Non-invasive blood glucose detection method and system based on grid search optimization of gaussian process regression

Abstract

The present application relates to non-invasive blood glucose medical detection technical field, disclose a kind of non-invasive blood glucose detection method and system based on grid search optimization Gaussian process regression, including S1, the physiological characteristic data of at least 50 subjects is collected;The physiological characteristic data includes the body surface temperature of left neck, right neck, left axillary fossa, right axillary fossa, navel, and heart rate, weight, age, gender;S2, the physiological characteristic data of subject is regarded as data set, and is divided into training set, test set according to preset proportion;S3, using the Matern kernel function of ν=5 / 2 as basic covariance kernel, composite constant kernel and white noise kernel form overall covariance function, construct Gaussian process regression model;S4, the hyperparameter of Gaussian process regression model is globally optimized by grid search etc.;The method and system solve the problem that existing technology cannot consider detection accuracy, prediction credibility quantification and clinical safety verification simultaneously.

Description

Technical Field

[0001] This invention relates to the field of non-invasive blood glucose medical detection technology, specifically to a non-invasive blood glucose detection method and system based on grid search optimized Gaussian process regression. Background Technology

[0002] Diabetes has become a prevalent chronic metabolic disease in China. Currently, the mainstream clinical methods for blood glucose monitoring are finger-prick and venous blood sampling, both invasive methods. Repeated punctures cause pain, and patient compliance with long-term monitoring is poor. Therefore, non-invasive blood glucose monitoring technology has become a research hotspot. Existing non-invasive blood glucose technologies mainly include three categories: near-infrared spectroscopy, Raman spectroscopy, and Metabolic Heat Conformation (MHC). Among them, MHC relies on the physiological mechanism of glucose oxidation and heat production, mapping blood glucose concentration to body surface temperature. It has advantages such as low sensor hardware cost, clear physiological mechanism, and stable signal acquisition. The theoretical basis of MHC is the law of conservation of energy metabolism. After carbohydrates ingested by the human body are digested and absorbed into the bloodstream, they are completely broken down in cells through aerobic oxidation, releasing energy. Approximately 40% is stored as high-energy phosphate bonds in ATP, and approximately 60% is dissipated as heat. This metabolic heat is conducted to the body surface via blood circulation, forming a characteristic temperature distribution pattern related to blood glucose concentration in different parts of the body. By measuring temperatures at multiple body sites and combining them with physiological parameters such as heart rate, a mapping relationship from multiple physiological parameters to blood glucose concentration can be established. However, existing non-invasive blood glucose modeling schemes based on metabolic heat integration still have many shortcomings: traditional linear regression and support vector machine models are unable to characterize the complex nonlinear relationship between body surface temperature, heart rate, and basic physiological parameters and blood glucose, resulting in insufficient prediction accuracy; the hyperparameters of existing Gaussian process regression non-invasive blood glucose models rely on manual experience for configuration, lacking a systematic optimization scheme, and regularization strength and iteration restart number are prone to unreasonable settings, leading to overfitting or underfitting of the model and weak generalization ability; conventional models only output a single blood glucose prediction value, making it impossible to quantify the reliability of the prediction results, difficult to distinguish between high-reliability and low-reliability detection data, and lacking corresponding clinical risk warning mechanisms; at the same time, there is a lack of standardized sample partitioning rules adapted to multi-source metabolic heat physiological data, which can easily lead to data leakage problems caused by cross-datasets of samples from the same subject, and the model performance relies solely on MAE (mean absolute error) and R... 2 Numerical metrics such as the coefficient of determination were not evaluated, and clinical safety verification was not completed through Clark error grid and Bland-Altman consistency tests. Furthermore, the combination of body surface temperature measurement points and input features lacked an optimal selection scheme at the physiological level, and the absence or redundancy of effective features further reduced the model's prediction accuracy. Existing technologies cannot simultaneously meet the three major requirements of detection accuracy, quantifiable prediction reliability, and clinical safety verification. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a non-invasive blood glucose detection method and system based on grid search optimized Gaussian process regression. This method and system solve the problem that existing technologies cannot simultaneously achieve detection accuracy, predictive reliability quantification, and clinical safety verification.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A non-invasive blood glucose detection method based on grid search-optimized Gaussian process regression includes: S1, collecting physiological characteristic data from at least 50 subjects; the physiological characteristic data includes surface temperature of the left and right sides of the neck, left and right armpits, and navel, as well as heart rate, weight, age, and gender; S2, using the subjects' physiological characteristic data as a dataset, dividing it into training and test sets according to a preset ratio; S3, using a Matérn kernel function with ν=5 / 2 as the basic covariance kernel, and combining a composite constant kernel with a white noise kernel to form the overall covariance function, constructing a Gaussian process regression model; S4, globally optimizing the hyperparameters of the Gaussian process regression model through grid search to obtain the optimal hyperparameters; the hyperparameters include a regularization parameter α. S5, retrain the model using the optimal hyperparameters on the entire training set to obtain the optimized Gaussian process regression prediction model; S6, input the physiological characteristic data of the subjects to be tested into the optimized Gaussian process regression prediction model, and simultaneously output the blood glucose prediction value and the prediction uncertainty quantification result; S7, clinically validate the model prediction performance using a multi-dimensional indicator system; the multi-dimensional indicator system includes mean absolute error, root mean square error, coefficient of determination, mean absolute relative deviation, Clark error grid analysis, and Bland-Altman consistency analysis; S8, construct an intelligent safety early warning mechanism based on the prediction uncertainty quantification result, and output an early warning prompt when the prediction uncertainty is greater than a preset threshold.

