Funding: This study was funded by the Natural Science Foundation of Shandong Province of China, grant number ZRGQ HW received the grant.

Competing interests: The authors have declared that no competing interests exist. Type 2 diabetes mellitus T2DM has rapidly become one of the most common non-communicable diseases NCDs globally, and one of the most challenging public health issues [ 1 ].

The International Diabetes Federation IDF estimated that million people had diabetes worldwide in , and this will rise to million by [ 2 ].

In China, a rapid increase in the prevalence of diabetes has been reported. Comparison of the latest national diabetes survey in with the first in shows diabetes prevalence has increased fold [ 3 — 4 ].

T2DM, as a chronic disease, affects people throughout their lifetime. T2DM patients are tasked with performing effective self-management. According to the Global Guideline for Type 2 Diabetes, four basic diabetes self-management skills were recommended for T2DM patients, which included dietary control, physical exercise, regular medication and self-monitoring of blood glucose [ 6 ].

Multiple clinical trials and reviews have confirmed the vital role of self-management in reducing blood glucose and improving quality of life among DM patients [ 7 — 9 ].

In the literature, many factors have been reported affecting self-management behaviors among T2DM patients; these include socioeconomic positions, diabetes knowledge, heath beliefs, attitudes, motivation and social support, etc [ 13 — 18 ].

One of the key factors is self-efficacy. Many studies have also shown the direct influence of self-efficacy on physical and mental health [ 21 ]. In China, self-efficacy has been shown to be positively associated with medication adherence and walking exercise among hypertension patients [ 22 — 23 ].

However, no study has identified the relationship between self-efficacy and the cluster of self-management behaviors recommended for diabetes management. This study therefore aimed to examine the association between self-efficacy and self-management behaviors among Chinese T2DM patients, which might provide evidence to inform effective self-management interventions for them in China.

This was a cross-sectional study in Shandong Province, China. Shandong contains 17 prefectures and counties districts , and had a total population of nearly 99 million in The estimated number of T2DM patients was about , prevalence: 9.

This study selected respondents from those patients registered in the NCDs management system. A multi-stage stratified randomized sampling was employed as follows. First, four representative prefectures Qingdao, Weifang, Jinan and Heze were selected based on their geographic location and economic development status within the province.

Then, three subdistricts urban and three towns rural were randomly selected from each of the four prefectures. Furthermore, three communities from each subdistrict and three villages from each town were randomly selected. In total, 36 urban communities and 36 rural villages were selected.

Finally, 35 T2DM patients were randomly recruited from each selected community and village. This study was conducted in accordance with the Declaration of Helsinki. The protocol was approved by the ethics committee of School of Health Care Management, Shandong University, China.

All participants were informed prior to investigation, and consent forms were signed by participants themselves. The study was conducted between August and October in Patient information was collected using a structured questionnaire.

Trained data collectors from Shandong University delivered the questionnaire by face-to-face interview. To ensure the questionnaire understood by the patients appropriately, all interviewers were required to make uniform explanations to the interviewees based on their local culture and customs.

To ensure the quality, the dedicated supervisors carefully checked all completed questionnaires each day after interviews. Self-developed questions were used to measure the four behaviors based on expert consensus and previous works [ 28 , 29 ].

In accordance with recommendations from the China guidelines for type 2 diabetes [ 30 ], all dimensions of self-management behaviors were treated as binary indicators. The Cronbach's alpha coefficient of internal consistency reliability of DES-SF was 0.

Total scores ranged from 8 to 40, with a higher score indicating a higher self-efficacy level. Covariates included demographics and clinical variables. Comorbidity meant the presence of any diabetes-related diseases or conditions i.

Descriptive statistics mean [M] and standard deviation [SD] was presented for numerical variables. Categorical variables such as residence, gender, age groups and duration of disease groups were presented as their frequency and percentage.

Latent class analysis LCA was used to explore the observed classes of self-management behaviors [ 33 ]. We began with a baseline one-class model and proceeded to test models with successively larger numbers of classes.

The choice of the optimal number of classes was based on the comparison of the various class-size models using the following criteria: 1 the Bayesian Information Criteria BIC , where smaller BIC value indicates better fit model; 2 Akaike Information Criteria AIC , where smaller AIC value indicates better fit model.

Further, we also assessed the entropy of each model as another indicator for class separation, and performed Lo-Mendell-Rubin Likelihood Ratio tests LMRLR to test the significance of the difference in the likelihoods of two models [ 35 ]. Univariate and multivariate logistic regression analyses were used to examine the associations between self-efficacy and self-management behaviors, and to identify other factors associated with self-management behaviors.

