The global prevalence of diabetes in 2021 among individuals aged between 20 and 79 years was estimated to be 10.5% (536.6 million people), with this figure expected to rise to 12.2% (783.2 million) by 2045. Global diabetes-related health expenditures were estimated at US $966 billion in 2021 and are projected to reach US $1054 billion by 2045 [1]. In 2020, approximately 38.1 million adults aged ≥18 years, or 14.7% of all adults in the United States, had diabetes [2]. The global prevalence of hyperlipidemia in adults is 39% (37% in men and 40% in women), and as of 2020, approximately 10% (about 25 million people) of adults aged ≥20 years in the United States had total cholesterol levels of ≥240 mg/dL and 33% (about 86 million people) had total cholesterol levels of ≥200 mg/dL [3]. Several primary hyperlipidemias (eg, familial combined hyperlipidemia, familial hypertriglyceridemia, and abnormal β lipoproteinemia) have been reported to be associated with an increased risk of type 2 diabetes [4].

Wearable devices are becoming increasingly capable of measuring a range of physiological data. Real-world data, which collect and analyze activity and physiological measurements from participants in clinical studies, can provide more sensitive measures of disease activity than traditional scales, thereby enabling faster and more objective readings in clinical trials [5]. Wearable devices can continuously collect measures of multiple physiological functions in real-time. Depending on the size and complexity of the raw data obtained from the device, data preprocessing, feature extraction, and selection are performed through data processing, such as data mining, and are applied toward abnormality detection, prediction, diagnosis, and decision-making [6]. Conversely, data repositories have advanced in recent years, making large, open databases available for wearable devices. Data mining techniques have advanced significantly in the last few years, and with the availability of these open data, opportunities emerged to devise algorithms suitable for wearable sensors [6].

Table S1 in Multimedia Appendix 1 depicts studies that applied existing open datasets. Several studies were conducted to establish algorithms and prediction models based on machine learning from sensor data [7-24].

Open datasets are available on the following websites: (1) IEEE Data Port, a research data platform designed to store research data and provide global access to research data across various fields [25]; (2) the Open Wearables Initiative, which aims to promote the effective use of high-quality sensor-generated health measurements in clinical research by openly sharing algorithms and datasets [26]; (3) PhysioNet, a searchable database containing a collection of cardiopulmonary, neurological, and other biomedical signals from healthy individuals and patients with several serious health conditions, including congestive heart failure, epilepsy, gait disturbance, and sleep apnea [27].

The BIG IDEAs Lab Glycemic Variability and Wearable Device Data (version 1.1.1) by Cho et al [28,29], a wearable database containing blood glucose–related indicators from PhysioNet, was selected for the study. Two previous studies have used the same database [30,31]. The first study demonstrated the feasibility of predicting blood glucose changes by continuously detecting individualized blood glucose deviations and determining the contribution of each variable to interstitial glucose prediction. The LOPOCV random forest regression model was used to examine the importance of the characteristics, resulting in the extraction of diet, circadian rhythm, stress, activity, body temperature, heart rate, electrodermal activity, biological sex, and HbA_1c [30]. The second study evaluated methods for detecting prediabetes and estimating glycated hemoglobin (HbA_1c) and glucose variability using digital biomarkers from wearables [31]. The relationships between features extracted from wearables and blood glucose variability and HbA_1c were investigated, and the results showed that glucose variability indices and HbA_1c could be estimated with high accuracy. The HbA_1c estimation model developed from a noninvasive wrist-worn wearable was as accurate as the invasive continuous glucose monitor (CGM)–based estimated A_1c (as recommended by the American Diabetes Association). Notably, all the sensors used in this study (triaxial accelerometer-derived acceleration [ACC], heart rate [HR], electrodermal activity [EDA], and skin temperature [TEMP]) were important for estimating glucose variability indices and HbA_1c, although EDA and TEMP were the most important indicators when estimating HbA_1c [31]. Similar to the other studies based on existing open datasets (Table S1 in Multimedia Appendix 1), these 2 studies were conducted to establish a prediction model for blood glucose-related indicators using machine learning and used the random forest regression model for analysis. This analysis can be difficult to interpret in a specific clinical context.