[0005] In this invention, preferably, the expression for the Matérn kernel function in S3 is: , in, For the Matérn kernel function, For signal variance, The Euclidean distance between the input features. is the feature length scale parameter.

[0006] In this invention, preferably, the comprehensive kernel function of the Gaussian process regression model is a constant kernel multiplied by a Matérn kernel and then superimposed with a white noise kernel, expressed as: , Where, kMatern (x, x') is the Matérn kernel function, C is a constant scaling factor for the output signal amplitude, and σ 2 n Let be the variance of white noise, and δ(x, x') be the Kronecker delta function.

[0007] In this invention, preferably, the grid search in S4 employs 3-fold cross-validation, and the regularization parameter alpha has a search range of {1×10^3}. -4 1×10 -3 1×10 -2 The search range for the number of restarts was optimized to {5, 10}, with a total of 6 candidate parameter combinations; the negative mean absolute error of 3-fold cross-validation was used as the evaluation criterion for the quality of the parameters.

[0008] In this invention, preferably, the optimal hyperparameters are: regularization parameter alpha = 0.001, and optimized restart count = 10.

[0009] In this invention, preferably, the rules for collecting the physiological characteristic data of the subjects in S1 are as follows: each subject sits quietly for 5 minutes and then takes 3 consecutive measurements, with an interval of 3 to 5 minutes between each measurement, and 3 independent samples are generated for each subject.

[0010] In this invention, preferably, the preset ratio in S2 is 8:2, and during the division of the training set and the test set, the same subject sample is prohibited from entering the training set and the test set at the same time.

[0011] In this invention, preferably, the performance standards for model prediction in S7 are: mean absolute error ≤ 0.9 mmol / L, root mean square error ≤ 1.45 mmol / L, coefficient of determination ≥ 0.75, mean absolute relative deviation ≤ 11%, and the proportion of samples in the Clark error grid A+B safe interval ≥ 99%.

[0012] In this invention, preferably, the preset threshold in S8 is 2.0 mmol / L; when the predicted standard deviation is greater than 2.0 mmol / L, the device triggers an audible and visual alarm, prompting the user to repeat the measurement or use venous blood sampling to verify the blood glucose.