Data were analyzed using Stata 15 StataCorp LLC, College Station, TX, USA for Windows. As shown in Table 1 , a total of T2DM patients were included in the analysis, comprising patients from urban areas and from rural areas.

The mean age of the participants was Majority of the participants were women With respect to education level, Overall, As shown in Table 2 , the mean DES-SF score was The score of each item ranged from 3.

The proportions of patients performing well self-management in medication adherence, dietary control, physical exercise and self-monitoring of blood glucose were As shown in Table 3 , models with one to four latent classes were estimated, where the one-class model was deemed a baseline.

Models were then compared based on the LCA criteria outlined in the analyses section. Fig 1 presented the estimated item-response probabilities for self-management behaviors in each of the two latent classes.

It showed that class 2 had relatively higher probabilities of self-management than class 1 in all aspects of self-management behaviors, including dietary control We named class 1 and class 2 the inactive and active self-management groups.

The estimated proportions of the two classes were Abbreviations : DC, Dietary Control; PE, Physical Exercise; RM, Regular Medication; SMBG, Self-Monitoring of Blood Glucose.

As shown in Table 4 , higher DES-SF score was significantly associated with higher possibility of active self-management behaviors. Other factors significantly associated with active self-management behaviors were residence, education level, household income per capita, duration of diabetes and diabetes comorbidity.

Our study applied LCA to classify self-management behaviors among Chinese T2DM patients into active and inactive self-management groups. Based on self-management behaviors classification, this study further explored the relationship between self-efficacy and self-management behaviors.

Our study found that the performance of self-management among T2DM patients was poor, with only This result was similar with a study in Chongqing province, China, which showed that only a half of DM patients have good self-management behaviors [ 12 ]. Meanwhile, our study revealed that DM patients had worst self-management performance in self-monitoring of blood glucose when compared with other dimensions of self-management behaviors.

However, all these insurance schemes did not pay for the blood glucose monitoring equipment and the test strips, which may bring a heavy financial burden for the patients. From the perspective of delivery of primary care services, DM management program, as one of the key contents of the national essential public health services EPHS , was implemented since when the new round of health system reform was initiated.

Patients registered in EPHS can freely receive lifestyle and medication guidance from their contracted primary health workers [ 37 ]. However, self-monitoring of blood glucose guidance was hardly available for these patients. In addition, fear of needles may also be a barrier of self-monitoring of blood glucose among T2DM patients [ 38 ].

Our results hinted that more educational and financial support should be given to these patients to improve their self-monitoring blood glucose. Consistent with previous studies, in which self-efficacy was found as a strong predictor of self-management behaviors among patients with hypertension, arthritis and cardiovascular diseases [ 20 , 39 , 40 ], this study showed that self-efficacy was positively related to self-management behaviors among Chinese T2DM patients.

Some possible explanations for that were that patients with higher self-efficacy were more likely to acquire disease-related knowledge and seek help from the doctors and their family members, which was greatly useful for them to maintain the self-management behaviors.

Moreover, previous studies confirmed that self-efficacy was associated with better quality of life, less depression and lower HbA1c in diabetes patients [ 41 ].

Therefore, it is necessary to promote the self-efficacy intervention in clinical setting. On the one hand, physician or health educators could use existing measures e. In addition, some innovative delivery ways can be introduced to the self-efficacy intervention, such as messaging interventions and family-based intervention [ 45 , 46 ], which may facilitate the exchange of information between the patients and physicians and help the patient get more support from family members or physicians.

Our study found that rural patients had worse performance in self-management than the urban ones. This disparity may be related to the different socioeconomic development levels between urban and rural areas.

Patients in urban areas usually had higher incomes and education levels, which had been found positively associated with self-management behaviors. Furthermore, this disparity may be due to urban-rural inequity in quality of healthcare services. In China, primary health care for diabetes in urban areas was mainly provided by general practitioners GPs , but by village doctors in rural areas.

Compared with village doctors, GPs usually have more comprehensive medical knowledge and skills, which may positively influence the quality of diabetes care.

Our study suggested that more measures should be taken to eliminate the difference in quality of primary care between and urban and rural areas. Consistent with Sarkar [ 47 ], our study showed that longer duration of diabetes and having a diabetes comorbidity were associated with better self-management behaviors.

