicting the blood glucose concentration and identifying abnormal data points; correcting a model prediction result through residual analysis and weighted adjustment, and performing secondary correction on abnormal data points in combination with event data input by a patient; and outputting the optimized blood glucose data to a terminal display device in a time sequence curve form. Through the process, the influence of abnormal state data on blood glucose monitoring is reduced, and the data stability is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for reducing data errors in continuous glucose monitoring systems. Background Technology

[0002] Current continuous glucose meters use filtering algorithms, such as simple mean filtering, to reduce noise in the current signal collected by the sensor. They convert the signal into blood glucose concentration through a fixed-parameter regression model, perform only baseline temperature compensation, rely on a single threshold to identify abnormal data, and do not incorporate patient dietary, exercise, or other event data for calibration.

[0003] When the body temperature is higher than the standard value, the lack of dynamic compensation leads to a larger signal and a higher predicted blood glucose value, typically with a deviation of 1.5-2.0 mmol / L. After the sensor has been implanted for more than 168 hours, its sensitivity decays, and without calibration, the signal becomes smaller, resulting in a lower predicted blood glucose value, with a daily error of 0.3-0.5 mmol / L. Furthermore, signal noise caused by diet and exercise is misinterpreted, leading to fluctuating blood glucose levels, with an error of 1.3-2.1 mmol / L during exercise. Fixed models do not adapt to individual differences, resulting in errors 30%-50% higher during periods of rapid change. These issues are not effectively addressed in current commercially available continuous glucose monitoring devices; they simply collect blood samples at regular intervals to measure glucose concentration.

[0004] Therefore, there is an urgent need for a method to reduce the data error of continuous glucose monitoring systems. This method can eliminate systematic errors caused by temperature and humidity and sensitivity decay through dynamic compensation; extract multi-dimensional features and use machine learning to improve adaptability to nonlinear signals; and combine event data to correct anomalies and reduce error fluctuations. Summary of the Invention

[0005] This application provides a method and system for reducing data errors in continuous glucose monitoring systems to solve the above-mentioned problems.

[0006] On the one hand, this application provides a method for reducing the data error of a continuous glucose monitoring system, the method comprising the following steps:

[0007] Step S1: Real-time acquisition of current signals in subcutaneous interstitial fluid using an implantable sensor, and simultaneous recording of sensor operating environment parameters;

[0008] Step S2: Denoise and normalize the current signal, and perform temperature compensation and drift correction on the signal according to the working environment parameters.

[0009] Step S3: Extract time series features and frequency domain features from the signal processed in step S2;

[0010] Step S4: Train a machine learning model based on time series feature data and frequency domain feature data. The machine learning model is used to predict blood glucose concentration and identify abnormal data points.

[0011] Step S5: Correct the model prediction results through residual analysis and weighted adjustment, and at the same time, perform secondary correction on abnormal data points by combining the event data input by the patient;

[0012] Step S6: Output the optimized blood glucose data to the terminal display device in the form of a time series curve.

[0013] In one implementation of this application, step S2 specifically includes:

[0014] The Kalman filter algorithm is used to dynamically filter the current signal. Its state equation and observation equation are as follows:

[0015] State equation: x k =Ax k-1 +Bu k +w k ;

[0016] Observation equation: z k =Hx k +v k ;

[0017] Where, x k Let u be the state vector. k For input control, w k v k Let A, B, and H be noise, and A, B, and H be system matrices.

[0018] The current signal is linearly compensated based on a standard temperature reference value. The compensation function is: I 补偿 =I 原始 ·(1+α·ΔT), where α is the temperature compensation coefficient, ΔT is the difference between the current temperature and the standard temperature, and I 原始 I refers to the raw current signal acquired by the sensor without temperature compensation. 补偿 This refers to the current signal after temperature correction.

[0019] In one implementation of this application, step S3 specifically includes: calculating the mean, variance, kurtosis, and skewness of the signal; converting the signal to the frequency domain using a fast Fourier transform, and extracting the dominant frequency, secondary frequency, and power spectral density.

[0020] In one implementation of this application, the machine learning model includes a support vector machine regression model or a long short-term memory neural network model, and its loss function is:

[0021] Where N is the number of samples, yi This represents the actual blood glucose level. These are the model's predicted values.