[0013] A non-invasive blood glucose monitoring system based on grid search-optimized Gaussian process regression includes: a data acquisition module for collecting physiological characteristic data from at least 50 subjects; the physiological characteristic data includes body surface temperature from the left and right sides of the neck, left and right armpits, and navel, as well as heart rate, weight, age, and gender; a dataset partitioning module for dividing the subjects' physiological characteristic data into training and test sets according to a preset ratio; a model building module for constructing a basic Gaussian process regression model using a Matérn kernel function with ν=5 / 2 as the basic covariance kernel, a composite constant kernel, and a white noise kernel to form the overall covariance function; and a grid search optimization module for globally optimizing the hyperparameters of the Gaussian process regression model through grid search to obtain the optimal hyperparameters; the hyperparameters include regularization... The model includes several modules: a hyperparameter alpha and a model optimization restart count; a training module for retraining the model on the entire training set using the optimal hyperparameters to obtain an optimized Gaussian process regression prediction model; a prediction output module for inputting the physiological characteristic data of the subjects to be tested into the optimized Gaussian process regression prediction model and simultaneously outputting the predicted blood glucose value and the quantification result of prediction uncertainty; a performance verification module for clinically verifying the model's prediction performance using a multi-dimensional indicator system, including mean absolute error, root mean square error, coefficient of determination, mean absolute relative deviation, Clark error grid analysis, and Bland-Altman consistency analysis; and an early warning module for constructing an intelligent safety early warning mechanism based on the quantification result of prediction uncertainty, outputting an early warning prompt when the prediction uncertainty exceeds a preset threshold.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This method and system construct a composite Gaussian process regression model using a ν=5 / 2 order Matérn kernel and automatically optimize hyperparameters using grid search combined with three-fold cross-validation. Compared to conventional models with manually configured hyperparameters, this effectively avoids overfitting and underfitting, significantly improving the model's generalization ability. After optimization, the mean absolute relative error (MARD) of the model is reduced to 10.67%, meeting the accuracy requirements of non-invasive blood glucose devices. The mean absolute error (MAE) is no higher than 0.9 mmol / L, and the coefficient of determination (R²) is [not specified in the original text]. 2 Achieving a value above 0.75; simultaneously outputting the predicted mean and standard deviation of blood glucose, quantifying the uncertainty of each test result; establishing a tiered early warning mechanism based on preset thresholds to promptly prompt verification of low-reliability measurement results, mitigating the risk of missed clinical detections; splitting the dataset into layers by subject ID to prevent data leakage caused by samples from the same subject crossing sets; simultaneously selecting a nine-dimensional physiological feature set including five body surface temperatures, heart rate, and basic human information, conforming to the metabolic pattern of glucose thermogenesis in the human body, reducing accuracy loss caused by feature redundancy or missing effective features; employing MAE (mean absolute error), RMSE (root mean square error), and R...2 The model was validated using a multi-dimensional quantitative evaluation system combining the coefficient of determination (CDO), MARD (mean absolute relative deviation), and Clark error grid. After optimization, 99.37% of the samples were within the safe range of Clark error grid A+B, with no high-risk treatment error range. Paired tests showed no significant systematic bias, and the prediction results were interchangeable with the true values ​​from venous blood sampling. The entire algorithm process was standardized and reproducible, requiring no manual parameter tuning. It could be directly embedded into non-invasive blood glucose testing hardware, balancing detection accuracy, clinical safety, and the practicality of long-term home monitoring. Attached Figure Description

[0015] Figure 1 This is a flowchart of the non-invasive blood glucose detection method based on grid search and optimized Gaussian process regression according to the present invention.

[0016] Figure 2 This is a Clark error grid analysis diagram of the Gaussian process regression model in the non-invasive blood glucose detection method based on grid search optimization of Gaussian process regression in this invention, applied to 317 test samples.

[0017] Figure 3 This is a Bland-Altman consistency analysis diagram of the Gaussian process regression model in the non-invasive blood glucose detection method based on grid search optimization of Gaussian process regression in this invention.

[0018] Figure 4 This is a schematic diagram of the non-invasive blood glucose detection system based on grid search and optimized Gaussian process regression according to the present invention.

[0019] In the attached diagram: 1. Data acquisition module; 2. Dataset partitioning module; 3. Model building module; 4. Grid search optimization module; 5. Training module; 6. Prediction output module; 7. Performance verification module; 8. Early warning module. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Please see Figure 1 A preferred embodiment of the present invention provides a non-invasive blood glucose detection method based on grid search optimized Gaussian process regression, comprising: S1. Collect physiological characteristic data from at least 50 subjects. The physiological characteristic data includes body surface temperature of the left and right sides of the neck, left and right armpits, and navel, as well as heart rate, weight, age, and gender.