Some common explanations for this were that patients with longer duration of diabetes may be more susceptible to diabetes complications, more dependent on self-management and may have longer time to develop a habit of self-management behaviors [ 10 ].

We noticed that hypoglycemia is a minority class in the dataset, which accounts for 2. In general, the classifier is less sensitive to detecting a minority class.

Nevertheless, in this work, the sensitivity can be further enhanced by reducing the thresholds of lower bounds at a cost of potential alarm fatigue.

This trade-off can be decided by clinicians on an individual case basis. We used the MAML approach to train population models and personalized models, which outperformed the transfer learning approach with a small amount of available data.

This fast adaption feature of the MAML approach can mitigate the cold-start issues when we provide the software to new T1D users with limited personal data. It is a common scenario in actual clinical settings since data collection is expensive and time-consuming.

Moreover, the MAML also improved the final average RMSE results in the ablation analysis Fig. The chronological partition of training, validation and testing set in this work was carefully selected. Random cross-validation can be found in previous work, which trained and validated machine learning models on the same dataset 24 , However, during the experiments, we noticed that there were temporal dependencies between the data points from nearby locations, especially in adjacent ones.

The features were derived with the small resolution of CGM, so the difference between consecutive time steps is sometimes negligible. In this regard, the use of random or stratified splitting methods would introduce underlying temporal correlation into training and testing sets, which could result in serious overestimation of model accuracy The ARISES app Supplementary Fig.

The source code of the app is not publicly available. We analyzed the performance of the app on an iPhone XS Max over 50 runs. The whole app has an initial storage size of The trained deep learning models were converted to mobile compatible format via TensorFlow Lite, which has a storage size of 1.

When the app received a new CGM measurement, it took 5. Model fine-tuning is performed by Amazon S3 buckets and SageMaker in the Amazon Web Services AWS cloud Fig.

Our results suggest that measurements obtained from wearable physiological wristband data sensors could be integrated alongside CGM data to improve the prediction of glucose levels and adverse glycemic events.

Interestingly, the IBI measured by the sensor wristband is the only predictor that has significant effect on both hypoglycemia and hyperglycemia prediction Fig. It indicates IBI or other heart rate variability HRV could be useful biomarkers in T1D decision support, which accords with the findings of previous studies 48 , 49 , However, the sensors in Empatica E4 are quite sensitive to motion artifacts, so it is difficult to obtain accurate measurements with too many hand movements In future work, an algorithm to detect exercise and reduce measurement error for the wristband will be developed.

Meanwhile, data extracted from manually recorded daily events have the potential to be used for the analysis of the drivers and patterns of the changes in plasma or interstitial glucose concentrations. During feature pre-processing, we calculated insulin on board and the carbohydrate on board with fixed duration i.

Nonlinear insulin on board and carbohydrate on board based on physiological models with personalized parameters. We collected the dietary data from the T1D participants under free-living conditions, so the dietary reporting is variable in quality but reflects the real-world use of carbohydrate counting and self-management.

It is noted that the percentages of time spent below range Table 1 are small, and there is a modest carbohydrate intake of — grams per day Supplementary Table 2. Although these values are not unusual for people living with T1D, especially for those who use CGM to visualize post-prandial glucose peaks, it is a potential limitation in the development of the algorithm to predict hypoglycemic events.

Future work will include validating the proposed system on a T1D cohort with greater variance in carbohydrate intake and glycemic variability. Currently, there are no publicly available T1D datasets containing all the data fields that we need in the ARISES model, but it is important to further test the generalization of the proposed algorithm using an independent validation dataset with a larger cohort size.

In this case, we also recommend to analyze covariates in the T1D population, such as age and insulin delivery mode. In addition, there is a lack of system testing of the whole ARISES in real-world settings.

It might be challenging to simultaneously administer the multiple wearable devices, smartphone app, and cloud services with reliable wireless connectivity. A deep learning model with only CGM input and daily entries needs to be implemented as a sub-optimal solution when the wristband data is not available, e.

This was a six-week longitudinal prospective study NCT ID: NCT using a clinically validated real-time physiological data acquisition sensor Empatica E4 and CGM Dexcom G6 to identify correlations between measurable physiological parameters and glycemia.

Under free-living conditions, twelve adults 18 years old and older with a median age IQR of 40 years 30—50 were equally stratified by gender and mode of insulin delivery MDI and CSII.