Several factors influence blood glucose levels, including diet [32], physical activity, exercise [33], stress [34], circadian rhythm [35], and HR [36]. For example, regarding diet, following the dietary approaches to stop hypertension (DASH) diets [37], low carbohydrate diets [38,39], and high consumption of phytochemicals and polyphenols [40,41] prevent type 2 diabetes. However, although some factors associated with blood glucose variability, such as HR [36], body temperature [42], and autonomic functions [43], including sweating motor response [44], have been reported [30], these factors have not been evaluated in real-world setting. Therefore, we conducted an exploratory study of the association between blood glucose and each physiological and nutritional index, using existing data from PhysioNet as well as simple analytical methods, such as correlation and multiple regression analyses. These analysis methods were used for their simplicity compared to the random forest model and may increase the explanatory potential as the contribution of each variable is clarified.

A database search was conducted using the word “wearable” from PhysioNet, which yielded 11 hits. Among these, we focused on wristband-type wearable devices and searched the relevant databases because the major market share is dominated by wristband- and watch-type devices [45]. The search resulted in 7 (64%) of 11 studies with wristband-type wearable devices (Empatica [Empatica Inc]: n=5, 45%; Apple watch [Apple Inc]: n=2, 18%; Fitbit [Fitbit, Inc]: n=2, 18%; Garmin [Garmin Ltd]: n=1, 9%; Samsung Galaxy watch [Samsung, Inc]: n=1, 9%; Xiaomi [Xiaomi Corp]: n=1, 9%; and Biovotion Everion [Biofourmis Inc]: n=1, 9% study; Table S2 in Multimedia Appendix 1).

We selected 1 study that investigated Empatica E4, the most commonly used wristband-type wearable device. The dataset, BIG IDEAs Lab Glycemic Variability and Wearable Device Data (version 1.1.1) by Cho et al [28,29] was selected to examine the correlation between each physiological index and blood glucose as well as conduct multiple regression analysis and 1-way ANOVA of the slope in postprandial glucose over time with each physiological index and nutrient value. The main eligibility criteria in the study included men and postmenopausal women aged 35-65 years, with point-of-care HbA_1c measurements between 5.2% and 6.4%.

A total of 16 patients (n=7, 44% men and n=9, 56% women) with HbA_1c in the high normal and prediabetic range (5.3%-6.4%, mean 5.73%, SD 0.28%) were included and monitored for 8-10 days using the Dexcom G6 CGM and Empatica E4 wrist-worn wearable-type device [28].

The demographic characteristics of the participants are listed in Table 1.

Notably, all data were time-shifted (by date) to prevent reidentification; the Dexcom G6 measured interstitial glucose concentration (mg/dL) every 5 minutes using a CGM, and the Empatica E4 measured photoelectric volumetric pulse wave, electrical activity: EDA, TEMP, and ACC, for 7 functions. The photoelectric volumetric pulse wave was sampled at 64 Hz, and HR and blood volume pulse (BVP) signals were obtained every second, from which the interbeat interval (IBI) data were calculated. Of these, EDA is known as a psychological factor and a measure of sympathetic activation related to stress [46,47]. In addition, HR variability, a related index of HR and IBI, was used as a psychological stress indicator [47,48]. EDA and TEMP were sampled at 4 Hz, whereas accelerometry was sampled at 32 Hz. For ACC, triaxial data were calculated using the Euclidean norm as a measure of average motion in the 3 axes using the following formula [49]:

Each physiological index (ACC, HR, TEMP, EDA, BVP, and IBI) collected using Empatica was also extracted at 5-minute intervals to match blood glucose, which had the longest measurement interval (5 minutes).

In addition, when the Cho et al [28] dataset was updated from version 1.0.0 to version 1.1.0 on March 6, 2023, the results of nutrient value calculations from the dietary survey records were added to the analysis dataset. The parameters of the calculated nutritional values were calories, total carbohydrate (carbon), dietary fiber, sugar, protein, and total fat. The values of these nutritional assessment indices at each time point were summed.