[0023] Glucose is oxidized and decomposed within the body, releasing heat. Body surface temperature distribution directly reflects the body's glucose metabolism level; therefore, it is necessary to collect data from locations that reflect core body temperature and peripheral blood flow temperature. First, five core temperature measurement areas were selected: the left and right sides of the neck, the left and right sides of the armpits, and the navel. These locations are close to the aorta and the core abdominal cavity, less affected by external environmental interference, and the temperature data can accurately reflect metabolic heat production. In addition, four basic physiological indicators were collected: heart rate, weight, age, and gender. Heart rate represents the rate of blood circulation, while weight, age, and gender affect the body's basal metabolic rate. These factors all alter heat conduction efficiency, collectively forming a complete nine-dimensional input feature. In terms of data collection, subjects were required to sit quietly for 5 minutes to stabilize their metabolic level, avoiding temperature fluctuations caused by exercise. Each subject was measured three times consecutively, with an interval of 3 to 5 minutes between adjacent measurements. Three independent samples were generated for each subject to ensure sample stability. Gender characteristics were binary encoded, with males assigned a value of 1 and females a value of 0, transforming the categorical variable into a numerical variable recognizable by the model. Other parameters, such as temperature, heart rate, age, and weight, were directly represented using measured continuous values, completing the construction of a structured sample and providing a standardized raw dataset for subsequent modeling. A larger number of subjects results in a more comprehensive dataset and better training performance; therefore, at least 50 subjects were required. In this implementation, 528 subjects were actually included.

[0024] S2 uses the physiological characteristic data of the subjects as a dataset and divides it into training set and test set according to a preset ratio.

[0025] First, all measurement samples from the subjects were collected, totaling 1584 valid samples, corresponding to 528 subjects. The partitioning process employed stratified random sampling based on the subject's unique ID, strictly prohibiting the simultaneous allocation of samples from the same subject to both the training and test sets. This prevented the model from learning individual subject characteristics rather than the general patterns between blood glucose and physiological parameters. The samples were split according to a preset ratio, ideally between 7:3 and 9:1, with an optimal ratio of 8:2. This means 80% of the samples (1267 samples) were allocated to the training set for model training and parameter tuning, and the remaining 20% ​​(317 samples) were allocated to the independent test set solely for final model performance evaluation. After partitioning, all 9-dimensional features underwent Z-score standardization to eliminate dimensional differences between features such as temperature, weight, and heart rate. This ensured that the values ​​of each feature were within the same range, guaranteeing that the Gaussian process regression kernel function would not be dominated by any single feature with an excessively large value when calculating the Euclidean distance, thus ensuring the objectivity and validity of the distance calculation results.

[0026] S3 uses the Matérn kernel function with ν=5 / 2 as the basic covariance kernel, and the composite constant kernel and the white noise kernel form the overall covariance function to construct the basic model of Gaussian process regression.

[0027] Gaussian process regression (GPR) is a nonparametric probabilistic model based on Bayesian theory. A Gaussian process is completely determined by its mean function m(x) and covariance function (kernel function) k(x, x')

[11] . Assuming the mean function is zero, consider the observation model y = f(x) + ε with Gaussian noise, where ε ~ N(0, σ 2Given training data, the predicted mean μ* and predicted variance Σ* for the test point x* can be analytically calculated using the covariance matrix of the training data, without the need for explicit model training. First, the mean function of the Gaussian process is set to zero to simplify the model's computational complexity. Then, a composite covariance kernel function is constructed, using a ν=5 / 2 order Matérn kernel as the basic kernel function. The sample curve corresponding to this kernel function possesses second-order differentiable smoothness, avoiding both the excessively volatile curves of lower-order Matérn kernels and the overly smooth curves of infinitely smooth RBF kernels, thus accurately matching the physiological pattern of continuous and gradual changes in human body surface temperature. The reason for choosing the Matérn 5 / 2 kernel instead of the Radial Basis Function (RBF) kernel is that MHC temperature data exhibits finite smoothness in space—differences in blood perfusion at different body surface sites, changes in skin thickness, and other factors mean that the spatial gradient of temperature is not infinitely differentiable. The Matérn kernel provides flexible control over the smoothness of the function through the parameter ν, where ν = 5 / 2 corresponds to a twice-differentiable sample function. This assumption of moderate smoothness matches the actual characteristics of MHC temperature data. Specifically, the Matérn kernel function expression is: , in, For the Matérn kernel function, For signal variance, The Euclidean distance between the input features. is the feature length scale parameter. The feature length scale parameter is used to measure the similarity between two sets of physiological feature vectors. The smoothness of the model fitting curve and the signal variance. The overall fluctuation range of the predicted blood glucose value is constrained. Based on this, the Matérn kernel is multiplied by a constant kernel, and then a white noise kernel is superimposed to form a comprehensive kernel function. The constant kernel is used to scale the covariance amplitude, and the white noise kernel is used to fit the random errors caused by sensor noise and minor human perturbations during the measurement process. The Kronecker function ensures that noise terms are introduced only when calculating the same sample. Finally, a complete covariance matrix is ​​formed, establishing a nonlinear mapping relationship between physiological input features and blood glucose output values.