Participants were recruited from the Imperial College Healthcare NHS trust T1D outpatient clinics, registered research databases, and interested participants who contact us. Participants were asked to log daily events such as, insulin doses in units, meal macronutrient composition in grams, alcohol intake in units, stress, illness, and exercise in the mySugr smartphone app, which are used to develop the input features of glucose prediction models.

Different from most of the previous studies using CGM and daily manual logs 20 , 21 , 22 , 23 , 24 , 26 , an objective of this work is to better understand the effect of the non-invasive physiological data on the prediction of glycemic events.

Using the package lme4 in R, a mixed effects logistic regression was applied calculate the logarithm of ORs to interpret the relationship between physiological measurements and the binary outcome of adverse glycemic events i.

The measured physiological variables applied to the regression analysis include the mean values, standard deviation, range, and maximum and minimum differential values of EDA, IBI, acceleration, and skin temperature signals.

As a clinically validated, commercially available, and non-invasive device, the Empatica E4 wristband uses a photoplethysmography sensor to measure blood volume pulse BVP , two electrodes to obtain EDA, a pair of accelerometers and a gyroscope to detect the level of physical activity, and a peripheral temperature sensor to monitor skin temperature.

In previous clinical studies, BVP and ECG signals are the primary sources to identify IBIs for HRV analysis In particular, we applied a band-pass Butterworth filter to remove noise in BVP signals and employed a slope sum function 55 to detect the local maxima.

Then we used a sliding window with decision rules 55 to search peaks, as the systoles in cardiac cycles. The IBIs were computed by the difference of consecutive peaks. We extracted short-term HRV features with a 5-minute window to indicate early HRV changes 48 in temporal and frequency domains To obtain skin conductance levels SCLs and skin conductance responses SCRs , we continuously decomposed EDA data into tonic and phasic components via a high-pass filter There are two open-source software tools involved in EDA and BVP processing 58 , Together with physical activity levels and skin temperature, the outcomes of these features in the past five minutes were averaged and aligned with the time steps of CGM readings.

HRV is an established indicator that reflects cardiac autonomic activities, while EDA is related to the status of the nervous system.

These biomarkers have been used in previous studies to predict and detect hypoglycemia for T1D 48 , 50 , The daily entries were converted to insulin on board and carbohydrate on board via physiological models. We assumed insulin bolus lasts for four hours in the human body with different slopes, as a common setting used by many commercialized pumps 24 , Similarly, the carbohydrate was assumed to be absorbed at a rate of 2.

We obtained a total of 20 features from the pre-processed multi-modal data Supplementary Table 3. There are some inevitable errors in the sensor data, e. To this end, we performed feature selection in the following steps.

First, we analyzed the missing fraction of CGM and wristband measurements to identify the quality of features. The median value of the missing percentages of CGM and wristband data are 3.

We linearly interpolated the gaps that occurred in the middle of input sequences and extrapolated the gaps at the tail to guarantee that future information is not involved in current predictions.

Then, min-max normalization was adopted to scale the selected features to [0, 1]. Finally, we performed collinearity analysis, considering correlated bias is prone to degrade the stability and interpretability of machine learning models We noted that features derived from the same measurement exhibited strong a correlation with each other.

Hence, each time we retained one feature in IBI or EDA feature group Supplementary Table 3 and selected the best combination according to the error scores that summed up RMSE results for the four prediction horizons in model validation. Considering the personalized models are provided to the T1D subjects at a midterm clinical visit Fig.

To simulate a clinical scheme with two phases Fig. Data of each subject in the population set were used to optimize the population model. The population models were validated with leave-one-subject-out cross validation. Then we used the training data of the personalized set to fine-tune the population model to develop a personalized model, and used the testing data of the personalized set for evaluation.

The Hyperband algorithm 63 of Keras tuner was used to select the best hyperparameters of the deep learning models Supplementary Table 7. Besides, we used early stopping to mitigate overfitting and improve generalization.

With the population and personalized data sets, we applied a well-established MAML framework to develop population models Each subject is regarded as a learning task in the inner loop of MAML. Then, the learned parameters for each task guided the population model to achieve meta-optimization via stochastic gradient descent in the outer loop.

The first-order approximation was performed to accelerate the training process In the personalization phase, we fine-tuned the meta-model with a small learning rate We performed an experiment to compare the MAML population model with the pre-trained model by classic transfer learning For each subject, the data collected on the fist day of the trial, i.