[0022] In one implementation of this application, step S5 involves correcting the model prediction results through residual analysis and weighted adjustment, specifically as follows:

[0023] Calculate the residual between the model's predicted value and the actual value. If the absolute value of the residual exceeds a preset threshold, it is marked as an outlier data point.

[0024] By assigning lower weights to outlier data points and retraining the model, the impact of these outliers on the overall prediction results can be reduced.

[0025] In one implementation of this application, the events input by the patient include diet, exercise, and insulin injection information. In step S5, the secondary correction step includes:

[0026] Match event data with timestamps of outlier data points;

[0027] Adjust model parameters based on event type to dynamically correct outlier data points.

[0028] In one implementation of this application, the sensor operating environment parameters include temperature, humidity, and sensor implantation time. In step S2, the drift correction step includes:

[0029] The current signal is exponentially compensated based on the sensor sensitivity decay curve. The compensation formula is as follows:

[0030] I 补偿 =I 原始 ·e -βt

[0031] Where β is the attenuation coefficient and t is the sensor implantation time;

[0032] The background current component is separated from the original signal, and the effective signal is processed separately.

[0033] In one implementation of this application, the method for identifying abnormal data points in step S5 includes:

[0034] If the rate of change in blood glucose levels between adjacent data points exceeds a preset threshold, it is considered abnormal.

[0035] The smoothed value of the signal is calculated by a sliding window. If the difference between the current value and the smoothed value exceeds the threshold, it is judged as abnormal.

[0036] In one implementation of this application, the method further includes:

[0037] Generate a blood glucose fluctuation trend chart and mark the high-glycemia and low-glycemia risk areas;

[0038] When blood glucose levels exceed the preset safe range, an audible and visual alarm will be triggered via the terminal device.

[0039] On the other hand, this application also provides a system for reducing data errors in continuous glucose monitoring systems, the system comprising:

[0040] The sensor module is used to collect current signals and environmental parameters of the subcutaneous interstitial fluid;

[0041] Data processing module: includes a preprocessing unit, a feature extraction unit, a model building unit, and an error optimization unit, used to execute the aforementioned methods;

[0042] Storage module: Used to store raw data, model parameters, and patient event data;

[0043] Display module: Used to output optimized blood glucose data, trend analysis results, and alarm prompts;

[0044] Communication module: Used for data interaction with terminal devices.

[0045] The method and system for reducing data error in a continuous glucose monitoring system provided in this application have the following beneficial effects:

[0046] (1) By using preprocessing techniques such as Kalman filtering, temperature compensation and attenuation correction, combined with residual analysis and weighted adjustment of machine learning models, interference such as environmental noise and sensor drift can be effectively suppressed, the impact of abnormal data on blood glucose monitoring can be reduced, and data stability can be improved.

[0047] (2) By leveraging the dynamic matching mechanism of patient event data and anomalies, and combining the dual anomaly identification algorithm of physiological model and data-driven approach, secondary calibration of blood glucose data in scenarios such as diet and exercise can be achieved, significantly improving the robustness of the model and the accuracy of clinical monitoring. Attached Figure Description

[0048] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0049] Figure 1 A flowchart illustrating a method for reducing data error in a continuous glucose monitoring system, as provided in this application embodiment;

[0050] Figure 2 This is a system composition diagram for reducing data error in a continuous glucose monitoring system, provided as an embodiment of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] Currently, although filtering algorithms and simple regression models can be used for data processing, they lack dynamic identification and weighted optimization of abnormal data points, and do not incorporate real-time event data input by patients for secondary calibration. As a result, the models have poor robustness and cannot meet the needs of precise clinical monitoring.

[0053] Therefore, there is an urgent need for a method that can comprehensively process environmental parameters, extract multi-dimensional signal features, and dynamically optimize errors through machine learning models to reduce the impact of abnormal data on blood glucose monitoring results and improve the measurement accuracy and reliability of continuous glucose monitoring (CGM) meters. To this end, this application provides a method and system for reducing data errors in CGM meters. The technical solution proposed in this application will be described in detail below with reference to the accompanying drawings.

[0054] Figure 1 This is a flowchart illustrating a method for reducing data error in a continuous glucose monitoring system, as provided in an embodiment of this application. Figure 1 As shown, the method mainly includes the following steps:

[0055] Step S1: The current signal of the subcutaneous interstitial fluid is collected in real time by an implanted sensor, and the working environment parameters of the sensor are recorded simultaneously.