The correlation of each physiological index (ACC, HR, TEMP, EDA, BVP, and IBI) was calculated and examined to determine whether their mean values or SDs affected the mean value or SD of blood glucose. Mean values and SDs were calculated for blood glucose and each physiological index at three 5-minute points in the same individual (10 minutes in total), and their correlation coefficients were calculated. To examine the impact of each physiological indicator on blood glucose before and after the time of collection of blood glucose data, lag data (data on physiological indicators before glucose data collection at 8 points [120, 105, 90, 75, 60, 45, 30, and 15 minutes] and after blood glucose data collection at 8 points [15, 30, 45, 60, 75, 90, 105, and 120 minutes]) were collected, and the correlation coefficient between blood glucose and each physiological indicator was calculated for each physiological index.

Lag data for physiological indicators before glucose data collection were created by time-shifting each physiological indicator every 15 minutes until 120 minutes (Figure S1 in Multimedia Appendix 1). Lag data for physiological index data after blood glucose data collection were time-shifted by 15 minutes for each glucose reading to 120 minutes (Figure S2 in Multimedia Appendix 1). The purpose of creating lag data was to calculate the correlation between glucose levels and ACC (ACC values at 15, 30, 45, 60, 75, 90, 105, and 120 minutes) before and after measurement. The correlation coefficient was calculated using Spearman correlation.

To examine the relationship between postprandial blood glucose rise and fall and physiological and dietary nutritional assessment indices, multiple regression analysis was performed on the relationship between the slope before and after the peak in postprandial glucose over time and physiological and dietary nutritional indices. Multiple regression analysis was performed using objective and explanatory variables.

As a supplementary analysis to the multiple regression analysis, a 1-way ANOVA was performed to compare the relationship between the upward and downward slopes of blood glucose and the groups above and below the median for each indicator. The groups below the median were used as comparison controls. The following variables were used in the 1-way ANOVA.

No ethical approval was required since this study was exclusively based on published literature.

We reviewed the evolution of the correlation coefficients, including the lag data for glucose, to determine whether the impact of each physiological indicator on glucose was influenced by physiological indicators before and after the time the glucose data were collected (Table 2).

The results showed that some indices such as ACC, TEMP, EDA, and IBI showed data correlations before and after the collection of blood glucose data. Activity, exercise, and stress are factors that influence blood glucose levels. Factors such as HR, body temperature, and autonomic nervous system function, including the sweating motor response, were associated with blood glucose variability [30]. The mean blood glucose and ACC SDs were negatively correlated; however, the uptake of blood glucose by mild to moderate physical activity is reported to cause decreased blood glucose levels and increased glucose production in the liver, leading to increased blood glucose [56]. This may be attributed to fluctuations in physical activity. The mean values of blood glucose and HR were negatively correlated; however, previous reports have indicated a negative relationship between blood glucose and HR variability, as sympathetic dominance increases with increasing blood glucose [57,58], which contradicts previous results. This could be attributed to, among others, increased in HR due to fasting-related hypoglycemia [58] and mild to moderate physical activity [56].

Regarding the correlation between mean blood glucose and mean TEMP, intravenous administration of blood glucose increased heat production by 20%, accompanied by an increase in TEMP after 55 minutes, which was presumed to be caused by this effect [59]. Regarding the correlation between blood glucose and EDA, EDA can increase during stress and is mediated by stress-induced activation of adrenergic hormones and cortisol. This increases blood glucose production and is thus positively related. Blood glucose levels increase in some individuals and decrease in others in response to stressful situations. Naturally occurring daily stressors may be associated with increased glycemic instability from hypoglycemia and decreased food intake, possibly due to these factors [60].

The mean blood glucose and SD of the physiological indices were negatively correlated.

This is because the mean and SD of physiological indicators were calculated from the mean and SD of the 3 points at individual 5-minute intervals, and the variation in physiological indicators was greater before and after the peak rise in glucose, whereas the variation in relevant physiological indicators was smaller at the peak of the rise in blood glucose. Meanwhile, the SD of blood glucose and physiological indicators were positively correlated, which was attributed to the increased variability in the SD of physiological indicators and blood glucose at times of blood glucose fluctuation (prepeak and peak transition).