[0028] The combined kernel function of the Gaussian process regression model is a constant kernel multiplied by a Matérn kernel, then superimposed with a white noise kernel. The expression is: , Where, k Matern (x, x') is the Matérn kernel function, C is a constant scaling factor for the output signal amplitude, and σ 2 nLet be the variance of white noise, and δ(x, x') be the Kronecker delta function.

[0029] S4. The hyperparameters of the Gaussian process regression model are globally optimized through grid search to obtain the optimal hyperparameters. The hyperparameters include the regularization parameter alpha and the number of model optimization restarts.

[0030] The regularization parameter alpha and the number of optimization restarts in a Gaussian process regression model directly determine the model's fitting performance. Manual parameter tuning is prone to getting stuck in local optima. Therefore, a grid search is used to traverse candidate parameters, combined with cross-validation to select the optimal combination of hyperparameters. First, the parameter search grid is defined: the candidate set for the regularization parameter alpha is {1×10...} -4 1×10 -3 1×10 -2 A larger alpha value indicates stronger model regularization constraints, making overfitting less likely. The candidate set for the number of optimization restarts is {5, 10}. Multiple restarts allow the kernel hyperparameter iterations to escape local optima, resulting in a total of 6 candidate parameter combinations. Subsequently, 3-fold stratified cross-validation is performed only within the training set, dividing the training set into 3 equal subsets. In each round of training, 2 subsets are used to train the model, and the remaining subset serves as a temporary validation set. Strict adherence to subject stratification rules is maintained to prevent data sharing. Negative MAE (mean absolute error) is used as the evaluation score for each parameter combination; a higher score indicates a smaller prediction error on the validation set. After fully traversing all 6 parameter combinations, the average scores of each 3-fold cross-validation combination are compared to select the optimal parameter combination: alpha = 0.001, number of optimization restarts = 10. This balances the model's fitting strength while effectively avoiding local optima, eliminating the need for repeated manual parameter adjustments, and the optimization process is reproducible.

[0031] S5. Retrain the model on the entire training set using the optimal hyperparameters to obtain the optimized Gaussian process regression prediction model.

[0032] After determining the optimal hyperparameters, the regularization parameters and the number of restarts were assigned to the composite kernel Gaussian process regression model. The entire training set of 1267 samples was then used for retraining without further data splitting. The model solved for the covariance matrix based on the input 9-dimensional physiological feature matrix. After iterative convergence, all internal parameters of the Matérn kernel—length scale, signal variance, and white noise variance—were fixed, and the entire set of model parameters was saved. Training with the full set of training samples fully utilized the data to learn the correlation between physiological features and blood glucose levels. Compared to models trained using only a subset of cross-validation samples, the final trained model exhibited stronger fitting ability and generalization performance, providing a stable inference model for subsequent non-invasive blood glucose prediction.

[0033] S6 inputs the physiological characteristic data of the subject to be tested into the optimized Gaussian process regression prediction model, and simultaneously outputs the predicted blood glucose value and the quantification result of the prediction uncertainty.

[0034] The standardized 9-dimensional physiological feature vector of the test subject is input into a trained Gaussian process regression model. Unlike ordinary machine learning models, Gaussian process regression simultaneously outputs two types of results: the predicted mean and the predicted variance. The predicted mean corresponds to the non-invasive blood glucose concentration value of this test, which is the final blood glucose test result. The square root of the predicted variance yields the predicted standard deviation, which serves as a quantitative indicator of the uncertainty of the prediction result. If the subject's physiological characteristics fall within the coverage of the model's training data, the predicted standard deviation is smaller, indicating high reliability of the test. If the test sample deviates from the distribution of the training samples, the predicted standard deviation will increase significantly, indicating insufficient reliability of the prediction result. This simultaneous output of predicted values ​​and reliability indicators overcomes the limitation of traditional models that can only output a single value, providing a quantitative basis for subsequent clinical risk grading and early warning.

[0035] S7 employs a multi-dimensional indicator system to clinically validate the model's predictive performance. This multi-dimensional indicator system includes mean absolute error, root mean square error, coefficient of determination, mean absolute relative deviation, Clark error grid analysis, and Bland-Altman consistency analysis.