Then, we evaluated the performance of the fine-tuned models using the testing data of the personalized set. The recurrent structure is well-suited to learn short and long-term temporal dependencies in sequence processing. Thus, RNN-based models are emerging in the literature of diabetes management and have been shown to exhibit superior performance in glucose prediction 26 , 33 , 36 , However, the vanilla RNNs face the challenge of gradient exploding and vanishing, which largely limits the learning performance on long-term temporal dependencies.

Fortunately, long short-term memory LSTM 67 and gated recurrent units GRUs 68 were proposed to solve this problem. The GRU uses reset and update gate functions with less parameters than LSTM Fig. A stack of bidirectional RNN Bi-RNN and RNN with GRU cells is used to extract hidden representations from the input multivariate time series.

With the weighted state vector by attention layer, the evidential output computes glucose predictions along with model uncertainty. After pre-processing the features, we developed an attention-based RNN with GRUs for glucose prediction and hypo- and hyperglycemia detection.

The multivariate input data for the RNN model were selected according to validation performance, which include CGM, carbohydrate amount, insulin bolus, time index, IBIs, and SCRs.

The output of the bidirectional RNN is sent to another GRU-based RNN layer to obtain high-level hidden representations, of which the cell output is denoted as h t.

We employed the attention mechanism, as one of the latest advances in deep learning, to extract temporal dependencies regardless of distance.

Introducing attention in deep neural networks has shown success in a variety of tasks, especially in natural language processing Instead of only using the output of the final state, the attention mechanism assigns attention weights to the hidden state h t on each time step and then combines them to compute the final representation vector.

To determine the reliability and confidence of the predictions, model uncertainty is estimated by a higher-order evidential distribution Assuming the predictions are drawn from a Gaussian distribution with unknown mean μ and variance σ 2 , i. The output of the evidential layer evid and epistemic uncertainty i.

Clinicians are allowed to adjust the thresholds to obtain specific clinical efficacy. For instance, increasing the value of k can enhance the sensitivity of the classifier to avoid missing the warnings of adverse glycemic events.

During the model training, the negative log-likelihood loss function to optimize the parameters with maximum likelihood estimation can be solved by a Student-t distribution according to Bayesian probability theory The glucose predictions were estimated by the mean values of the evidential distribution of the model output.

The regression performance was evaluated by the RMSE, gRMSE, MAE, MAPE, and the time lag. In particular, gRMSE penalizes the prediction errors that could lead to harmful events, such as overestimation in hypoglycemia and underestimation in hyperglycemia, to demonstrate clinical impact 71 , which is defined as follows:.

The time lag is derived by the cross-correlation of predicted glucose levels and actual CGM measurements 34 , 35 , which denotes the time-shift between two time series. A smaller time lag indicates a faster response of the prediction method to the changes in CGM trends and thus better prediction performance.

The thresholds of lower and upper bounds were selected in model validation according to MCC scores, which are respectively used to detect hypoglycemia and hyperglycemia. In particular, a hypoglycemic or hyperglycemic event is defined as three consecutive CGM measurements i.

A true positive means that an adverse glycemic event is correctly identified, while a false negative indicates a missed prediction.

We evaluated the classification performance of hypo- and hyperglycemia prediction using a set of standard metrics, including accuracy, sensitivity, specificity, precision, and MCC 22 , 23 , Good MCC scores can be obtained only if the classifier performs well in all confusion matrix categories, which is a more reliable and informative score than accuracy and the F1 score in binary classification In addition, we introduced MD scores calculated by the MAE for the glucose sequences in missed predicted hypoglycemic or hyperglycemic events.

We used the results for the minute prediction horizon as the primary outcomes, since predicting glucose over such a long prediction horizon is challenging. The converted TensorFlow Lite models were evaluated to simulate on-device inference in the ARISES app.

To compare the proposed model with existing approaches, we employed a set of classic machine learning and deep learning baseline methods Supplementary Tables 2 , 5 and 6 , including support vector regression SVR with the RBF kernel 21 , artificial neural networks ANNs with three fully-connected layers 20 , bidirectional long short-term memory Bi-LSTM 32 , and CRNNs Besides, we also used a statistical model, the ARMA with exogenous inputs 44 , and a physiological model, the PKM, which is based on the composite minimal model of plasma glucose and insulin kinetics with personalized insulin sensitivity, time to maximum glucose rate of appearance, and time to maximum insulin absorption 45 , The input features of baseline models were identical to those of the proposed model, except that the PKM only used the information of CGM measurements, carbohydrate intake and insulin bolus.

