Materials and Methods. A retrospective analysis of 524 continuous glucose monitoring (CGM) profiles of patients with type 1 diabetes mellitus was performed. CGM was performed using the Medtronic iPro2 system for 6-7 days, and the night periods of CGM were analyzed to identify regular NH. The study included 239 patients, of whom 65 (27.1%) experienced regular NH. We constructed the models of 7-point glycemic profiles, the data from which were uploaded to the DiaLog GM mobile application to calculate conventional glucose monitoring parameters. The prognostic model of NH was developed using the logistic regression method.
Results. Based on the regression analysis, the most significant predictors of NH included in the prognostic model were glycated hemoglobin level (p=0.001), use of insulin pump therapy (p=0.001), time below the target time in range (TIR) for blood glucose content of level 1 (p<0.001), and the coefficient of variation for glucose content (p=0.02). The area under the ROC curve for the prediction model was 0.917; the optimal cut-point value for the predicted probability of NH was 0.317, at which the sensitivity of the model was 86%, and its specificity was 90%.
Conclusion. Due to its higher predictive ability, the developed prediction model based on the data of a specialized mobile application allows improving existing approaches to assessing the risk of NH.
Introduction
Hypoglycemia in patients with diabetes mellitus (DM) receiving insulin therapy is a life-threatening condition associated with a significant reduction in their quality of life, increasing risk of developing hypoglycemic coma, and severe vascular complications of DM [1]. In conditions of self-monitoring of blood glucose (SMBG) based on using a glucometer, nocturnal hypoglycemia (NH) is a noteworthy diagnostic problem, especially in combination with the phenomenon of impaired awareness of hypoglycemia [2]. A more recent method of self-monitoring known as continuous glucose monitoring (CGM) has substantial advantages in the diagnosis of NH due to the measurement of interstitial fluid glucose every 1-5 minutes in order to investigate daily trends in blood glucose (BG) level [3].
Despite the fact that CGM systems were introduced into clinical practice over 20 years ago, most patients with DM still use SMBG as the main technique of self-monitoring [4]. The key limitations to the widespread use of CGM systems are the high cost of sensors, the need to train patients and medical personnel in the principles of CGM, and the difference between blood and interstitial fluid glucose values. The last listed issue may affect the accuracy of CGM [5, 6]. Thus, the insufficient sensitivity of CGM to NH remains a pressing issue requiring improvement of the conventional method of glycemic control in DM. The most promising developments in this field are prognostic algorithms for NH based on the values of glycemic control parameters [7-9]. Another direction in the development of CGM is the introduction of glucose monitoring information technologies into clinical practice, such as the use of specialized mobile applications and computer programs for glucometers [10]. Such software allows calculating mean glycemic values, glycemic variability, time spent in target and non-target glycemic ranges, and keeping track of the doses of administered insulin [11]. However, when reviewing recent publications, we did not find information about mobile applications for SMBG capable of predicting the risks of occurrence of NH episodes based on self-monitoring data.
Objective – To develop a predictive algorithm for NH based on data collected with a mobile application for glucose monitoring.
Materials and Methods
We performed a retrospective analysis of a database consisting of 524 CGM profiles. The analysis was conducted at the public health care institution, Saratov City Clinical Hospital No. 9 of Saratov, Saratov, Russia, from September 2016 through August 2023. Our study complied with the principles of Good Clinical Practice and was approved by the Ethics Committee of V.I. Razumovsky Saratov State Medical University of the Russian Federation Ministry of Healthcare (protocol #4 of November 7, 2023). Clinical and epidemiological parameters of patients were studied based on electronic patient medical records. The study included the results of CGM of patients with type 1 DM that met the following criteria: age range of 18-45 years; CGM device represented by Medtronic iPro2; Enlite MMT-7008 CGM sensor; 6 or 7 days of patient medical examination. CGM sensor calibration performed at least 4 times a day using a glucometer or biochemical analyzer. Exclusion criteria were as follows: pregnancy, severe somatic and infectious diseases, and glucocorticosteroid therapy at the time of CGM. To describe the CGM results, we used reports from the official Medtronic CareLink Professional web platform.
Our study included 239 patients with type 1 DM; the total number of CGM sensor measurements included in the analysis was 425,908. An analysis of 1,321 nights was performed, in which regular NH was detected in 65 people (27.1% of the total number of patients) (Group 1). The remaining patients (174 individuals) without regular NH formed the control group of the study (Group 2). The fact of the presence of regular NH was established when glucose level, according to the CGM sensor data, decreased to less than 3.9 mmol/L from midnight (00:00) to 06:00 for 2 days or more during the observation period of the patient.