[0056] In this embodiment, the sensor's operating environment parameters include temperature, humidity, and sensor implantation time. Specifically, the sensor is an implantable electrochemical sensor with an integrated glucose oxidase electrode, which acquires the current signal generated by the glucose oxidation reaction in the subcutaneous interstitial fluid in real time at a frequency of 5 minutes per acquisition. This signal typically ranges from 10 to 20 nA, with a dynamic response time ≤ 2 seconds.

[0057] Furthermore, the synchronously acquired sensor operating environment parameters are achieved through a miniature detection module built into the sensor: temperature parameters are measured using high-precision thermocouple elements with an accuracy of ±0.1℃, capable of capturing temperature fluctuations of 0.5℃ caused by human activity or environmental changes; humidity parameters are for monitoring the subcutaneous microenvironment, covering a range of 30%-90%RH with a measurement error ≤2%RH; sensor implantation time is precisely timed from the moment of implantation completion, with the accumulated time recorded in hours, achieving a recording accuracy of 0.1 hours. All acquired data is accompanied by a unified timestamp, with a time synchronization error ≤1 second, providing a consistent foundation of data for subsequent temperature compensation, drift correction, and other processing.

[0058] Step S2: Denoise and normalize the current signal, and perform temperature compensation and drift correction on the signal according to the working environment parameters.

[0059] In this embodiment, the Kalman filter algorithm is used to dynamically filter the current signal. Its state equation and observation equation are as follows:

[0060] State equation: x k =Ax k-1 +Bu k +C ek +w k ;

[0061] Observation equation: z k =Hx k +v k ;

[0062] Where, x k Let u be the state vector. k For input control, w k v k For noise, e k =[ΔTk,ΔHk] T Let C represent the temperature and humidity changes, and C be the environmental influence coefficient matrix, obtained by fitting experimental data, such as temperature coefficient c1 = 0.02 and humidity coefficient c2 = 0.01. A, B, and H are the system matrices. This setting allows the filtering model to perceive environmental changes in real time and dynamically adjust the noise covariance matrices Q and R (increasing the noise weight when environmental fluctuations are large), improving the denoising stability in complex environments. During periods of drastic environmental change (such as a sudden rise in body temperature after exercise), the signal-noise variance is reduced by 40%, and the deviation between the filtered signal and the true value is reduced to ±0.3 nA.

[0063] The current signal is linearly compensated based on a standard temperature reference value. The compensation function is: I 补偿 =I 原始 ·(1+α·ΔT), where α is the temperature compensation coefficient, ΔT is the difference between the current temperature and the standard temperature, and I原始 I refers to the raw current signal acquired by the sensor without temperature compensation. 补偿 This refers to the current signal after temperature correction.

[0064] Furthermore, the original linear compensation formula I 补偿 =I 原始 ·(1+α·ΔT), we can also introduce a modified nonlinear characteristic: I 补偿 =I 原始 ·(1+α·ΔT+β·(ΔT) 2 ), where β=0.0005, is calibrated using the sensor's temperature response curve, and the quadratic term is automatically activated when ΔT>2℃. This setting allows for a more accurate fit to the nonlinear decay of sensor sensitivity in scenarios with large body temperature fluctuations (such as fever or strenuous exercise). When the temperature difference exceeds 3℃, the compensation error decreases from the original 1.2% to below 0.5%, and the blood glucose prediction deviation is reduced by ±0.3mmol / L.

[0065] In this embodiment of the application, the drift correction step includes:

[0066] The current signal is exponentially compensated based on the sensor sensitivity decay curve. The compensation formula is as follows:

[0067] I 补偿 =I 原始 ·e -βt

[0068] Where β is the attenuation coefficient, the value of which is determined by experiments on the attenuation characteristics of sensor sensitivity with implantation time. Due to biocompatibility reactions such as subcutaneous tissue encapsulation and electrode oxidation, the sensitivity of this type of implantable sensor decreases exponentially with time; t is the sensor implantation time;

[0069] Background current components are separated from the raw signal, and the effective signal is processed separately. To separate the background current components, a reference signal is first obtained through sensor baseline calibration. This signal contains interfering components such as electrode polarization current (approximately 2-5 nA) and non-specific oxidation reaction current in subcutaneous tissue fluid. An adaptive noise canceller is used, with the baseline signal as the reference input and the raw acquired signal as the main input. The LMS algorithm dynamically adjusts the filter coefficients (convergence factor set to 0.01) to achieve real-time separation of background current and effective signal. The separated effective signal (related only to glucose oxidation reaction, typically with an amplitude of 8-18 nA) undergoes separate 50Hz notch filtering and a 10x gain adjustment to further improve the signal-to-noise ratio, providing a clean signal source for subsequent feature extraction.

