Device for predicting whether there is rise in blood sugar

Abstract

The purpose of the present disclosure is to provide means for identifying incipient diabetics. The present disclosure provides a device for predicting whether there is a rise in the blood sugar of a subject. More specifically, provided is a device for predicting whether there is a rise in the blood sugar of a subject, wherein the device includes: a data acquisition unit for acquiring electrocardiogram data of the subject; a blood sugar rise prediction unit for predicting whether there is a rise in blood sugar by inputting the electrocardiogram data into a machine learning model; and a results presentation unit for presenting the prediction results.

Description

Device for predicting the presence or absence of blood sugar increase

[0001] The present disclosure relates to a device, a method, and a program for predicting the presence or absence of blood sugar increase in a subject.

[0002] With changes in lifestyle and social environment, the number of diabetic patients has been increasing rapidly. In order to suppress the onset of complications, it is extremely important to detect diabetes at the pre-onset stage. Diabetes significantly reduces the quality of life (QOL) of patients due to various complications. Although prognosis can be improved by treatment intervention, since the disease state is irreversible at the time of diagnosis of diabetes, it is rarely curable.

[0003] Non-Patent Document 1 reports on the long-term risk assessment of type 2 diabetes onset based on deep learning using electrocardiograms in a non-diabetic population. It is stated that in the non-diabetic group, those whose electrocardiograms are classified as diabetes by DLM have a higher risk of type 2 diabetes onset compared to those who are not, but it does not teach a method for detecting blood sugar increase from electrocardiogram data.

[0004] Patent Document 1 discloses the following method for diagnosing myocardial infarction using an electrocardiogram. However, it does not teach predicting blood sugar increase from an electrocardiogram.

[0005] WO2024019584A1

[0006] eClinicalMedicine 2024;68: 102445, Published Online 1 February 2024, https: / / doi.org / 10.1016 / j.eclinm.2024.102445

[0007] Diabetes is an independent risk factor for heart failure, and it has been reported that the onset risk of heart failure is already significantly high at the stage of borderline diabetes, that is, the pre-onset stage of diabetes onset. This suggests that some influence is being exerted on the heart from the stage of blood sugar increase before reaching diabetes, reminding us of the possibility that this influence is manifested in the electrocardiogram.

[0008] If diabetes can be diagnosed at the stage of elevated blood glucose (i.e., the early "pre-diabetic" stage), the condition is considered reversible, and the onset can be prevented through therapeutic intervention. However, since there are no symptoms at the elevated blood glucose stage, it is unlikely that people will seek medical attention at this stage. While detection in regular health checkups is technically possible, HbA1c levels, one of the main indicators of diabetes, are not always measured in regular health checkups. Furthermore, regular health checkups are generally conducted only once a year, and it is not uncommon for people to have already developed diabetes by the time they receive a checkup. Moreover, there are a significant number of people who do not even receive a health checkup. As described above, the purposeful detection of early pre-diabetes in medical institutions is not easy. If methods could be developed that would incidentally reveal diabetes (or pre-diabetes) when a person visits a medical institution for another purpose, or if simple methods that can be performed at home could be developed, it would be possible to detect more early pre-diabetes and contribute to preventing the onset of the disease.

[0009] The inventors focused on the finding that diabetes increases the risk of heart failure from an early stage and developed a method to detect elevated blood glucose levels from electrocardiogram data (i.e., a method to detect early-stage "pre-diabetic individuals").