To calculate the statistical significance with respect to the considered baseline results, we performed paired t tests after evaluating the normality by Shapiro—Wilk tests.

Further information on research design is available in the Nature Research Reporting Summary linked to this article. The dataset used in this study is not publicly available due to the proprietary nature of the data and privacy concerns.

Interested researchers should contact the corresponding authors to inquire about the access. The free and open-source programming languages R 3. The source code of the deep learning model and smartphone app is available upon reasonable request from the corresponding authors.

Saeedi, P. et al. Global and regional diabetes prevalence estimates for and projections for and Results from the International Diabetes Federation Diabetes Atlas. Practice , Article Google Scholar. Katsarou, A. Type 1 diabetes mellitus. Disease Primers 3 , 1—17 Google Scholar.

Yale, J. Can J Diabetes 42 , S—S Article PubMed Google Scholar. Gregg, E. The changing face of diabetes complications. Lancet Diabetes Endocrinol. Rodbard, D. Continuous glucose monitoring: a review of successes, challenges, and opportunities.

Diabetes Technol. Therapeutics 18 , S2—3 Article CAS Google Scholar. Juvenile Diabetes Research Foundation Continuous Glucose Monitoring Study Group. Continuous glucose monitoring and intensive treatment of type 1 diabetes. Heinemann, L.

Real-time continuous glucose monitoring in adults with type 1 diabetes and impaired hypoglycaemia awareness or severe hypoglycaemia treated with multiple daily insulin injections HypoDE : a multicentre, randomised controlled trial.

Lancet , — Article CAS PubMed Google Scholar. Herrero, P. A bio-inspired glucose controller based on pancreatic β -cell physiology. Diabetes Sci. Article PubMed PubMed Central Google Scholar.

Oliver, N. Open source automated insulin delivery: addressing the challenge. npj Digital Med. Kirwan, M. Diabetes self-management smartphone application for adults with type 1 diabetes: randomized controlled trial. Internet Res.

Ryan, E. Improved A1C levels in type 1 diabetes with smartphone app use. Diabetes 41 , 33—40 Sevil, M. Determining physical activity characteristics from wristband data for use in automated insulin delivery systems. IEEE Sensors J. Ozaslan, B. Impact of daily physical activity as measured by commonly available wearables on mealtime glucose control in type 1 diabetes.

Article CAS PubMed PubMed Central Google Scholar. Wu, Y. Mobile app-based interventions to support diabetes self-management: a systematic review of randomized controlled trials to identify functions associated with glycemic efficacy. JMIR mHealth uHealth 5 , e35 Lithgow, K.

Smartphone app use for diabetes management: evaluating patient perspectives. JMIR Diabetes 2 , e2 Mathieu, C. Insulin analogues in type 1 diabetes mellitus: getting better all the time.

Battelino, T. Prevention of hypoglycemia with predictive low glucose insulin suspension in children with type 1 diabetes: a randomized controlled trial. Because the NHANES III was a cross-sectional analysis of patient behavior and glycemic control, a cause-and-effect relationship would be difficult to determine.

Prospective study designs need to be employed to better understand the role of SMBG in all patients with diabetes. Nevertheless, most diabetologists agree that self-management of diabetes needs to incorporate some SMBG data, and that motivated patients can benefit from the increased empowerment that SMBG yields.

Diabetes specialists now recommend that patients use SMBG data for day-to-day regimen changes and that health care professionals use SMBG data to guide alterations in medication regimens. The American Diabetes Association has sanctioned efforts to teach people with diabetes to use SMBG data actively as part of a patient-centered self-management program.

SMBG works by having patients perform a number of glucose tests each day or each week. The test most commonly involves pricking a finger with a lancet device to obtain a small blood sample, applying a drop of blood onto a reagent strip, and determining the glucose concentration by inserting the strip into a reflectance photometer for an automated reading.

People with diabetes can be taught to use their SMBG results to correct any deviations out of a desired target range by changing their carbohydrate intake, exercising, or using more or less insulin.

The frequency with which patients with diabetes should monitor their blood glucose level varies from person to person.

Most experts agree that insulin-treated patients should monitor blood glucose at least four times a day, most commonly fasting, before meals, and before bed. In addition, patients using insulin can benefit by obtaining postprandial blood glucose readings to help them more accurately adjust their insulin regimen.

A positive correlation between frequency of SMBG and glycemic control among patients with insulin-treated type 1 or type 2 diabetes has been demonstrated. For patients with type 2 diabetes, optimal SMBG frequency varies depending on the pharmaceutical regimen and whether patients are in an adjustment phase or at their target for glycemic control.