[0070] Furthermore, the original compensation function formula is: I 补偿 =I 原始 ·e -βtUsing a fixed β does not account for the varying sensitivity decay rates caused by individual differences in the subcutaneous environment (such as tissue fluid pH). Therefore, a secondary correction is necessary. An individual calibration factor k (range 0.8-1.2, determined through initial blood collection calibration) is introduced, and the formula is optimized as follows:

[0071] I 补偿 =I 原始 ·e -kβt

[0072] Among them, k is adaptively adjusted based on the blood glucose-current correlation of the user in the first 3 days (e.g., k = 1.1 when the tissue fluid viscosity is high).

[0073] This optimization allows for adaptation to the differences in subcutaneous microenvironments among different users, making the attenuation correction more closely match the actual performance degradation pattern of individual sensors. Seven days after sensor implantation, the inter-individual correction error decreased from ±0.4 mmol / L to ±0.15 mmol / L, and long-term monitoring stability improved by 50%.

[0074] It should be noted that, to address the issue of low convergence efficiency caused by weight oscillations in small sample scenarios, a dynamic learning rate scheduling method combining cosine annealing and warm-up can be further introduced, as shown in the following formula:

[0075]

[0076] Where, η t Let η be the learning rate at step t; max η represents the maximum learning rate (typically 0.001–0.01, such as 0.005 for a BiLSTM model). min The minimum learning rate is set to 1 / 100 of the given value, i.e., 0.00005; T cur T is the number of steps within the current training iteration; max The total number of steps in a single round (determined by the dataset size and batch size, e.g., when batch=32 and N=1000). ); The function is an indicator function (1 if the condition is met, 0 otherwise); t warmup Preheating steps (the first 500 wires are linearly heated to avoid initial gradient explosion); η warmup The initial learning rate for warm-up is set to 1 / 20 of the initial value, which is 0.00025.

[0077] This improvement linearly increases the learning rate during the warm-up phase, quickly converging to a better parameter region. During the cosine annealing phase, the learning rate is periodically adjusted to escape local optima and enhance generalization ability. The training loss oscillation amplitude is reduced by 60%, the convergence speed is increased by 40% (reducing the number of iterations to achieve the same loss), and the test set MSE is reduced by 15%.

[0078] Step S3: Extract time series features and frequency domain features from the signal processed in step S2.

[0079] In this embodiment, the mean, variance, kurtosis, and skewness of the signal are calculated. The signal is converted to the frequency domain using a Fast Fourier Transform (FFT) to extract the dominant frequency, secondary frequency, and power spectral density. For the preprocessed current signal, a 10-minute sliding window (containing 12 sampling points) is used to calculate the time series characteristics: the mean reflects the overall signal level, the variance reflects the fluctuation amplitude, the kurtosis quantifies the steepness of the signal distribution (reference threshold 3.0), and the skewness characterizes the symmetry of the distribution (range -1 to 1). The signal is converted to the frequency domain using a Fast Fourier Transform with a sampling frequency of 1 Hz, and the dominant and secondary frequencies (usually concentrated in 0.03-0.07 Hz) with a power share exceeding 30% are extracted. The power spectral density in the 0-0.1 Hz band is calculated to capture the frequency characteristics of blood glucose fluctuations.

[0080] Step S4: Train a machine learning model based on time series feature data and frequency domain feature data. The machine learning model is used to predict blood glucose concentration and identify abnormal data points.

[0081] In this embodiment of the application, the machine learning model includes a support vector machine regression model or a long short-term memory neural network model, and its loss function is:

[0082] Where N is the number of samples, y i This represents the actual blood glucose level. These are the model's predicted values.