[0010] This disclosure provides the following as more specific examples: [Embodiment 1] A device for predicting whether or not blood glucose levels rise in a subject, comprising: a data acquisition unit for acquiring electrocardiogram data of the subject; a blood glucose elevation prediction unit for inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels rise; and a result presentation unit for presenting the prediction results. [Embodiment 2] The device according to Embodiment 1, which is a wearable device. [Embodiment 3] The device according to Embodiment 1, wherein the electrocardiogram data is 12-lead electrocardiogram data or I-lead electrocardiogram data. [Embodiment 4] The device according to Embodiment 1, wherein the machine learning model is a pre-trained machine learning model. [Embodiment 5] The device according to Embodiment 1, wherein the machine learning model is a pre-trained machine learning model obtained through supervised learning. [Embodiment 6] The device according to Embodiment 1, wherein the machine learning model is CatBoost. [Embodiment 7] The device according to Embodiment 4, wherein, during training of the pre-trained machine learning model, data including pairs of the subject's electrocardiogram data and data indicating the subject's fasting blood glucose, HbA1c, or diabetes treatment status is input to the machine learning model. [Embodiment 8] The apparatus according to Embodiment 7, wherein the number of subjects is 10,000 or more. [Embodiment 9] The apparatus according to Embodiment 7, wherein the electrocardiogram data of a subject and the data indicating the subject's fasting blood glucose, HbA1c, or diabetes treatment status are data obtained from measurements performed on the same day, or are data for a subject on the same day. [Embodiment 10] The apparatus according to Embodiment 7, wherein the electrocardiogram data of a subject is feature data extracted from the electrocardiogram. [Embodiment 11] The apparatus according to Embodiment 10, wherein feature extraction from the feature data extracted from the electrocardiogram is performed using software. [Embodiment 12] The apparatus according to Embodiment 11, wherein the software is an electrocardiogram feature extraction program. [Embodiment 13] The apparatus according to Embodiment 10, wherein the number of features in the feature data extracted from the electrocardiogram is 100 or more and 500 or less. [Embodiment 14] The apparatus according to Embodiment 10, wherein the feature data extracted from the electrocardiogram is calculated for each of the 12 leads.[Embodiment 15] The apparatus according to Embodiment 10, wherein the feature data extracted from the electrocardiogram includes only the feature data of lead I for items related to heart rate. [Embodiment 16] The apparatus according to Embodiment 10, wherein the feature data extracted from the electrocardiogram does not include items related to heart rate. [Embodiment 17] The apparatus according to Embodiment 10, wherein the feature data extracted from the electrocardiogram does not include features where a predetermined percentage or more of the total number of data sets containing pairs of the subject's electrocardiogram data and the subject's fasting blood glucose, HbA1c, or data indicating that the subject is undergoing diabetes treatment are missing values. [Embodiment 18] The apparatus according to Embodiment 10, wherein the feature data extracted from the electrocardiogram excludes features where a predetermined percentage or more of the total number of data sets containing pairs of the subject's electrocardiogram data and the subject's fasting blood glucose, HbA1c, or data indicating that the subject is undergoing diabetes treatment are missing values. [Embodiment 19] The apparatus according to Embodiment 7, wherein the data indicating the subject's fasting blood glucose, HbA1c, or the status of diabetes treatment includes data indicating classification based on whether the fasting blood glucose or HbA1c is above a predetermined value. [Embodiment 20] The apparatus according to Embodiment 7, wherein, in training a pre-trained machine learning model, data including pairs of the subject's electrocardiogram data and the subject's fasting blood glucose, HbA1c, or the status of diabetes treatment is divided into training data and test data in a ratio of 10:1 to 7:1. [Embodiment 21] The apparatus according to Embodiment 4, wherein, in training a pre-trained machine learning model, the parameters of the machine learning model are set by an optimization tool. [Embodiment 22] The apparatus according to Embodiment 4, wherein, in training a pre-trained machine learning model, a predetermined number of training sessions are performed using a predetermined number of patterns obtained by further dividing the training data into training data and validation data. [Embodiment 23] The apparatus according to Embodiment 22, wherein the predetermined number is 5 or more. [Embodiment 24] The apparatus according to Embodiment 4, wherein, in training a pre-trained machine learning model, a setting is made to impose a larger penalty for prediction errors in minority classifications as a training parameter.[Embodiment 25] The apparatus according to Embodiment 4, wherein training is performed after optimization of parameter settings in training with a pre-trained machine learning model. [Embodiment 26] The apparatus according to Embodiment 25, wherein a predetermined number of training sessions are performed using a predetermined number of patterns obtained by further dividing the training data into training data and validation data in training performed after optimization of parameter settings. [Embodiment 27] The apparatus according to Embodiment 4, wherein data including a set of the subject's electrocardiogram data, the subject's fasting blood glucose, HbA1c, or data indicating that they are undergoing diabetes treatment, and data obtained from the subject's health checkup is input to the machine learning model in training with a pre-trained machine learning model. [Embodiment 28] The apparatus according to Embodiment 27, wherein the data obtained from the subject's health checkup is data indicating one or more subject information selected from the group consisting of the subject's age, sex, height, weight, BMI, waist circumference, smoking status, and frequency of alcohol consumption. [Embodiment 29] The apparatus according to Embodiment 28, wherein the data indicating one or more subject information is data indicating four or more subject information. [Embodiment 30] The apparatus according to Embodiment 7, wherein the electrocardiogram data of the subject is data generated from data recorded by an electrocardiograph. [Embodiment 31] The apparatus according to Embodiment 7, wherein the electrocardiogram data of the subject is image data of an electrocardiogram waveform as raw data, numerical data or image data obtained by processing to convert to an electrocardiogram waveform, raw electrocardiogram potential data, electrocardiogram potential data of 12 sampled leads, feature data extracted from electrocardiogram potential data, or data vectorized from electrocardiogram potential data using a deep learning model. [Embodiment 32] A method for supporting the early detection of prediabetes, comprising the steps of acquiring target electrocardiogram data, inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated, and presenting the prediction results, wherein the prediction of elevated blood glucose levels serves as an indicator that the person is prediabetic. [Embodiment 33] The method according to Embodiment 32, further comprising the step of confirming the actual measured blood glucose level of the subject.[Embodiment 34] A program for predicting whether or not blood glucose levels are elevated in a subject, wherein one or more processors perform the steps of acquiring electrocardiogram data of the subject, inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated, and presenting the prediction results. [Embodiment 35] A non-temporary computer-readable recording medium storing the program described in Embodiment 34. [Embodiment 36] A trained machine learning model used in a device for predicting whether or not blood glucose levels are elevated in a subject, wherein the trained machine learning model is generated by training using training data, with the subject's electrocardiogram data as the explanatory variable and the subject's fasting blood glucose, HbA1c, or data indicating that they are undergoing diabetes treatment as the dependent variable. [Embodiment 37] A device for predicting whether or not blood glucose levels are elevated in a subject, comprising: a data acquisition unit for acquiring electrocardiogram data of the subject; a blood glucose elevation prediction unit for inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated; and a result presentation unit for presenting the prediction results, wherein blood glucose elevation is defined as including fasting blood glucose ≥ 110 mg / dL or HbA1c ≥ 6.0%, the presence of blood glucose elevation indicates that the subject may be pre-diabetic, and the features input into the machine learning model include at least features that suggest signs of left ventricular hypertrophy or the standard deviation of the RR interval. [Embodiment 38] The device according to Embodiment 37, which predicts whether or not blood glucose levels are elevated in a subject in order to predict whether the subject is pre-diabetic. [Embodiment 39] The device according to Embodiment 37 or 38, wherein the features that suggest signs of left ventricular hypertrophy include features selected from the group consisting of the R wave height of lead aVL, the R wave height of lead I, and the S wave height of lead V4. [Embodiment 40] A device for predicting whether a target person is pre-diabetic, comprising: a data acquisition unit for acquiring electrocardiogram data of the target person; a blood glucose elevation prediction unit for inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated; and a result presentation unit for presenting the prediction results.[Embodiment 41] A device for predicting whether a target is pre-diabetic, comprising: a data acquisition unit for acquiring electrocardiogram data of the target; a blood glucose elevation prediction unit for inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated; and a result presentation unit for presenting the prediction results, wherein blood glucose elevation is defined as including fasting blood glucose ≥ 110 mg / dL or HbA1c ≥ 6.0%, the presence of blood glucose elevation indicates that the target may be pre-diabetic, and the features input to the machine learning model include at least features that suggest signs of left ventricular hypertrophy or the standard deviation of the RR interval. [Embodiment 42] A program for predicting whether or not blood glucose levels are elevated in a target in order to support the early detection of pre-diabetes, comprising: causing one or more processors to perform the steps of acquiring electrocardiogram data of the target; inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated; and presenting the prediction results.

[0011] Figure 1 shows a schematic configuration of an exemplary computer that may be used to implement an embodiment of the present disclosure. The exemplary computer (100) includes a control unit (101), a memory unit (102), a peripheral device interface unit (103), an input unit (104), a display unit (105), a communication unit (106), and a bus (110). This computer (100) may connect to an external server (130) and a database (140) via a network (120). Figure 2 shows the predictive performance of a CatBoost model trained using 269 features of a 12-lead electrocardiogram. AUROC stands for Area Under the ROC Curve, and FPR stands for False Positive Rate. Figure 3 shows the predictive performance of five models trained using 269 features of a 12-lead electrocardiogram, compared to CatBoost. AUROC stands for Area Under the ROC Curve, meaning the area under the ROC curve. FPR stands for False Positive Rate, meaning the false positive rate. Figure 4 shows the predictive performance of a CatBoost model trained using 28 electrocardiogram features from lead I only. AUROC stands for Area Under the ROC Curve, meaning the area under the ROC curve. FPR stands for False Positive Rate, meaning the false positive rate. Figure 5 shows the SHAP values ​​calculated for the contribution of each feature in a predictive model trained using 269 features from a 12-lead electrocardiogram, and shows the top 25 features in terms of contribution. The figure on the left shows the direction of contribution of each feature, with the scale from gray to black indicating the value of each feature, and the spread from left to right indicating the magnitude of the contribution. The histogram on the right shows the contribution of the corresponding feature to the prediction as a percentage. Figure 6 shows the change in the predictive performance of the CatBoost model when the classification threshold is fixed at fasting blood glucose = 110 mg / dL and the HbA1c value is varied. Figure 7 is a table showing the predictive performance (AUROC) when the classification criteria are changed. The vertical axis corresponds to the HbA1c threshold, and the horizontal axis corresponds to the fasting blood glucose (FPG) threshold.Figure 8 shows the predictive performance when a total of 36 items are input, combining the electrocardiogram features of lead I only (Figure 4) with eight items of subject information (age, sex, height, weight, BMI, waist circumference, smoking status, and frequency of alcohol consumption). AUROC stands for Area Under the ROC Curve, meaning the area under the ROC curve. FPR stands for False Positive Rate, meaning the false positive rate. Figure 9 shows the contribution of features to prediction when using lead I electrocardiogram. The central figure shows the direction of contribution of each feature shown on the left, with the gray to black scale indicating the value of each feature and the spread to the left and right indicating the magnitude of the contribution. The histogram on the right shows the contribution of the corresponding feature to the prediction as a percentage. Figure 10 is a graph showing the relationship between the number of features used and AUROC. The vertical axis corresponds to AUROC, and the horizontal axis corresponds to the number of features used.