To investigate the relationship between the slopes of the blood glucose rise and fall peaks and the mean values of physiological indices and nutritional assessment indices, as well as the relationship between the slope of the blood glucose rise and fall peaks and the mean values of physiological indices, a multiple regression analysis was conducted. A summary of the results is shown in Table 3.

The physiological and nutritional assessment indices associated with the slope of the peak blood glucose increase were TEMP, calories, carbon, dietary fiber, protein, ACC, HR_30, and EDA_90. For TEMP, a positive association was observed; however, this was presumably due to the reported association between increased glucose and TEMP [59]. For carbon, a positive association was found, which was thought to be because carbohydrates contribute to the increase in blood glucose, although they are not absorbed as rapidly as glucose [61].

Although the upward slope of blood glucose and the calorie mean were negatively associated, the 1-way ANOVA revealed a positive association, as did carbon. Therefore, this result may be due to the multicollinearity effect of simultaneously introducing carbon and calories (r=0.78; P<.001), which are strongly correlated as explanatory variables.

Dietary fiber intake lowers postprandial and average daily blood glucose levels [62]. Protein intake does not increase plasma glucose levels but rather decreases them; thus, protein intake with glucose suppresses the postprandial increase in glucose [63], contrasts with the expected result. This was inferred to be due to increased carbohydrate intake since total carbon and protein intake were positively correlated (r=0.49; P<.001). A negative association was found for ACC and HR, which was presumed to be physical activity-related increased HR, accompanied by decreased blood glucose levels [56]. A positive relationship was found for EDA_90. EDA increases during stress and is mediated by stress-induced activation of adrenergic hormones and cortisol, which increases gluconeogenesis [60].

The physiological and nutritional assessment indices associated with the slope of the descending glucose peak were ACC, sugar, HR_af, and HR_90; positive associations were found between the downward slope of blood glucose and ACC and HR mean values. The decrease in blood glucose is potentially due to an increase from mild to moderate physical activity (HR also increased with physical activity), resulting in a smaller upward slope of blood glucose and therefore creating a smaller downward slope [56]. Although a negative relationship was observed between the downward slope of blood glucose and sugar, similar to carbohydrates, sugar intake has been reported to increase blood glucose [64], The upward slope of blood glucose is greater when carbohydrate intake is higher, and therefore the downward slope is also greater.

As a supplementary analysis to the multiple regression analysis, a 1-way ANOVA was performed to compare the relationship between the upward and downward slopes of blood glucose and the groups above and below the median for each indicator, using variables created by patterned group combinations by physiological indicator mean and dietary nutrient value. In addition, below-median groups were used as comparison controls. A summary of the results is presented in Table S33.

The results of the multiple regression analysis showed that the physiological indicators associated with the slope of rising and falling blood glucose were TEMP, ACC, HR, and EDA, whereas the nutritional indicators were carbon, protein, calories, sugar, and dietary fiber; therefore, these indicators were of interest.

The combination pattern of physiological indicators TEMP:ACC:HR:EDA, which included TEMP and EDA, showed that the upward and downward slopes of blood glucose were greater in the group above than in the group below the median. The effect on the upward slope of blood glucose was particularly large, similar to the results of multiple regression analysis. The combination pattern TEMP:ACC:HR:EDA=1011, which had a large upward slope of blood glucose, was due to stress, with high EDA and HR. The combination patterns TEMP:ACC:HR:EDA=1101 and 1111 also reported high glucose levels during and after moderate-intensity exercise [65] and high EDA due to exercise-induced sweating [66]. TEMP and EDA are strongly associated with autonomic nervous system function, which, in turn, is sensitive to glucose fluctuations, especially hyperglycemia. Therefore, we inferred that a high association exists between these indicators and the increase and fall in blood glucose slopes [31].