[0083] This machine learning model takes extracted time-series features (mean, variance, etc.) and frequency domain features (dominant frequency, power spectral density, etc.) as input, and outputs a predicted blood glucose concentration value through a hidden layer of 64 neurons. The parameters are iteratively optimized using a mean squared error loss function. When identifying anomalies, if the rate of change in blood glucose between adjacent data points exceeds a preset threshold of 2.0 mmol / L / h, or if the difference between the smoothed value calculated using a 5-point sliding window and the current value exceeds a threshold of 1.5 mmol / L, then the data point is considered an anomaly.

[0084] Furthermore, the original mean squared error loss function assigns equal weight to outlier data (such as jump points caused by exercise or diet), leading to the model overfitting noise.

[0085] To address this, an outlier data weighting factor w is introduced. i Improve the loss function:

[0086]

[0087] Among them, w i =exp(-γ·|r i |), The residual is γ = 0.8, which is the attenuation coefficient. The more significant the anomaly, the higher the attenuation coefficient. i The smaller the value, the lower the value is, down to 0.1.

[0088] This optimization dynamically reduces the weight of outlier data in model training, minimizing the interference of extreme values ​​on parameter updates and enhancing the model's ability to fit normal physiological fluctuations. When outlier data accounts for more than 10%, the mean squared error (MSE) of the model's prediction decreases by 35%, and the prediction accuracy for postprandial blood glucose peaks improves by 20%.

[0089] Furthermore, to constrain model parameter complexity and address overfitting in small-sample scenarios, such as model weight oscillations caused by short-term user anomalies, an L2 regularization term can be introduced to balance fitting accuracy and model complexity. The improved loss function is:

[0090]

[0091] Where λ is the regularization strength coefficient (range [0, 0.1], the optimal value is determined by 5-fold cross-validation, typical value λ = 0.03); θ j M represents the trainable parameters of the model (such as elements of the neural network weight matrix, coefficients of the Kalman filter state transition matrix, etc.); M represents the total number of model parameters (automatically counted by the model structure, such as the number of neurons in a BiLSTM layer × the input dimension).

[0092] By penalizing large parameter values, the model is limited from overfitting to abnormal data while retaining its ability to fit normal physiological signals, thus improving generalization across users and scenarios. With a small training set (sample size <50), the model's MSE on the test set is reduced by 25%, and the blood glucose prediction bias for new users is reduced to ±0.25 mmol / L.

[0093] Step S5: Correct the model prediction results through residual analysis and weighted adjustment, and at the same time, perform secondary correction on abnormal data points by combining the event data input by the patient.

[0094] In this embodiment, the model prediction results are corrected through residual analysis and weighted adjustment. Specifically, the residual between the model prediction value and the actual value is calculated. If the absolute value of the residual exceeds a preset threshold, it is marked as an abnormal data point. The abnormal data point is assigned a lower weight, and the model is retrained to reduce its impact on the overall prediction results.

[0095] In this embodiment of the application, the events input by the patient include diet, exercise, and insulin injection information.

[0096] In this embodiment, the secondary correction step includes: matching event data with the timestamps of abnormal data points; adjusting model parameters according to the event type to dynamically correct abnormal data points. First, the system's built-in time synchronization module achieves accurate matching of event data and abnormal data points, controlling the timestamp error within ±30 seconds to ensure a temporal correlation between events such as diet, exercise, and insulin injection and the corresponding abnormal data points. Then, the model parameters are dynamically adjusted for different event types: for diet events, the carbohydrate coefficient correction module is invoked, adjusting the blood glucose rise slope parameter from the default 0.5 mmol / (L·g) to 0.3-0.7 mmol / (L·g) (depending on the food's GI value); for exercise events, the metabolic rate correction factor is enabled, widening the blood glucose change rate threshold from 2.0 mmol / L / h to 3.5 mmol / L / h; for insulin injection events, the action delay coefficient is adjusted, reducing the model's prediction deviation for blood glucose decrease within 30 minutes after injection by 40%, ultimately achieving dynamic correction of abnormal data points.

[0097] In this embodiment, the method for identifying abnormal data points includes: if the rate of change of blood glucose in adjacent data points exceeds a preset threshold, it is determined to be abnormal; specifically, for adjacent data points at 5-minute intervals, the rate of change of blood glucose (unit: mmol / L / h) is calculated. If this value exceeds a preset threshold of 2.0 mmol / L / h (e.g., a sudden increase from 5.0 mmol / L to 8.0 mmol / L, with a rate of change of 36 mmol / L / h, is considered abnormal), a smoothed signal value is calculated using a sliding window. If the difference between the current value and the smoothed value exceeds a threshold, it is determined to be abnormal. Specifically, a 10-minute sliding window (containing 12 sampling points) is used, and the smoothed signal value is calculated using an arithmetic mean. If the difference between the current blood glucose value and the smoothed value exceeds a threshold of 1.5 mmol / L (e.g., the smoothed value is 6.2 mmol / L, and the current value is 8.0 mmol / L), it is marked as an abnormal data point.