[0012] This disclosure relates to a device for predicting the presence or absence of elevated blood glucose in a subject, comprising a data acquisition unit for acquiring electrocardiogram data of the subject, a blood glucose elevation prediction unit for inputting the electrocardiogram data into a machine learning model to predict the presence or absence of elevated blood glucose, and a result presentation unit for presenting the prediction results. The presence or absence of elevated blood glucose in a subject may indicate that the subject may be pre-diabetic. In other words, this disclosure relates to a device for predicting whether a subject is pre-diabetic. In addition to electrocardiogram data, attribute information including the subject's age and sex may be further input to the machine learning model. Electrocardiogram data may include, for example, waveform features such as amplitude, duration, and shape of P waves, QRS complexes, and T waves; ST segment deviation; time intervals such as QT intervals and PR intervals; and variability of R-R intervals (heart rate variability); frequency domain features such as wavelet transform coefficients and frequency components obtained by Fourier transform; statistical features such as mean, variance, skewness, and kurtosis of each waveform; autoregressive model coefficients; and nonlinear features such as entropy, fractal dimension, and Lyapunov exponent. The device relating to this disclosure may be, for example, a wearable device, particularly a wearable device capable of acquiring a single-lead electrocardiogram. Wearable devices include, for example, smartwatches, fitness trackers, and smart rings. Specific products include, for example, Apple Watch (manufactured by Apple Inc.), Withings ScanWatch (manufactured by Withings Inc.), AliveCor KardiaMobile (manufactured by AliveCor Inc.), and QardioCore (manufactured by Qardio Inc.).

[0013] In some embodiments, a machine learning model is trained to output a value indicating the possibility of elevated blood glucose. The presence of elevated blood glucose in a subject may indicate that the subject is pre-diabetic. In other words, the machine learning model is trained to output a value indicating that the subject is pre-diabetic. Training a machine learning model generally involves the following steps: 1. Data collection and preprocessing: First, data appropriate to the problem is collected. This data is used to train the model. After collecting the data, preprocessing is performed, such as removing unnecessary parts, handling missing values, and normalizing the data. Feature extraction may also be performed on the collected data. 2. Data splitting: The data is split into three datasets: a training dataset, a validation dataset, and a test dataset. The training dataset is used to train the model, and the validation dataset is used for adjusting the model parameters and hyperparameter tuning. The test dataset is used for the final evaluation of the model. 3. Model selection: An appropriate model is selected depending on the nature of the problem. For example, logistic regression, random forests, and neural networks are commonly used for classification problems. 4. Model construction: The selected model is implemented and fitted to the training dataset. 1. Model training: The model is trained on the training dataset. In this process, the model learns patterns and relationships in the data. Optimization algorithms such as gradient descent are used for training. 2. Model evaluation: The model is evaluated on the validation dataset to assess its performance. Evaluation metrics include accuracy, recall, precision, and F1 score. 3. Model tuning: The model's hyperparameters and structure are adjusted to improve its performance on the validation dataset. This is expected to enable the model to make better predictions. 4. Final evaluation: The final model is evaluated on the test dataset to verify its performance. This step confirms how well the model can generalize to unknown data.9. Deployment: Finally, the trained model is deployed to a production environment and ready to make predictions and classifications on new data.

[0014] In some embodiments, the machine learning model may be selected from the group consisting of CatBoost, decision trees, random forests, LightGBM, XGBoost, stacking (e.g., LR, RF, LGBM stacking), logistic regression, lasso regression, support vector machines, multilayer perceptrons, neural networks, and combinations thereof.

[0015] CatBoost (short for Categorical Boosting) is an implementation of the gradient boosting algorithm developed by Yandex, and is particularly known for its effective handling of categorical data. Like other gradient boosting algorithms (such as XGBoost and LightGBM), CatBoost is a method that builds a strong learner by sequentially combining multiple weak learners (usually decision trees). Because CatBoost can process categorical data directly, encoding is not required. This means that preprocessing such as target encoding or one-hot encoding is not necessary. Furthermore, CatBoost often achieves higher accuracy than other gradient boosting algorithms, which is due to its unique techniques for preventing overfitting and its ingenious method of calculating gradients. CatBoost is efficiently parallelized and optimized, allowing for fast training even on large datasets, and often exhibits high performance even with default settings, thus requiring less effort for parameter tuning. Furthermore, CatBoost offers robust handling of missing and outlier values, as well as features to enhance model interpretability (e.g., SHAP value calculation) and specialized functions for handling time-series data.

[0016] Decision trees are a method for recognizing patterns by dividing data into groups along a tree structure using conditional branching. Decision trees learn patterns and relationships in data by dividing the data based on a set of data points and simple questions and their conditions. This allows them to solve classification and regression problems. A decision tree has a tree structure, starting from a root node, with each node dividing the data based on a specific question. Each node generates new nodes (child nodes) through its branching, eventually reaching a leaf node. The leaf node outputs the final classification or prediction. Decision trees are intuitively easy to understand, making it easy to interpret the constructed model. Furthermore, the importance of each feature can be evaluated during the model construction process. Additionally, they are applicable even when the data has non-linear relationships.

[0017] Random Forest is a machine learning ensemble learning method. Ensemble learning is a method that combines multiple models to make a final prediction. Random Forest is constructed by combining multiple decision trees. Random Forest constructs multiple decision trees and combines their results to make a final prediction. Each decision tree is trained independently. In Random Forest, multiple datasets are generated by randomly sampling from the original dataset. This method is called bootstrap sampling. Each decision tree is trained using different bootstrap samples. In the splitting of each decision tree, only randomly selected features are considered. This prevents individual decision trees from overfitting and improves the generalization performance of the entire model. After all decision trees have made predictions, Random Forest makes a final prediction by taking a majority vote in the case of classification, or by taking the mean in the case of regression. Random Forest exhibits high predictive performance even with relatively little data preprocessing and is applied to various types of problems. It is also known as a highly interpretable model because it is easy to evaluate the importance of features.

[0018] LightGBM is a gradient boosting framework. Gradient boosting builds weak predictive models (usually decision trees) sequentially, learning so that each model corrects the errors of the previous one. LightGBM is known as a particularly fast and efficient implementation. LightGBM uses a histogram-based algorithm, building a gradient histogram for each feature in the dataset and employing an approximation algorithm to find efficient partitions. LightGBM can perform effective modeling even on large datasets and data with high-dimensional features.

[0019] Logistic regression is a machine learning model used for classification problems. While primarily used for classifying two classes, it can be extended to multi-class classification. Logistic regression predicts the probability that input data belongs to each class by applying a linear combination of input variables to a logistic function (or sigmoid function). The sigmoid function returns a value between 0 and 1, which can be interpreted as a probability. Logistic regression works well for linearly separable problems but is unsuitable for problems with non-linear relationships. In such cases, other methods such as support vector machines using kernel methods or neural networks are considered.

[0020] Lasso regression is a widely used technique in statistics and machine learning, and is particularly effective for data with a large number of predictor variables. Lasso stands for "Least Absolute Shrinkage and Selection Operator," and it is a type of linear regression that introduces a penalty term (regularization term) based on the sum of the absolute values ​​of the parameters. Lasso regression has the ability to exclude unnecessary predictor variables (features) from the model and select only the important variables. This is achieved by the penalty term suppressing some coefficients to zero. While ordinary linear regression can encounter problems when multiple predictor variables are highly correlated, Lasso regression helps mitigate this. Furthermore, the reduced number of variables makes the model simpler, facilitating interpretation and analysis.

[0021] Support Vector Machines (SVMs) are one of the most powerful algorithms for classification and regression in machine learning. While primarily used for classification problems, they can also be applied to regression problems. The basic idea of ​​SVMs is to find the optimal decision boundary (hyperplane) for classifying data. This hyperplane is a boundary in a multidimensional space, such as a line or plane, that best divides the data. SVMs attempt to find a decision boundary with the largest margin (distance) between different classes of data, thus improving their generalization performance. Furthermore, SVMs can be applied to nonlinear datasets using the kernel trick, which allows for mapping data into a high-dimensional space to find nonlinear boundaries. SVMs possess excellent generalization capabilities with a relatively low risk of overfitting.