If a patient is on a stable oral regimen with HbA 1c concentration within the target range, infrequent SMBG monitoring is appropriate. In such cases, patients can use SMBG data as biofeedback at times of increased stress or changes in diet or physical activity.

For patients whose diabetes is out of control or for those having medication initiated, however, SMBG data can be helpful in creating or modifying the diabetes management regimen.

Persistent fasting hyperglycemia, for example, might indicate excessive hepatic glucose output, and patients experiencing this problem might derive benefit from using metformin Glucophage , which has been shown to decrease nocturnal hepatic glucose output.

Additionally, patients with persistent postprandial hyperglycemia might derive benefit from taking a short-acting oral agent with meals to either decrease carbohydrate absorption i.

People with type 2 diabetes who use insulin should perform SMBG at least four times per week, including at least two fasting and two postprandial values. Additional measurements at bedtime and before meals can also be obtained.

Thoughtful interpretation of SMBG data will assist patients and health care providers in selecting appropriate pharmaceutical and lifestyle regimens. There are now seven manufacturers and more than 20 types of meters available on the market.

Meters vary in size, weight, test time, memory capabilities, and special features. Most meters can measure blood glucose with only a one- or two-step process.

Most also incorporate no-wipe technology, which means users do not have to wipe off excess blood after applying a blood drop to the reagent strip. In addition, many meters now require only a very small amount of blood, thus decreasing the pain of deep wounds from the lancet.

A few of the newer meters offer the option of obtaining blood samples from alternate sites, such as a forearm instead of a fingertip. This can benefit patients who find constant lancet wounds on their fingers difficult to tolerate.

Although there has been concern that the accuracy of alternate site testing is inferior to detect hypoglycemia, 4 the judicious use of this alternative may help to improve adherence to an SMBG regimen. More complex meters have features to aid in identifying trends and to graph reports for more comprehensive data tracking, particularly for patients who test several times a day.

Table 1 provides a summary of the more popular blood glucose meters. Recent reviews of available meters have been published elsewhere. SMBG can play an important role in improving metabolic control in patients with diabetes.

It is recommended for patients treated with insulin and is desirable for all patients with diabetes. Judicious use of SMBG data can help to improve glycemic control, select an anti-diabetic regimen, and provide powerful feedback to patients wishing to improve metabolic control.

Benjamin, MD, FACP, is an assistant professor of medicine at Tufts University School of Medicine and Director of Healthcare Quality at Baystate Medical Center in Springfield, Mass.

Sign In. Skip Nav Destination Close navigation menu Article navigation. Volume 20, Issue 1. Previous Article Next Article. SMBG Use and Frequency. Recommending a Meter. Article Navigation. Practical Pointer January 01 Self-Monitoring of Blood Glucose: The Basics Evan M.

Benjamin, MD, FACP Evan M. Benjamin, MD, FACP. This Site. Google Scholar. Clin Diabetes ;20 1 — Get Permissions. toolbar search Search Dropdown Menu. toolbar search search input Search input auto suggest. SMBG can aid in diabetes control by:.

Table 1. Blood Glucose Meters. View large. View Large. Evans JMM, Newton RW, Ruta DA, MacDonald TM, Stevenson RJ, Morris AD: Frequency of blood glucose monitoring in relation to glycemic control: observational study with diabetes database.

Franciosi M, Pellegrini F, De Bernardis G, Belfiglio M, Nicolucci A: The impact of blood glucose self-monitoring on metabolic control and quality of life in type 2 diabetic patients.

Diabetes Care. Harris MI: Frequency of blood glucose monitoring in relation to glycemic control in patients with type 2 diabetes. Jungheim K, Koschinsky T: Risky delay of hypoglycemia detection by glucose monitoring at the arm.

Self care of diabetes. Consumer Rep. Blood Glucose Monitors and Data Management in Buyers Guide Diabetes Forecast. American Diabetes Association.

View Metrics. Email alerts Article Activity Alert. Online Ahead of Print Alert. Latest Issue Alert. Follow-up appointments were made for patients before discharge, and the follow-up communication was conducted with each patient. Each intervention was measured on its effectiveness.

By utilizing continuous glucose monitoring in primary care settings, Nurse Practitioners can reduce the incidence of hospitalizations by effectively treating patients with type 2 diabetes. Angeles, Nikki Rose, "Optimizing Diabetes Self-Management Using Continuous Glucose Monitoring and Improving the Transition of Care for Adult Patients Discharged From the Hospital" Doctor of Nursing Practice.