[0098] Step S6: Output the optimized blood glucose data to the terminal display device in the form of a time series curve.

[0099] In this embodiment, a blood glucose fluctuation trend chart is generated and high and low blood glucose risk areas are marked; when the blood glucose value exceeds the preset safe range, an audible and visual alarm is issued through the terminal device.

[0100] The above is a method for reducing data error in a continuous glucose monitoring system provided by embodiments of this application. Based on the same inventive concept, embodiments of this application also provide a system for reducing data error in a continuous glucose monitoring system. Figure 2 A system composition diagram for reducing data error in a continuous glucose monitoring system provided in this application embodiment is shown below. Figure 2As shown, the system mainly includes: a sensor module 201, used to collect current signals and environmental parameters of subcutaneous interstitial fluid; and a data processing module 202, including a preprocessing unit 2021, a feature extraction unit 2022, a model building unit 2023, and an error optimization unit 2024, used to execute the method described in any one of claims 1-9.

[0101] Storage module 203: Used to store raw data, model parameters, and patient event data;

[0102] Display module 204: Used to output optimized blood glucose data, trend analysis results, and alarm prompts;

[0103] Communication module 205: Used for data interaction with terminal devices.

[0104] The method and system provided in this application utilize preprocessing techniques such as Kalman filtering, temperature compensation, and attenuation correction, combined with residual analysis and weighted adjustment of machine learning models, to effectively suppress interference from environmental noise and sensor drift, reduce the impact of abnormal data on blood glucose monitoring, and improve data stability. By leveraging a dynamic matching mechanism between patient event data and outliers, combined with a dual anomaly identification algorithm driven by physiological models and data, secondary calibration of blood glucose data under scenarios such as diet and exercise is achieved, significantly improving model robustness and clinical monitoring accuracy.

[0105] Below is an example of this application in a specific application scenario.

[0106] In a routine blood glucose monitoring scenario for a diabetic patient, the continuous glucose monitoring system collects the current signal of subcutaneous interstitial fluid every 5 minutes via an implanted sensor, simultaneously recording the sensor surface temperature (accuracy ±0.1℃), implantation time (cumulative in hours), and ambient humidity (percentage). On a certain morning at 8:00 AM, the patient consumed breakfast (containing 50g of carbohydrates), performed 30 minutes of moderate-intensity exercise at 9:00 AM, and injected rapid-acting insulin at 11:00 AM. These events were manually entered into the system via a terminal device, corresponding to timestamps of 8:15, 9:00, and 11:10, respectively.

[0107] The initial current signal acquired by the sensor is affected by metabolic changes after breakfast, exhibiting high-frequency noise between 8:20 and 8:40. The original signal value fluctuates between 12.3 nA and 15.8 nA, while the sensor surface temperature rises by 1.2°C due to human activity. In the preprocessing stage, a Kalman filter algorithm is used to process the signal during this period. The state vector x in the state equation... k Including signal trend and noise terms, the filtered signal range is narrowed to 13.5nA to 14.8nA through recursive calculation, and the noise variance is reduced by 60%. The temperature compensation stage is based on formula I. 补偿 =I 原始·(1+α·ΔT), taking α=0.002 / ℃, ΔT=1.2℃, the overall signal is reduced by 2.4% after compensation, eliminating the influence of temperature rise on sensor sensitivity. For sensors implanted for 168 hours, the drift correction module applies exponential compensation formula I. 补偿 =I 原始 ·e -βt According to the table, β = 0.01 / hour and t = 168 hours at this time. The signal is further compensated by 40%, which corrects the sensitivity decay caused by long-term sensor implantation.