[0022] A neural network is a type of machine learning model based on a mathematical model inspired by the workings of the human brain. A neural network consists of multiple layers, each layer potentially consisting of multiple neurons (nodes). These neurons receive input from the previous layer, weight each input, sum them up, and apply an activation function to the sum to generate an output. A neural network generally consists of the following main layers: 1. Input layer: The layer that receives data, with each input associated with a specific neuron. 2. Hidden layer: A layer between the input and output layers that models the complex relationships between data. When there is one or more hidden layers, these are called a "deep neural network." 3. Output layer: The final layer of the neural network that outputs the model's prediction or classification.

[0023] Neural network training typically follows these steps: 1. Data preparation: Prepare input data and their corresponding correct labels (training data). 2. Network definition: Define the structure of the neural network and select the appropriate number of layers and neurons. 3. Model training: Use the input data to adjust the network's weights and biases to bring the output closer to the correct labels. This is usually done using gradient descent or its derivative algorithms. 4. Model evaluation: Apply the trained model to test or validation data and evaluate its performance. This is done using metrics such as classification accuracy or prediction precision.

[0024] A multilayer perceptron (MLP) is a type of artificial neural network (ANN) characterized by the presence of multiple layers (hidden layers). An MLP consists of an input layer, multiple hidden layers, and an output layer. Each layer consists of multiple neurons (nodes) connected to neurons in adjacent layers. An MLP has the following structure and operation: 1. Input layer: Data given to the model is input to this layer. The number of nodes in the input layer is equal to the number of features. 2. Hidden layer: There is one or more hidden layers, each containing multiple neurons. Each neuron is connected to all neurons in the previous layer, and each connection is assigned a weight. An activation function is applied, introducing nonlinearity. Common activation functions include sigmoid, ReLU, and tanh (hyperbolic tangent). 3. Output layer: This layer generates the results that the model outputs. The number of nodes in the output layer varies depending on the type of problem. For regression problems, there is one node, and for classification problems, there are the same number of nodes as the number of classes. Common activation functions include linear functions for regression problems, sigmoid functions for binary classification problems, and softmax functions for multi-class classification problems. MLPs can model complex nonlinear relationships between inputs and outputs by learning the appropriate combination of weights. The training process typically uses an algorithm called backpropagation, which updates the weights so that the model's predictions minimize the error from the target.

[0025] Stacking machine learning models is a technique for building a more powerful predictive model by combining multiple different models. For example, LR (Logistic Regression), RF (Random Forest), and LGBM (LightGBM) are machine learning models with different algorithms, and are not limited to these, but stacking these models allows us to leverage the strengths of each model and compensate for their weaknesses. Any model, including the models disclosed herein, can be used for stacking. Stacking is generally performed in the following steps: 1. First stage: Train multiple basic models using training data. For example, different types of models that tend to capture different features, such as LR, RF, and LGBM, are used. 2. Second stage: Treat the prediction results of each basic model trained in the first stage as new features. Train another model called a metamodel based on the training data. This model takes the prediction results of the basic models as input and makes the final prediction. 3. Prediction: First, make predictions on the test data using each basic model from the first stage. Then, use the prediction results to make the final prediction with the metamodel from the second stage.

[0026] In some embodiments, the computer program relating to this disclosure may be stored on a computer-readable recording medium. Thus, in one embodiment, this disclosure relates to a computer-readable recording medium on which the computer program relating to this disclosure is recorded. Computer-readable recording media include, but are not limited to, hard disk drives (HDDs), solid-state drives (SSDs), USB flash drives, CD-ROMs, DVD-ROMs, Blu-ray® discs, memory cards, magnetic tapes, and the like.

[0027] In some embodiments, the apparatus for prediction includes means for inputting data. Examples of means for inputting data include electrocardiographs, wearable devices such as smartwatches, keyboards, and mice.

[0028] In some embodiments, the apparatus according to the present disclosure includes a central processing unit (CPU) connected to a wearable device such as an electrocardiograph or smartwatch that acquires data, a keyboard or mouse, a hard disk or flash memory as a storage unit, and a memory (storage means) such as ROM or RAM.

[0029] Means for outputting data including prediction results include, for example, monitors and printers. Alternatively, means for storing the data in storage devices such as hard disks, flash memory, ROM, and RAM can also be considered output methods.

[0030] The apparatus relating to this disclosure may include means for storing the computer program relating to this disclosure. Examples of means for storing the program include hard disks and flash memory. Such storage means may be connected by a communication line. That is, the apparatus relating to this disclosure may be part of a system obtained by connecting a device including the means for storing the program via a communication line.

[0031] Figure 1 is a schematic diagram showing an exemplary embodiment of the apparatus according to the present disclosure. In Figure 1, 100 is a computer, comprising a control unit 101, a storage unit 102, a peripheral device I / F unit 103, an input unit 104, a display unit 105, and a communication unit 106, which are connected by a bus 110. Note that this configuration is illustrative, and various configurations can be adopted as appropriate.

[0032] The control unit 101 consists of a CPU (Central Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), etc. The CPU calls programs stored in the storage unit 102, ROM, recording medium, etc., into the work memory area on the RAM and executes them, drives and controls each device connected via the bus 110, and realizes the processing that the computer performs. The ROM is a non-volatile memory that holds the boot program and BIOS of the computer 100, as well as other programs and data. The RAM is a volatile memory that temporarily holds programs and data loaded from the storage unit 102, ROM, recording medium, etc., and also has a work area used by the control unit 101 when performing various processing. The storage unit 102 is, for example, an HDD (Hard Disk Drive) and stores programs executed by the control unit 101 and various other data.

[0033] The peripheral device interface (I / F) section 103 is a port for connecting the computer 100 to peripheral devices. The peripheral device interface section 103 consists of USB, IEEE 1394, RS-232C, etc. The connection method to peripheral devices can be wired or wireless. The input section 104 has input devices such as a keyboard, mouse, pointing device, and numeric keypad, and provides operation instructions, action instructions, and data input to the computer 100. The display section 105 is a logic circuit or device driver for displaying images and videos on a display device such as an LCD panel. The input section 104 and the display section 105 can also be configured as an integrated touch display.

[0034] The communication unit 106 has a communication control device, a communication port, etc., and is a wired or wireless communication interface that mediates communication with the network 120. The bus 110 is a communication path that mediates the exchange of control signals, data signals, etc. between each device. The network 120 may be further connected to an external server 130 or a database (or network storage) 140.

[0035] This disclosure provides a device for predicting whether or not blood glucose levels will rise in a subject, the device including a data acquisition unit for acquiring electrocardiogram data of the subject, a blood glucose elevation prediction unit for inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels will rise, and a result presentation unit for presenting the prediction results.

[0036] The device provided in this disclosure may be a wearable device.

[0037] The electrocardiogram data may be, for example, 12-lead electrocardiogram data or lead I electrocardiogram data.

[0038] More specifically, a machine learning model is a pre-trained machine learning model.

[0039] More specifically, a machine learning model is a pre-trained machine learning model obtained through supervised learning.

[0040] The machine learning model is preferably CatBoost.

[0041] Preferably, during training of a pre-trained machine learning model, data including pairs of the subject's electrocardiogram data and data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing diabetes treatment is input to the machine learning model.

[0042] Here, the number of subjects is, for example, 10 or more, 100 or more, 1,000 or more, 2,000 or more, 3,000 or more, 4,000 or more, 5,000 or more, 6,000 or more, 7,000 or more, 8,000 or more, 9,000 or more, 10,000 or more, 11,000 or more, 12,000 or more, 13,000 or more, 14,000 or more, 15,000 or more, 16,000 or more, 17,000 or more, 18,000 or more, 19,000 or more, or 20,000 or more.