Endocrine System Diseases Commons , Family Practice Nursing Commons. Advanced Search. Home About FAQ My Account Accessibility Statement. Privacy Copyright. Skip to main content. Tansey M , Laffel L , Cheng J , et al. Satisfaction with continuous glucose monitoring in adults and youths with type 1 diabetes.

Diabet Med. Rasbach LE , Volkening LK , Markowitz JT , Butler DA , Katz ML , Laffel LM. Youth and parent measures of self-efficacy for continuous glucose monitoring: survey psychometric properties. Beck RW , Bergenstal RM , Cheng P , et al. The relationships between time in range, hyperglycemia metrics, and HbA 1c.

Battelino T , Danne T , Bergenstal RM , et al. Clinical targets for continuous glucose monitoring data interpretation: recommendations from the International Consensus on Time in Range.

Battelino T , Phillip M , Bratina N , Nimri R , Oskarsson P , Bolinder J. Effect of continuous glucose monitoring on hypoglycemia in type 1 diabetes. Oxford University Press is a department of the University of Oxford.

It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide. Sign In or Create an Account. Navbar Search Filter Journal of the Endocrine Society This issue Endocrine Society Journals Endocrinology and Diabetes Books Journals Oxford Academic Mobile Enter search term Search.

Endocrine Society Journals. Advanced Search. Search Menu. Article Navigation. Close mobile search navigation Article Navigation. Volume 4. Article Contents Abstract. Materials and Methods.

Statistical Methods. Additional Information. References and Notes. Journal Article. Diabetes Telehealth Solutions: Improving Self-Management Through Remote Initiation of Continuous Glucose Monitoring. Robin L Gal , Robin L Gal. Jaeb Center for Health Research. Gal, MSPH, Jaeb Center for Health Research, Amberly Drive, Suite , Tampa, FL , USA.

E-mail: rgal jaeb. Oxford Academic. Nathan J Cohen. Davida Kruger. Henry Ford Health System. Roy W Beck. Richard M Bergenstal. International Diabetes Center Park Nicollet.

Peter Calhoun. Terra Cushman. Amanda Haban. UW-Madison Dept. of Family Medicine and Community Health: Wisconsin Research and Education Network WREN. Korey Hood. Stanford University School of Medicine. Mary L Johnson. Teresa McArthur , Teresa McArthur. Cecelia Health. Beth A Olson. Lagoon Health.

Ruth S Weinstock. SUNY Upstate Medical University. Sean M Oser. Tamara K Oser. Brian Bugielski. Heidi Strayer. Grazia Aleppo. Northwestern University Feinberg School of Medicine.

Corrected and typeset:. PDF Split View Views. Select Format Select format. ris Mendeley, Papers, Zotero. enw EndNote. bibtex BibTex. txt Medlars, RefWorks Download citation. Permissions Icon Permissions. Close Navbar Search Filter Journal of the Endocrine Society This issue Endocrine Society Journals Endocrinology and Diabetes Books Journals Oxford Academic Enter search term Search.

Abstract The purpose of this study was to evaluate feasibility of initiating continuous glucose monitoring CGM through telehealth as a means of expanding access. telehealth , continuous glucose monitoring , T1D , T2D.

Table 1. Participant characteristics at enrollment a. b Missing for 2 participants with T2D. Open in new tab. Table 2.

Continuous glucose monitoring use during the study a. Baseline HbA 1c. Table 3. Continuous glucose monitoring and glycemic control metrics during the study. Figure 1. Open in new tab Download slide. Table 4. Participant-reported outcomes a.

P for change from baseline to month 3 d. CGM Self Efficacy Scale Mean Score NA 5. Certified Diabetes Care and Education Specialist.

Wisconsin Research and Education Network. Google Scholar Crossref. Search ADS. Juvenile Diabetes Research Foundation Continuous Glucose Monitoring Study Group.

OpenURL Placeholder Text. For commercial re-use, please contact journals. permissions oup. Issue Section:. Download all slides. Views 6, More metrics information. Total Views 6, Month: Total Views: June 85 July 75 August September October November December January February March April May June July August September October November December January February March April May June July 99 August September October November December 72 January 99 February March April May June 61 July 64 August 58 September 75 October 96 November December 65 January February Email alerts Article activity alert.