[0108] The feature extraction process processed 49 signal points from 8:00 to 12:00, yielding time-series features with a mean of 14.2 nA, variance of 0.85, kurtosis of 1.2, and skewness of 0.3. Frequency domain features, after FFT transformation, showed a dominant frequency concentrated at 0.05 Hz (corresponding to the blood glucose fluctuation cycle), with 75% of the power spectral density falling within the 0.03-0.07 Hz range. The machine learning model employed an LSTM neural network, with an input layer consisting of feature vectors from 10 time steps, 64 hidden neurons, and an output layer providing predicted blood glucose concentrations. The training data included 600 labeled samples from the patient over the past 3 days, and the loss function used was the mean squared error formula. During initial training, the unoptimized model failed to adequately identify abnormal data during the 9:00 exercise period. During this period, the actual blood glucose level dropped sharply from 7.8 mmol / L to 5.2 mmol / L, while the model's predicted value only dropped to 6.5 mmol / L, resulting in a residual of 1.3 mmol / L, which exceeded the preset threshold of 1.0 mmol / L.

[0109] In the error optimization phase, the system first identified three data points—9:00, 9:05, and 9:10—as anomalies using residual analysis. These corresponded to predicted values ​​of 6.5 mmol / L, 6.3 mmol / L, and 6.1 mmol / L, while the actual values ​​were 5.2 mmol / L, 5.5 mmol / L, and 5.8 mmol / L, respectively. During weighted adjustment, the weights of these three points were reduced from the default 0.9 to 0.3. After retraining the model, the predicted values ​​near the anomalies showed a 40% reduction in error in subsequent tests. In the secondary correction phase, the system matched the patient's input exercise event. Recognizing the exercise as moderate intensity, it invoked a preset exercise impact model, temporarily adjusting the blood glucose change rate threshold for that time period from the default 2.0 mmol / L / h to 3.5 mmol / L / h to prevent rapid blood glucose drops due to exercise from being misinterpreted as sensor malfunction. Simultaneously, in conjunction with the insulin injection event, within 30 minutes after 11:10, the model parameters dynamically adjusted the insulin action delay coefficient, reducing the prediction error from 1.2 mmol / L before injection to 0.8 mmol / L.

[0110] Furthermore, the error comparison before and after data optimization was quantified using the root mean square error (RMSE). The RMSE of the blood glucose value directly converted from the original signal was 2.1 mmol / L, which was reduced to 1.6 mmol / L after preprocessing. The initial RMSE predicted by the model was 1.3 mmol / L, and after residual weighting and event correction, the final output RMSE was reduced to 0.7 mmol / L. Specific time period comparisons showed that the error decreased from 1.8 mmol / L to 0.9 mmol / L within 30 minutes after breakfast (normal physiological fluctuations) after optimization; the error for abnormal data during exercise decreased from 1.3 mmol / L to 0.5 mmol / L; and the error after insulin injection decreased from 1.5 mmol / L to 0.6 mmol / L. The blood glucose trend chart generated by the system at 12:00 accurately marked the hypoglycemic risk area from 9:00 to 9:30, and triggered an audible and visual alarm when the blood glucose value dropped to 3.9 mmol / L at 11:45, issuing a warning 10 minutes earlier than the traditional system.

[0111] Throughout the data processing, the error optimization module achieves accurate identification of abnormal states through multi-dimensional data fusion: first, it corrects sensor physical errors using environmental parameters; second, it captures physiological signal characteristics through a machine learning model; and third, it establishes contextual associations by combining patient-inputted event data, ultimately forming a three-layer error suppression mechanism. Compared to traditional methods, the accuracy of abnormal data point identification in this embodiment increases from 65% to 89%, and the model's response delay to sudden physiological changes is reduced from 20 minutes to 5 minutes. This significantly improves the monitoring accuracy of the continuous glucose monitor in complex scenarios, especially in situations where blood glucose fluctuates drastically due to diet, exercise, or medication interventions. It effectively reduces the risk of misjudgment and provides more reliable data support for clinical blood glucose management. By updating model weights in real time and dynamically adjusting parameters, the system achieves adaptive learning of individual patient physiological characteristics. Even when sensor performance degrades over time, it maintains stable long-term monitoring accuracy, solving the problem of insufficient adaptability of traditional fixed-parameter models to time-varying signals.