[0043] Preferably, the electrocardiogram data of the subject and the data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing diabetes treatment are data obtained from measurements taken on the same day, or data for the subject on the same day.

[0044] Preferably, the subject's electrocardiogram data is feature data extracted from the electrocardiogram. Here, the feature extraction in the feature data extracted from the electrocardiogram is performed, for example, using software. Here, the software is, for example, an electrocardiogram feature extraction program. In the feature data extracted from the electrocardiogram, the number of features is 1 or more, 2 or more, 3 or more, 4 or more, 5 or more, 6 or more, 7 or more, 8 or more, 9 or more, 10 or more, 20 or more, 30 or more, 40 or more, 50 or more, 60 or more, 70 or more, 80 or more, 90 or more, 100 or more, 110 or more, 120 or more, 130 or more, 140 or more, 150 or more, 160 or more, 170 or more, 180 or more, 1 90 or higher, 200 or higher, 210 or higher, 220 or higher, 230 or higher, 240 or higher, 250 or higher, 260 or higher, 270 or higher, 280 or higher, 290 or higher, 300 or higher, 310 or higher, 320 or higher, 330 or higher, 340 or higher, 350 or higher, 360 or higher, 370 or higher, 380 or higher, 390 or higher, 400 or higher, 410 or higher, 420 or higher, 430 or higher, 440 or higher, 450 or higher, 460 or higher , 470 or more, 480 or more, 490 or more, 500 or more, 510 or more, 520 or more, 530 or more, 540 or more, 550 or more, 560 or more, 570 or more, 580 or more, 590 or more, 600 or more, 610 or more, 620 or more, 630 or more, 640 or more, 650 or more, 660 or more, 670 or more, 680 or more, 690 or more, 700 or more, 710 or more, 720 or more, 730 or more, 740 The above are 750 or more, 760 or more, 770 or more, 780 or more, 790 or more, 800 or more, 810 or more, 820 or more, 830 or more, 840 or more, 850 or more, 860 or more, 870 or more, 880 or more, 890 or more, 900 or more, 910 or more, 920 or more, 930 or more, 940 or more, 950 or more, 960 or more, 970 or more, 980 or more, 990 or more, or 1,000 or more.In the data of the feature quantities extracted from the electrocardiogram, the number of feature quantities is 1,000 or less, 990 or less, 980 or less, 970 or less, 960 or less, 950 or less, 940 or less, 930 or less, 920 or less, 910 or less, 900 or less, 890 or less, 880 or less, 870 or less, 860 or less, 850 or less, 840 or less, 830 or less, 820 or less, 810 or less, 800 or less, 790 or less, 780 or less, 770 or less, 760 or less, 750 or less, 740 or less, 730 or less, 720 or less, 710 or less, 700 or less, 690 or less, 680 or less, 670 or less, 660 or less, 650 or less, 640 or less, 630 or less, 620 or less, 610 or less, 600 or less, 590 or less, 580 or less, 570 or less, 560 or less, 550 or less, 540 or less, 530 or less, 520 or less, 510 or less, 500 or less, 490 or less, 480 or less, 470 or less, 460 or less, 450 or less, 440 or less, 430 or less, 420 or less, 410 or less, 400 or less, 390 or less, 380 or less, 370 or less, 360 or less, 350 or less, 340 or less, 330 or less, 320 or less, 310 or less, 300 or less, 290 or less, 280 or less, 270 or less, 260 or less, 250 or less, 240 or less, 230 or less, 220 or less, 210 or less, 200 or less, 190 or less, 180 or less, 170 or less, 160 or less, 150 or less, 140 or less, 130 or less, 120 or less, 110 or less, 100 or less, 90 or less, 80 or less, 70 or less, 60 or less, 50 or less, 40 or less, 30 or less, 20 or less, 10 or less, 9 or less, 8 or less, 7 or less, 6 or less, 5 or less, 4 or less, 3 or less, 2 or less, or 1 or less.

[0045] [[ID=​​​​​In the data of the feature quantities extracted from the electrocardiogram, the feature quantities preferably do not include a feature quantity in which a predetermined ratio or more of the total number of data including, in pairs, the electrocardiogram data of the subject and the data indicating the fasting blood glucose, HbA1c, or diabetes treatment of the subject is a missing value. Here, examples of the predetermined ratio include a ratio belonging to the range of 1% or more, 2% or more, 3% or more, 4% or more, 5% or more, 6% or more, 7% or more, 8% or more, 9% or more, 10% or more, 20% or more, 30% or more, 40% or more, 50% or more, 60% or more, 70% or more, 80% or more, or 90% or more. Also, examples of the predetermined ratio include a ratio belonging to the range of 90% or less, 80% or less, 70% or less, 60% or less, 50% or less, 40% or less, 30% or less, 20% or less, 10% or less, 9% or less, 8% or less, 7% or less, 6% or less, 5% or less, 4% or less, 3% or less, 2% or less, or 1% or less.

[0048] In the data of the feature quantities extracted from the electrocardiogram, the feature quantities are preferably those obtained by excluding a feature quantity in which a predetermined ratio or more of the total number of data including, in pairs, the electrocardiogram data of the subject and the data indicating the fasting blood glucose, HbA1c, or diabetes treatment of the subject is a missing value. Here, examples of the predetermined ratio include a ratio belonging to the range of 1% or more, 2% or more, 3% or more, 4% or more, 5% or more, 6% or more, 7% or more, 8% or more, 9% or more, 10% or more, 20% or more, 30% or more, 40% or more, 50% or more, 60% or more, 70% or more, 80% or more, or 90% or more. Also, examples of the predetermined ratio include a ratio belonging to the range of 90% or less, 80% or less, 70% or less, 60% or less, 50% or less, 40% or less, 30% or less, 20% or less, 10% or less, 9% or less, 8% or less, 7% or less, 6% or less, 5% or less, 4% or less, 3% or less, 2% or less, or 1% or less.

[0049] The data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing diabetes treatment preferably includes data indicating a classification based on whether the fasting blood glucose or HbA1c is above a predetermined value. For fasting blood glucose, a predetermined value could be, for example, 110 mg / dL. For HbA1c, a predetermined value could be, for example, 6.0%. In this disclosure, elevated blood glucose may be defined as including fasting blood glucose ≥ 110 mg / dL or HbA1c ≥ 6.0%.

[0050] In training a pre-trained machine learning model, data containing pairs of a subject's electrocardiogram data and data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing diabetes treatment is preferably divided into training data and test data in a predetermined ratio. Examples of predetermined ratios include 100:1 to 2:1, 50:1 to 3:1, 30:1 to 4:1, 20:1 to 5:1, 15:1 to 6:1, and 10:1 to 7:1.

[0051] Preferably, during training on a pre-trained machine learning model, the parameters of the machine learning model are set by an optimization tool.

[0052] Preferably, during training of a pre-trained machine learning model, a predetermined number of training sessions are performed using a predetermined number of patterns obtained by further dividing the training data into training data and validation data. The predetermined number is, for example, 1 or more, 2 or more, 3 or more, 4 or more, 5 or more, 6 or more, 7 or more, 8 or more, 9 or more, or 10 or more, for example, 100 or less, 90 or less, 80 or less, 70 or less, 60 or less, 50 or less, 40 or less, 30 or less, 20 or less, or 10 or less.

[0053] Preferably, during training of a pre-trained machine learning model, the training parameters are set to impose a larger penalty for prediction errors in minority classifications.