[0036] This step verifies the model's effectiveness from two perspectives: numerical error metrics and clinical safety. First, four numerical quantitative metrics are calculated: Mean Absolute Error (MAE) directly reflects the average deviation between the predicted value and the true venous blood glucose value; Root Mean Square Error (RMSE) amplifies the impact of large error samples and measures the level of extreme errors; and the Coefficient of Determination (R²)... 2 The model's ability to interpret variations in blood glucose data is demonstrated by the Mean Absolute Relative Deviation (MARD), which corresponds to the general accuracy standard in the medical device industry, used to verify whether it meets the technical requirements of non-invasive blood glucose devices. Two additional clinical verification methods are added: Clark error grid analysis divides the predicted and true values ​​into five risk intervals: A, B, C, D, and E. Interval A represents accurate predictions, interval B represents benign errors, and intervals C and E represent high-risk errors that could lead to diagnostic errors. The proportion of samples in interval A+B determines the clinical safety level; a higher proportion indicates better clinical safety. Bland-Altman consistency analysis examines whether there is a systematic bias between the predicted and true venous blood sampling values. It calculates the average deviation and 95% consensus threshold of the two sets of data, and then uses a paired t-test to determine whether there is a statistically significant systematic bias. This entire evaluation system balances numerical accuracy and clinical safety, avoiding ignoring clinical risks by solely relying on numerical index optimization, and fully validating the model's usability in medical scenarios.

[0037] In this embodiment, to verify the effectiveness of grid search hyperparameter optimization, a baseline GPR model was first established with default parameters (alpha = 1 × 10⁻⁶, optimization restarts = 0). The performance of the default GPR on the test set was: MAE = 0.951 mmol / L, RMSE = 1.512 mmol / L, R 2 =0.738, MARD=12.13%. The performance of the training set and the test set are similar, indicating that the model does not show significant overfitting.

[0038] After systematically evaluating six candidate parameter combinations using grid search cross-validation and 3-fold cross-validation, the optimal parameter configuration was determined to be: alpha = 0.001 and optimization restart count = 10. Retraining with this optimal configuration significantly improved the model's performance on the test set: MAE = 0.864 mmol / L (a 9.2% reduction compared to the default model), RMSE = 1.439 mmol / L (a 4.8% reduction), and R... 2 =0.765 (an improvement of 3.7%), MARD=10.67% (a decrease of 12.0%). Table 1 summarizes the performance comparison data of the model before and after optimization.

[0039] This result indicates that appropriately increasing the regularization parameter alpha helps enhance the numerical stability and generalization ability of the model; increasing the number of optimization restarts allows the optimization process of the kernel hyperparameters to explore the parameter space more fully and effectively avoid getting trapped in local optima.

[0040] Figure 2 The CEGA analysis results of Gaussian process regression on 317 test samples are presented. Of the 317 test samples, 82.97% (263 cases) were in zone A, 16.40% (52 cases) in zone B, and 0.63% (2 cases) in zone D, for a combined zone A+B ratio of 99.37%. No samples fell into zones C or E. The two zone D samples both showed reference values ​​of hyperglycemia (approximately 236 mg / dL, or 13.1 mmol / L) but significantly lower predicted values ​​(approximately 114–138 mg / dL). This bias may be related to insufficient training samples in the hypoglycemic zone, leading to model prediction bias. Overall, the combined 99.37% coverage of zones A+B indicates that the model's predictions have good clinical safety.

[0041] Bland-Altman consistency analysis was used to assess the degree of consistency between predicted GPR values ​​and reference values ​​for venous blood glucose concentration. Figure 3The Bland-Altman analysis results of the GPR tuning model are presented. The mean bias between the predicted and reference values ​​is -1.6 mg / dL, and the 95% agreement limit (LoA) is [-52.4, 49.3] mg / dL. The vast majority of sample points fall within the agreement limit, indicating good agreement between the GPR predictions and the reference method.

[17] The small absolute value of the systematic bias indicates that the model does not have a significant tendency to overestimate or underestimate at the population level.

[0042] Specifically, the performance criteria for model prediction are: Mean Absolute Error (MAE) ≤ 0.9 mmol / L, Root Mean Square Error (RMSE) ≤ 1.45 mmol / L, and Coefficient of Determination (R²). 2 ≥0.75, mean absolute relative deviation MARD≤11%, Clark error grid A+B safe interval sample proportion ≥99%.