[0112] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0113] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0114] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for reducing data error in a continuous glucose monitoring system, characterized in that, The method includes the following steps: Step S1: Real-time acquisition of current signals in subcutaneous interstitial fluid using an implantable sensor, and simultaneous recording of sensor operating environment parameters; Step S2: Denoise and normalize the current signal, and perform temperature compensation and drift correction on the signal according to the working environment parameters. Step S3: Extract time series features and frequency domain features from the signal processed in step S2; Step S4: Train a machine learning model based on time series feature data and frequency domain feature data. The machine learning model is used to predict blood glucose concentration and identify abnormal data points. Step S5: Correct the model prediction results through residual analysis and weighted adjustment, and at the same time, perform secondary correction on abnormal data points by combining the event data input by the patient; Step S6: Output the optimized blood glucose data to the terminal display device in the form of a time series curve.

2. The method for reducing data error in a continuous glucose monitoring system according to claim 1, characterized in that, Step S2 specifically includes: The Kalman filter algorithm is used to dynamically filter the current signal. Its state equation and observation equation are as follows: State equation: x k =Ax k-1 +Bu k +w k ; Observation equation: z k =Hx k +v k ; Where, x k Let u be the state vector. k For input control, w k v k Let A, B, and H be noise, and A, B, and H be system matrices. The current signal is linearly compensated based on a standard temperature reference value. The compensation function is: I 补偿 =I 原始 ·(1+α·ΔT), where α is the temperature compensation coefficient, ΔT is the difference between the current temperature and the standard temperature, and I 原始 I refers to the raw current signal acquired by the sensor without temperature compensation. 补偿 This refers to the current signal after temperature correction.

3. The method for reducing data error in a continuous glucose monitoring system according to claim 1, characterized in that, Step S3 specifically includes: calculating the mean, variance, kurtosis and skewness of the signal; converting the signal to the frequency domain through fast Fourier transform, and extracting the dominant frequency, secondary frequency and power spectral density.

4. The method for reducing data error in a continuous glucose monitoring system according to claim 1, characterized in that, The machine learning model includes a support vector machine regression model or a long short-term memory neural network model, and its loss function is: Where N is the number of samples, y i This represents the actual blood glucose level. These are the model's predicted values.

5. The method for reducing data error in a continuous glucose monitoring system according to claim 1, characterized in that, In step S5, the model prediction results are corrected through residual analysis and weighted adjustment, specifically as follows: Calculate the residual between the model's predicted value and the actual value. If the absolute value of the residual exceeds a preset threshold, it is marked as an outlier data point. By assigning lower weights to outlier data points and retraining the model, the impact of these outliers on the overall prediction results can be reduced.

6. The method for reducing data error in a continuous glucose monitoring system according to claim 1, characterized in that, The events input by the patient include diet, exercise, and insulin injection information. In step S5, the secondary correction specifically includes: Match event data with timestamps of outlier data points; Adjust model parameters based on event type to dynamically correct outlier data points.

7. The method for reducing data error in a continuous glucose monitoring system according to claim 1, characterized in that, The sensor's operating environment parameters include temperature, humidity, and sensor implantation time. In step S2, the drift correction step includes: The current signal is exponentially compensated based on the sensor sensitivity decay curve. The compensation formula is as follows: I 补偿 =I 原始 ·e -βt Where β is the attenuation coefficient and t is the sensor implantation time; The background current component is separated from the original signal, and the effective signal is processed separately.

8. The method for reducing data error in a continuous glucose monitoring system according to claim 1, characterized in that, In step S5, the method for identifying abnormal data points includes: If the rate of change in blood glucose levels between adjacent data points exceeds a preset threshold, it is considered abnormal. The smoothed value of the signal is calculated by a sliding window. If the difference between the current value and the smoothed value exceeds the threshold, it is judged as abnormal.

9. The method for reducing data error in a continuous glucose monitoring system according to claim 1, characterized in that, The method further includes: Generate a blood glucose fluctuation trend chart and mark the high-glycemia and low-glycemia risk areas; When blood glucose levels exceed the preset safe range, an audible and visual alarm will be triggered via the terminal device.

10. A system for reducing data errors in a continuous glucose monitoring system, characterized in that, The system includes: The sensor module is used to collect current signals and environmental parameters of the subcutaneous interstitial fluid; Data processing module: includes a preprocessing unit, a feature extraction unit, a model building unit, and an error optimization unit, used to execute the method described in any one of claims 1-9; Storage module: Used to store raw data, model parameters, and patient event data; Display module: Used to output optimized blood glucose data, trend analysis results, and alarm prompts; Communication module: Used for data interaction with terminal devices.

Images