[0054] Preferably, in training a pre-trained machine learning model, training is performed after optimizing the parameter settings.

[0055] Preferably, in the training performed after the optimization of parameter settings, a predetermined number of training sessions are performed using a predetermined number of patterns obtained by further dividing the training data into training data and validation data. The predetermined number is, for example, 1 or more, 2 or more, 3 or more, 4 or more, 5 or more, 6 or more, 7 or more, 8 or more, 9 or more, or 10 or more, for example, 100 or less, 90 or less, 80 or less, 70 or less, 60 or less, 50 or less, 40 or less, 30 or less, 20 or less, or 10 or less.

[0056] Preferably, during training of a pre-trained machine learning model, data including a set of the subject's electrocardiogram data, data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing diabetes treatment, and data obtained from the subject's health checkup is input to the machine learning model. The data obtained from the subject's health checkup is, for example, data showing one or more subject information selected from a group consisting of the subject's age, sex, height, weight, BMI, waist circumference, smoking status, and frequency of alcohol consumption. Here, the data showing one or more subject information is, for example, data showing subject information 1, 2, 3, 4, 5, 6, 7, or 8.

[0057] Preferably, the subject's electrocardiogram data is data generated from data recorded by an electrocardiograph.

[0058] Preferably, the subject's electrocardiogram data is image data of the electrocardiogram waveform as raw data, numerical data or image data obtained by processing to convert to an electrocardiogram waveform, raw electrocardiogram potential data, electrocardiogram potential data of 12 sampled leads, feature data extracted from electrocardiogram potential data, or data obtained by vectorizing electrocardiogram potential data using a deep learning model.

[0059] This disclosure provides a method for supporting the early detection of prediabetes, comprising the steps of acquiring electrocardiogram data of the subject, inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated, and presenting the prediction results, wherein a prediction of elevated blood glucose levels serves as an indicator that the person is prediabetic.

[0060] The method disclosed herein may further include a step of confirming the measured value of the target blood glucose level.

[0061] This disclosure provides a program for predicting whether or not blood glucose levels are elevated in a subject, which causes one or more processors to perform the steps of: acquiring electrocardiogram data of the subject; inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated; and presenting the prediction results.

[0062] This disclosure provides a non-temporary computer-readable recording medium storing the above-mentioned program.

[0063] This disclosure provides a pre-trained machine learning model used in a device for predicting whether or not blood glucose levels are elevated in a subject, the model being generated by training with training data, with the subject's electrocardiogram data as the explanatory variable and the subject's fasting blood glucose, HbA1c, or data indicating that they are undergoing diabetes treatment as the dependent variable.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by those skilled in the art to which the invention pertains. Any methods and materials similar or equivalent to those described herein may be used for carrying out or testing the invention, but several possible and preferred methods and materials are described herein. All publications referenced herein are incorporated herein by reference, and the methods and / or materials cited in relation to these publications are disclosed and described herein. In the event of any conflict, this disclosure shall prevail over the disclosure of the incorporated publication.

[0065] Where a range of values ​​is given, unless the context clearly indicates otherwise, each intermediate value between the upper and lower limits of that range, up to one-tenth of the lower limit unit, is also specifically disclosed. Each smaller range between any given value or intermediate value within the given range and any other given value or intermediate value within that given range is also included in this disclosure. The upper and lower limits of these smaller ranges may be independently included in or excluded from that range, and each range that includes either, either, or both of the limit values ​​in the smaller range is also included in the invention, but the limit values ​​specifically excluded in the given range are reserved. Where a given range includes one or both of the limit values, a range that excludes either or both of the included limit values ​​is also included in the invention. The term “about” with respect to numerical values ​​means within 5%.

[0066] The embodiments described herein are intended to be illustrative only, and those skilled in the art will be able to make numerous modifications and alterations without departing from the spirit of the invention. Certain modifications and alterations may yield satisfactory results, though not optimal. All such modifications and alterations are intended to fall within the scope of the invention as defined by the appended claims. Furthermore, any combination of the components disclosed herein, and any conversion of the expressions of this disclosure between methods, apparatus, systems, computer programs, data structures, recording media, etc., are also valid embodiments of this disclosure. Thus, the details described relating to the methods of this disclosure may be applied to systems, computer programs, data structures, recording media, etc.

[0067] This disclosure will be further understood by reference to the following embodiments. These embodiments are provided solely to illustrate the disclosure set forth in the claims, and the scope of this disclosure is not limited by the embodiments shown, which are intended only as examples of a single aspect of this disclosure. Any functionally equivalent method is included within the scope of this disclosure. In addition to those described herein, various modifications of this disclosure will be apparent to those skilled in the art from the foregoing description. Such modifications are intended to be within the scope of the appended claims.

[0068] Example 1: With changes in lifestyle and social environment, the number of diabetic patients is rapidly increasing, and detecting diabetes before it develops is extremely important in order to suppress the onset of complications. For example, diabetes is an independent risk factor for heart failure, and it has been reported that the risk of developing heart failure is already significantly higher at the prediabetic stage, that is, the stage before diabetes develops. This suggests that even at the stage of elevated blood glucose before diabetes develops, there is some effect on the heart, and it suggests that this effect may be reflected in the electrocardiogram. If even minor electrocardiogram changes that clinicians cannot recognize can be detected by machine learning, the electrocardiogram could become a means of detecting "elevated blood glucose (= early "pre-diabetic")," and the following investigation was conducted.

[0069] Subjects and Methods: The study included individuals who underwent electrocardiogram (ECG) testing during health checkups at Clinic A in 2022 and whose fasting blood glucose or HbA1c was measured on the same day. "Elevated blood glucose (= early 'pre-diabetic')" was defined as meeting any of the following criteria: fasting blood glucose ≥ 110 mg / dL, HbA1c ≥ 6.0%, or being under diabetes treatment. Features were extracted from the ECGs, and supervised learning was performed using multiple machine learning models, as described below, to examine the predictive performance of each classification. There were 16,766 ECGs available for analysis.

[0070] 1. Feature Generation Each electrocardiogram sample was obtained as a 5,000-row x 12-column CSV file containing 12-lead potential data recorded at 500 Hz for 10 seconds. This potential data was processed using ECG featurizer (an electrocardiogram feature extraction program), a freely available Python package, to obtain 384 feature quantities for each electrocardiogram. Each feature quantity is calculated for each of the 12 leads, but since the five items related to heart rate (R_HR_STD, R_HR_max, R_HR_median, R_HR_min, mean_heartrate_r) should be uniquely determined for each test, only the feature quantities for lead I were adopted, and the data for the other leads were excluded. This process resulted in the exclusion of 55 features. Furthermore, features that could not be calculated for any reason were recorded as missing values. Since a large number of missing values ​​can distort the analysis results, 60 features that were missing in more than 10% of the samples were excluded. A total of 115 features were excluded, resulting in 269 ECG features being used (Table 1). These excluded features can be included in the model, and their addition may contribute to improved predictive performance. There are 41 other features that can be extracted by the ECG featurizer, as shown in Table 2. However, ECG_baseline is the baseline potential and is not thought to show pathological findings, and the other 40 items have output values ​​of heart rate, so their significance was considered the same as the five heart rate-related items mentioned above, and therefore they were excluded this time. However, depending on the circumstances, adding these features may be an option to improve predictive performance. Furthermore, there are items shown in Table 3 that are not included in the standard features extracted by the ECG featurizer but are technically extractable by changing the algorithm, and there is room to add these.