[0043] S8 constructs an intelligent safety early warning mechanism based on the quantification results of prediction uncertainty, and outputs an early warning prompt when the prediction uncertainty is greater than a preset threshold.

[0044] First, a fixed threshold is defined based on the distribution of the predicted standard deviation of all samples in the test set. Combined with the actual measured data of this scheme, the warning threshold is set to 2.0 mmol / L. During real-time detection, the model automatically compares the predicted standard deviation with the preset threshold after each blood glucose prediction: when the predicted standard deviation is less than or equal to 2.0 mmol / L, the predicted blood glucose value is directly output for user reference; when the predicted standard deviation exceeds 2.0 mmol / L, the reliability of the measurement result is deemed insufficient, and the device immediately triggers a dual warning with a pop-up window and audio-visual alert, prompting the subject to sit still to stabilize metabolism and remeasure, or to go to a medical institution for a venous blood sample to verify the blood glucose value. This mechanism can automatically identify low-quality test results, proactively avoiding misdiagnosis and missed diagnosis caused by model prediction bias, and significantly improving the clinical safety of non-invasive blood glucose devices.

[0045] This invention also provides a non-invasive blood glucose detection system based on grid search optimized Gaussian process regression, such as... Figure 4 As shown, it includes: Data acquisition module 1 is used to collect physiological characteristic data from at least 50 subjects. The physiological characteristic data includes body surface temperature of the left side of the neck, the right side of the neck, the left side of the armpit, the right side of the armpit, and the navel, as well as heart rate, weight, age, and gender. Data set partitioning module 2 is used to divide the physiological characteristic data of the subjects into training set and test set according to a preset ratio. Model building module 3 is used to construct the basic model of Gaussian process regression by using the Matérn kernel function with ν=5 / 2 as the basic covariance kernel, the composite constant kernel and the white noise kernel to form the overall covariance function; The grid search optimization module 4 is used to globally optimize the hyperparameters of the Gaussian process regression model through grid search to obtain the optimal hyperparameters; the hyperparameters include the regularization parameter alpha and the number of model optimization restarts; Training module 5 is used to retrain on the entire training set using the optimal hyperparameters to obtain the optimized Gaussian process regression prediction model; The prediction output module 6 is used to input the physiological characteristic data of the subject to be tested into the optimized Gaussian process regression prediction model and simultaneously output the blood glucose prediction value and the prediction uncertainty quantification result. Performance validation module 7 is used to clinically validate the model's predictive performance using a multi-dimensional indicator system. The multi-dimensional indicator system includes mean absolute error, root mean square error, coefficient of determination, mean absolute relative deviation, Clark error grid analysis, and Bland-Altman consistency analysis. Early warning module 8 is used to construct an intelligent safety early warning mechanism based on the quantification results of prediction uncertainty, and outputs an early warning prompt when the prediction uncertainty is greater than a preset threshold.

[0046] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A non-invasive blood glucose detection method based on grid search optimized Gaussian process regression, characterized in that, include: S1, collect physiological characteristic data from at least 50 subjects; The physiological characteristic data include the surface temperature of the left side of the neck, the right side of the neck, the left side of the armpit, the right side of the armpit, and the navel, as well as heart rate, weight, age, and gender; S2, the physiological characteristic data of the subjects are used as a dataset and divided into training set and test set according to a preset ratio; S3 uses the Matérn kernel function with ν=5 / 2 as the basic covariance kernel, and the composite constant kernel and the white noise kernel form the overall covariance function to construct a Gaussian process regression model; S4. The hyperparameters of the Gaussian process regression model are globally optimized through grid search to obtain the optimal hyperparameters; the hyperparameters include the regularization parameter alpha and the number of model optimization restarts. S5. Retrain on the entire training set using the optimal hyperparameters to obtain the optimized Gaussian process regression prediction model. S6, input the physiological characteristic data of the subject to be tested into the optimized Gaussian process regression prediction model, and simultaneously output the blood glucose prediction value and the prediction uncertainty quantification result; S7. A multi-dimensional indicator system is used to clinically validate the model's predictive performance. The multi-dimensional indicator system includes mean absolute error, root mean square error, coefficient of determination, mean absolute relative deviation, Clark error grid analysis, and Bland-Altman consistency analysis. S8 constructs an intelligent safety early warning mechanism based on the quantification results of prediction uncertainty, and outputs an early warning prompt when the prediction uncertainty is greater than a preset threshold.