Conversely, for ACC and HR, the upward and downward slopes of blood glucose were slightly smaller in the group above than in the group below the median, accompanied by decreased blood glucose levels due to mild to moderate increases in physical activity [56] and physical activity-related increased HR similar to the results of the multiple regression analysis.

In the combination pattern of the nutrient indices carbon:protein:calories:sugar:fiber, the combination pattern including carbon, calories, and sugar showed a greater upward and downward slope of blood glucose in the group above than in the group below the median.

Dietary fiber reduces postprandial and mean daily blood glucose levels [62]; however, a group pattern of combinations with a greater rise and fall slope in blood glucose levels exists, with fiber above the median. The rise and fall slopes were smaller than those for groups of combinations below the median, similar to the results of the multiple regression analysis.

Two previous studies have used the wearable database in this study, BIG IDEAs Lab Blood Glycemic Variability and Wearable Device Data [30,31]. These studies were conducted to establish a prediction model for blood glucose-related indicators using machine learning and used a random forest regression model as the method of analysis. This analysis can be difficult to interpret in a specific clinical context.

In contrast, this study used simple analysis methods, such as correlation and multiple regression analyses, to explore the relationship between blood glucose and each physiological and nutritional index in a real-world setting. Despite the difference in analysis methods, the results for glucose and related indicators of our study were consistent with those of previous studies, reinforcing the results of this study. In addition, compared with the random forest model, these simpler analyses revealed more details of the relationship between blood glucose and each physiological and nutritional index.

Other publicly available open datasets on diabetes are shown in Table S1 in Multimedia Appendix 1. In total, 4 studies were conducted using PhysioNet’s D1NAMO Multimodal Dataset for Noninvasive Type 1 Diabetes Management Studies (2018) dataset. This dataset was collected to contribute to the development of data-centric algorithms and diabetes monitoring techniques by providing an openly available multimodal dataset. It was obtained from real patients in a nonclinical setting, containing electrocardiogram signals, respiratory signals, accelerometer output, blood glucose level information, and annotated food photographs [67]. Studies conducted with this dataset include the following: a study that used machine learning to predict blood glucose in patients with type 1 diabetes [20]; 1 study aimed at improving the accuracy of CGM systems [21]; an insulin absorption simulation study [22]; and a study for predicting diabetes [24], which is useful in several approaches.

This study has some limitations. First, as we used publicly available data, other data, such as detailed patient background data (height and weight) collected during a clinical study but not made publicly available, were not included in the analysis. Second, Empatica, a wearable device, continuously collects a vast amount of data on physiological indicators, whereas the Dexcom G6, which measures glucose, collects data at 5-minute intervals. For the analysis of the physiological indicators, data were extracted according to the glucose measurement interval, and data that were not extracted could not be considered. Third, as the data were from 16 participants, which is a small sample size, the calculation of correlation coefficients and multiple regression analysis were conducted; however, only exploratory studies were possible. Fourth, while data at the beginning of the meal were available for all cases, data at the end of the meal were available only for some cases. Therefore, in analyzing the relationship between the upward and downward slopes of postprandial glucose to the peak and the nutritional index of the meal, the nutritional index immediately before the peak was added together. Therefore, the effect of mealtime length may not have been considered.

Existing data from clinical studies on wearable-type devices (Dexcom 6 CGM and Empatica) from PhysioNet, a public open dataset, were used secondarily to examine the association of blood glucose with physiological and nutritional indices in 16 patients with borderline diabetes. The results showed that physiological indices associated with blood glucose were physical activity, HR, TEMP, and EDA, a stress indicator. In addition, physiological indices that were associated with the slope of the peak of the rise and fall of blood glucose were TEMP, physical activity, HR, and EDA. Nutritional measures associated with the slope of the peak rise and fall of blood glucose were carbohydrates, dietary fiber, and sugars. For the 3 analyses, the physiological measures associated with blood glucose were similar and consistent with previous reports.

The wearable-type device dataset allowed for the examination of the relationship of blood glucose with physiological and nutritional indicators. Research using existing data is expected to increase as open datasets of wearable device data become more readily accessible through data sharing and as it becomes possible to perform statistical analysis from various angles using such data.