[0071]

[0072] 2. 16,766 health checkup data sets, each containing training electrocardiogram data and either fasting blood glucose or HbA1c as a pair, were split into a training:test ratio of 9:1. The training data was used to train the model to distinguish between "elevated blood glucose (= early "pre-diabetes")" and "normal." The prediction performance of six models (logistic regression, random forest, LightGBM, XGBoost, CatBoost, and MLP) was compared (Table 4, Figures 2 and 3), and CatBoost, which showed the best prediction performance, was adopted. In comparing the performance of the models, AUROC, which is independent of the binary classification threshold, was used as an indicator. To explore the optimal parameter tuning for the model, 1,000 trials were performed using the optimization tool Optuna, and the parameter setting with the lowest prediction error was adopted. During training, the training data was further split into a training:validation ratio of 9:1, and 10 patterns were created and 10 training runs were performed using 10-fold cross-validation to improve prediction performance. Furthermore, only about 10% of the total data was classified as "elevated blood glucose," indicating a significant imbalance between normal and abnormal data. There was a concern that the system would learn primarily from the overwhelming majority, "normal," and classify everything as "normal." Therefore, a class weight was set as a training parameter to impose a greater penalty for misclassification of minority categories, aiming to correct this imbalance.

[0073] 3. The CatBoost model was trained again with the predictively optimized parameter settings. Ten-fold cross-validation was performed again to obtain ten trained models. Using these ten models, the test data, which had been initially split, was classified into "normal" and "elevated blood glucose," and the average of the ten predicted values ​​was output as the final predicted value. Since the predicted value is output as a decimal value between 0 and 1, it is necessary to set a boundary value for binary classification. This was determined using the Youden index, which is the boundary value where the difference between the true positive rate and the false positive rate is maximized by drawing a receiver operating characteristic curve (ROC curve) using the predicted values ​​from the training data. If the predicted value is lower than this boundary value, it was classified as "normal"; if it is higher than or equal to the boundary value, it was classified as "elevated blood glucose."

[0074] Results The CatBoost model showed the highest predictive performance for "elevated blood glucose," with an AUROC of 0.858, sensitivity of 86.4%, and specificity of 67.7% when using only a 12-lead electrocardiogram (ECG) (Figure 2). When analyzing using only a lead I ECG, with the use of smartwatches in mind, the AUROC was 0.841, sensitivity of 79.6%, and specificity of 72.3%. While the predictive performance was lower than that of the 12-lead ECG, it was still at a sufficient level (Figure 4, Table 5). Based on the prediction results from the 12-lead ECG, SHAP values ​​indicating the contribution of each feature to the prediction were calculated. Features that showed particularly large contributions included features suggesting signs of left ventricular hypertrophy, such as increased R wave amplitude in lead aVL, increased R wave amplitude in lead I, and a deep S wave in lead V4, as well as features such as a decrease in RR interval variability (Figure 5). These abnormalities are known as characteristic cardiac function changes in diabetic patients. However, among subjects who were judged to have elevated blood glucose, only 11 out of 621 positive cases (1.8%) and 4 out of 127 true positive cases (3.2%) were also judged to have left ventricular hypertrophy (including suspected cases) by automated analysis of the electrocardiograph. Furthermore, when eight subject information items (age, sex, height, weight, BMI, waist circumference, smoking status, and frequency of alcohol consumption) were added to the electrocardiogram of lead I only, an improvement in predictive performance was observed, with an AUROC of 0.881, sensitivity of 85.7%, and specificity of 73.8% (Figure 8, Table 6).

[0075] Figure 9 shows the SHAP values ​​calculated based on the prediction results from lead I, indicating the contribution of each feature to the prediction. In terms of feature contribution, the height of the R wave and the standard deviation of the RR interval, which also appeared in the 12-lead analysis, are among the most important. Since a small variation in the RR interval is associated with autonomic neuropathy, it is reasonable that a smaller variation leads to a higher prediction of prediabetes / diabetes.

[0076] Figure 10 is a graph showing the relationship between the number of features used in I-guided prediction and AUROC. When the number of features is 15, the AUROC is 0.842, suggesting that around 15 features are necessary to improve prediction accuracy.

[0077] Discussion We have successfully established a model that predicts "elevated blood glucose" using only a 12-lead electrocardiogram (ECG). Very recently, a report (eClinicalMedicine. 2024 Feb 1:68:102445.) from overseas has been published that predicts diabetes (fasting blood glucose ≥ 126 mg / dL or HbA1c ≥ 6.5%) using deep learning with ECGs, but the predictive performance using only a 12-lead ECG was AUROC 0.787, sensitivity 82.6%, and specificity 61.1%. The model created using machine learning with ECG features in this study is superior not only in its predictive performance but also in its ability to predict "elevated blood glucose (= early "pre-diabetes")" from the stage of "elevated blood glucose (= early "pre-diabetes")" (fasting blood glucose ≥ 110 mg / dL, HbA1c ≥ 6.0%). Furthermore, we have also succeeded in predicting "elevated blood glucose (= early "pre-diabetes")" with high performance using only a lead I ECG. Considering the contribution of each feature to prediction based on SHAP values, it was found that signs of left ventricular hypertrophy are treated as an important finding. However, only a small number of subjects exhibited significant left ventricular hypertrophy to the extent that it would be judged as (suspected) by automated analysis using an electrocardiograph. It is thought that the model makes its judgment by comprehensively considering mild signs of left ventricular hypertrophy associated with elevated blood glucose, as well as other signs of cardiac dysfunction. This is a task that is technically very difficult for physicians to perform. Predictive performance was improved by adding eight items of subject information to the I-lead electrocardiogram analysis. In the future, it is conceivable to use a combination of various subject data obtained from health checkups (including medical history, family history, physical findings, and blood test results) with the electrocardiogram.

[0078] 12-lead electrocardiograms are widely used as a basic examination in all medical departments, not just general health checkups, and are expected to be useful in detecting diabetes in many people at an early stage. Furthermore, if it were possible to incorporate lead I electrocardiograms into smartwatches, it would be possible to detect elevated blood glucose levels (i.e., early stages of prediabetes) at home without performing blood tests.

[0079] Furthermore, when examining the changes in predictive performance when the HbA1c threshold was altered, an improvement in predictive performance was observed at HbA1c levels of 5.4% to 6.0%, indicating that elevated blood glucose levels were reflected in the electrocardiogram even before a diagnosis of diabetes (HbA1c 6.5% or higher) was made (Figure 6). This is considered one basis for the idea that intervention is necessary from the "elevated blood glucose" stage.

[0080] There are several criteria for prediabetes, and to verify at what blood glucose and HbA1c levels prediction is it possible to predict diabetes, Figure 7 shows the results of examining the predictive performance by changing the classification criteria. The diagnostic criteria for diabetes are a fasting blood glucose (FPG) of 126 and an HbA1c of 6.5. At this point, the AUROC is 0.882, which is higher than the graph in Figure 6, indicating that the higher the blood glucose, the more pronounced the changes in the electrocardiogram become.

[0081] Furthermore, it was interesting to note that the predictive performance did not reach 100%, and there were a certain number of health checkup participants whose diagnoses based on blood glucose and HbA1c differed from those based on electrocardiograms. It is generally known that diabetic patients have a high risk of heart failure. On the other hand, there are individual differences (i.e., some people develop heart failure even with the same HbA1c level, while others do not), but in actual clinical practice, these individual differences only become apparent retrospectively after heart failure has occurred. The electrocardiogram-based diagnosis developed in this study is based on the changes occurring in each individual's myocardium when blood glucose levels are rising, so it is thought that it may be possible to predict the onset of heart failure and identify individuals who truly need preventive measures.