2. The non-invasive blood glucose detection method according to claim 1, characterized in that, The expression for the Matérn kernel function in S3 is: ， in, For the Matérn kernel function, For signal variance, The Euclidean distance between the input features. is the feature length scale parameter.

3. The non-invasive blood glucose detection method according to claim 2, characterized in that, The comprehensive kernel function of the Gaussian process regression model is a constant kernel multiplied by a Matérn kernel, then superimposed with a white noise kernel, and its expression is: ， Where, k Matern (x, x') is the Matérn kernel function, C is a constant scaling factor for the output signal amplitude, and σ 2 n Let be the variance of white noise, and δ(x, x') be the Kronecker delta function.

4. The non-invasive blood glucose detection method according to claim 3, characterized in that, The grid search in S4 employs 3-fold cross-validation, and the regularization parameter alpha searches within the range of {1×10⁻⁶}. -4 1×10 -3 1×10 -2 The search range for the number of restarts was optimized to {5, 10}, with a total of 6 candidate parameter combinations; the negative mean absolute error of 3-fold cross-validation was used as the evaluation criterion for the quality of the parameters.

5. The non-invasive blood glucose detection method according to claim 3, characterized in that, The optimal hyperparameters are: regularization parameter alpha = 0.001, and optimized restart count = 10.

6. The non-invasive blood glucose detection method according to claim 3, characterized in that, The rules for collecting the physiological characteristics data of the subjects in S1 are as follows: each subject sits quietly for 5 minutes and then takes 3 consecutive measurements, with an interval of 3 to 5 minutes between each measurement, and 3 independent samples are generated for each subject.

7. The non-invasive blood glucose detection method according to claim 3, characterized in that, The preset ratio in S2 is 8:

2. During the division of the training set and the test set, it is prohibited for the same subject sample to enter both the training set and the test set at the same time.

8. The non-invasive blood glucose detection method according to claim 3, characterized in that, The performance standards for model prediction in S7 are as follows: mean absolute error ≤ 0.9 mmol / L, root mean square error ≤ 1.45 mmol / L, coefficient of determination ≥ 0.75, mean absolute relative deviation ≤ 11%, and the proportion of samples in the Clark error grid A+B safe interval ≥ 99%.

9. The non-invasive blood glucose detection method according to claim 3, characterized in that, The preset threshold in S8 is 2.0 mmol / L; when the predicted standard deviation is greater than 2.0 mmol / L, the device triggers an audible and visual alarm, prompting the user to repeat the measurement or use venous blood sampling to verify the blood glucose.

10. A non-invasive blood glucose detection system based on grid search optimized Gaussian process regression, characterized in that, include: The data acquisition module is used to collect physiological characteristic data from at least 50 subjects; The physiological characteristic data include the surface temperature of the left side of the neck, the right side of the neck, the left side of the armpit, the right side of the armpit, and the navel, as well as heart rate, weight, age, and gender; The dataset partitioning module is used to divide the physiological characteristic data of the subjects into training set and test set according to a preset ratio. The model building module is used to construct the basic model of Gaussian process regression by using the Matérn kernel function with ν=5 / 2 as the basic covariance kernel, the composite constant kernel and the white noise kernel to form the overall covariance function; The grid search optimization module is used to globally optimize the hyperparameters of the Gaussian process regression model through grid search to obtain the optimal hyperparameters; the hyperparameters include the regularization parameter alpha and the number of model optimization restarts; The training module is used to retrain the Gaussian process regression prediction model on the entire training set using the optimal hyperparameters; The prediction output module is used to input the physiological characteristic data of the subject to be tested into the optimized Gaussian process regression prediction model, and simultaneously output the blood glucose prediction value and the prediction uncertainty quantification result. The performance validation module is used to clinically validate the model's predictive performance using a multi-dimensional indicator system, which includes mean absolute error, root mean square error, coefficient of determination, mean absolute relative deviation, Clark error grid analysis, and Bland-Altman consistency analysis. The early warning module is used to build an intelligent safety early warning mechanism based on the quantification results of prediction uncertainty. When the prediction uncertainty is greater than a preset threshold, an early warning prompt is output.

Images