[0082] In this example, prediction was performed using features extracted by ECG Featureizer as input. However, other possible input data formats include raw electrocardiogram waveform images, raw potential data, 12-lead sampling potential data as obtained in the above example, numerical and image data obtained by processing waveforms with Fourier transforms, features extracted from electrocardiogram potential data using a proprietary algorithm, and electrocardiogram potential data vectorized using a deep learning model. In short, any data that can be generated from data recorded by an electrocardiograph can be used as input data. Furthermore, analysis using a combination of this electrocardiogram-related data and subject data obtained from health checkups (including physical examination and blood tests) is also included in the detection of elevated blood glucose using electrocardiograms.

[0083] 100...Computer 101...Control Unit 102...Storage Unit 103...Peripheral Device I / F Unit 104...Input Unit 105...Display Unit 106...Communication Unit 110...Bus 120...Network 130...External Server 140...Database

Claims

1. A device for predicting the presence or absence of elevated blood glucose in a subject, comprising: a data acquisition unit for acquiring electrocardiogram data of the subject; a blood glucose elevation prediction unit for inputting the electrocardiogram data into a machine learning model to predict the presence or absence of elevated blood glucose; and a result presentation unit for presenting the prediction results, wherein elevated blood glucose is defined as including fasting blood glucose ≥ 110 mg / dL or HbA1c ≥ 6.0%, the presence of elevated blood glucose indicates that the subject may be pre-diabetic, and the features input into the machine learning model include at least features that suggest signs of left ventricular hypertrophy or the standard deviation of the RR interval.

2. The apparatus according to claim 1, which is a device for predicting whether or not a subject is at risk of developing diabetes, in order to predict whether or not a subject is at risk of developing diabetes.

3. The apparatus according to claim 1, wherein the feature quantity suggesting signs of left ventricular hypertrophy includes a feature quantity selected from the group consisting of the R wave amplitude of lead aVL, the R wave amplitude of lead I, and the S wave amplitude of lead V4.

4. The apparatus according to claim 1, which is a wearable device.

5. The apparatus according to claim 1, wherein the electrocardiogram data is 12-lead electrocardiogram data or I-lead electrocardiogram data.

6. The apparatus according to claim 1, wherein the machine learning model is a pre-trained machine learning model.

7. The apparatus according to claim 1, wherein the machine learning model is a pre-trained machine learning model obtained through supervised learning.

8. The apparatus according to claim 1, wherein the machine learning model is CatBoost.

9. The apparatus according to claim 6, wherein, during training of a pre-trained machine learning model, data including pairs of the subject's electrocardiogram data and data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing diabetes treatment is input to the machine learning model.

10. The apparatus according to claim 9, wherein the number of subjects is 10,000 or more.

11. The apparatus according to claim 9, wherein the electrocardiogram data of the subject and the data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing diabetes treatment are data obtained from measurements taken on the same day, or data relating to the subject on the same day.

12. The apparatus according to claim 9, wherein the electrocardiogram data of the subject is data of features extracted from the electrocardiogram.

13. The apparatus according to claim 12, wherein feature extraction is performed using software in the feature data extracted from an electrocardiogram.

14. The apparatus according to claim 13, wherein the software is an electrocardiogram feature extraction program.

15. The apparatus according to claim 12, wherein the number of features in the feature data extracted from the electrocardiogram is 100 or more and 500 or less.

16. The apparatus according to claim 12, wherein the feature data extracted from the electrocardiogram is calculated for each of the 12 leads.

17. The apparatus according to claim 12, wherein, in the feature data extracted from the electrocardiogram, the features include only the features of lead I for items related to heart rate.

18. The apparatus according to claim 12, wherein the feature data extracted from the electrocardiogram does not include items related to heart rate.

19. The apparatus according to claim 12, wherein the feature data extracted from the electrocardiogram does not include any features in which a predetermined proportion or more of the total number of data containing pairs of the subject's electrocardiogram data and the subject's fasting blood glucose, HbA1c, or data indicating that the subject is undergoing treatment for diabetes are missing values.

20. The apparatus according to claim 12, wherein the feature data extracted from the electrocardiogram excludes features in which a predetermined proportion or more of the total number of data containing pairs of the subject's electrocardiogram data and data indicating the subject's fasting blood glucose, HbA1c, or diabetes treatment are missing values.

21. The apparatus according to claim 9, wherein the data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing diabetes treatment includes data indicating a classification based on whether the fasting blood glucose or HbA1c is above a predetermined value.

22. The apparatus according to claim 9, wherein, in training a pre-trained machine learning model, data including pairs of electrocardiogram data of a subject and data indicating the subject's fasting blood glucose, HbA1c, or whether they are undergoing treatment for diabetes is divided into training data and test data in a ratio of 10:1 to 7:

1.

23. The apparatus according to claim 6, wherein, during training of a pre-trained machine learning model, the parameters of the machine learning model are set by an optimization tool.

24. The apparatus according to claim 6, wherein, in training a pre-trained machine learning model, a predetermined number of training sessions are performed using a predetermined number of patterns obtained by further dividing the training data into training data and validation data.

25. The apparatus according to claim 24, wherein the predetermined number is 5 or more.

26. The apparatus according to claim 6, wherein, in training a pre-trained machine learning model, a setting is made to impose a larger penalty for prediction errors in minority classifications as a training parameter.

27. The apparatus according to claim 6, wherein training is performed in a pre-trained machine learning model after optimization of parameter settings.

28. The apparatus according to claim 27, wherein, in training performed after optimization of parameter settings, a predetermined number of training sessions are performed using a predetermined number of patterns obtained by further dividing the training data into training data and validation data.

29. The apparatus according to claim 6, wherein, during training of a pre-trained machine learning model, data including a set of the subject's electrocardiogram data, the subject's fasting blood glucose, HbA1c, or data indicating that the subject is undergoing treatment for diabetes, and data obtained from the subject's health checkup, is input to the machine learning model.

30. The apparatus according to claim 29, wherein the data obtained from the health examination of the subject is data representing one or more subject information selected from groups consisting of the subject's age, sex, height, weight, BMI, waist circumference, smoking status, and frequency of alcohol consumption.

31. The apparatus according to claim 30, wherein the data indicating one or more subject information is data indicating four or more subject information.

32. The apparatus according to claim 9, wherein the electrocardiogram data of the subject is data generated from data recorded by an electrocardiograph.

33. The apparatus according to claim 9, wherein the electrocardiogram data of the subject is image data of an electrocardiogram waveform as raw data, numerical data or image data obtained by processing to convert to an electrocardiogram waveform, raw electrocardiogram potential data, electrocardiogram potential data of 12 sampled leads, feature data extracted from electrocardiogram potential data, or data obtained by vectorizing electrocardiogram potential data using a deep learning model.

34. A method for supporting the early detection of prediabetes, comprising the steps of acquiring electrocardiogram data of a target, inputting the electrocardiogram data into a machine learning model to predict whether or not blood glucose levels are elevated, and presenting the prediction results, wherein a prediction of elevated blood glucose levels serves as an indicator that the person is prediabetic.

35. The method according to claim 34, further comprising the step of confirming the measured value of the target blood glucose level.

36. A program for predicting the presence or absence of elevated blood glucose in a subject in order to support the early detection of prediabetes, the program causing one or more processors to perform the steps of acquiring electrocardiogram data of the subject, inputting the electrocardiogram data into a machine learning model to predict the presence or absence of elevated blood glucose, and presenting the prediction results.

37. A non-temporary computer-readable recording medium storing the program described in claim 36.

38. A trained machine learning model used in a device for predicting whether or not blood glucose levels are elevated in a subject, wherein the trained machine learning model is generated by training with training data, with the subject's electrocardiogram data as the explanatory variable and the subject's fasting blood glucose, HbA1c, or data indicating that they are undergoing diabetes treatment as the dependent variable.

Images