To develop the prognosis model, 7-point SMBG profiles were simulated using CGM data on interstitial fluid glucose at a certain point in time, calibrated by the BG level. The selection of measurements for the SMBG models was carried out according to the patient’s journal data: 5 minutes before the start of a meal, 2 hours after it, and before bedtime. In the absence of journal records, the standard times of breakfast, lunch, dinner, and bedtime in the Department of Endocrinology of Saratov City Clinical Hospital No. 9 were assumed as 9:00, 13:30, 18:30, and 23:00. The total number of measurements in the developed SMBG models was 10,936.
For each SMBG profile, the following parameters were calculated: mean BG level (BGmean); glucose management indicator (GMI); time in the target range of BG level (TIR); total time above the target BG range (TARtotal); time above target BG range of levels 1 and 2 (TAR-1, TAR-2); total time below the target BG range (TBRtotal); time below the target BG range of levels 1 and 2 (TBR-1, TBR-2); standard deviation for BG level (SD), coefficient of variation in BG level (CV), and mean amplitude of glycemic excursions (MAGE). These parameters were calculated by a specialized mobile application, DiaLog GM (A.N. Rusanov, T.I. Rodionova, computer program registration certificate #2022665169 of August 11, 2022), developed in accordance with domestic and international recommendations for glycemia monitoring [12, 13]. Ranges of TIR, TAR-1, TAR-2, TBR-1 and TBR-2 were 3.9-10.0 mmol/L, 10.0-13.9 mmol/L, >13.9 mmol/L, 3.0-3.9 mmol/L and <3.0 mmol/L, respectively. CV was calculated using the following formula:
CV =SD/BGmean×100%.
The MAGE parameter was calculated based on the original algorithm reported in a previous publication [14]. The user interface of the DiaLog GM application with a sample analysis of one of the 7-point profiles of the SMBG is shown in Figure 1.
Statistical data processing was performed via SPSS Statistics 25.0 and Microsoft Office Excel 2019 software packages. The data are presented as median, 1st and 3rd quartiles: Me [Q1; Q3]. The normality of the distribution was assessed using the Kolmogorov-Smirnov test. The significance of differences was evaluated by the Mann-Whitney U test. Differences were assumed significant at p<0.05. When developing and evaluating the effectiveness of the predictive model, we employed the logistic regression method with stepwise inclusion of predictors, as well as receiver operating characteristic (ROC) analysis.
Results
The clinical and epidemiological characteristics, along with CGM and SMBG data of the patients included in the study, are presented in Table 1. We revealed no statistically significant differences in age, gender, or DM duration between Group 1 and Group 2 (p>0.05). The level of glycated hemoglobin (HbA1c) was significantly lower in Group 1 vs. Group 2: 7.7 [6.9; 8.2] % vs. 8.6 [8.2; 9.2] %, respectively (p<0.001). Patients with regular NH had a lower body mass index (BMI) and used insulin pump therapy, also known as continuous subcutaneous insulin infusion (CSII), less frequently than those in the control group (p<0.001).
Table 1. General characteristics of the study groups and glucose monitoring results
Parameters | Group | p | |
1 (NH+), n=65 | 2 (NH-), n=174 | ||
Clinical and epidemiological parameters | |||
Age, years Female gender, % Duration of diabetes, years BMI, kg/m2 HbA1c, % Proportion of patients on CSII, % Daily insulin dose, IU | 26 [22; 33] 62 15 [10; 20] 22.2 [20.5; 24.8] 7.7 [6.9; 8.2] 31 50 [40; 83] | 25 [21; 29] 52 15 [11; 20] 24.6 [21.3; 28.5] 8.6 [8.2; 9.2] 42 65 [48; 83] | 0.475 0.176 0.752 <0.001 <0.001 <0.001 0.631 |
CGM results | |||
Number of sensor-based measurements (CGM) Number of CGM days Mean glucose average content (CGM), mmol/L GMI (CGM), % TIR (CGM), % TARtotal (CGM), % TBRtotal (CGM), % SD (CGM), mmol/L CV (CGM), % MAGE (CGM), mmol/L | 1797 [1566; 1924] 6 [6; 7] 8.9 [7.9; 10.5] 7.2 [6.7; 7.9] 51 [43; 62] 37 [25; 48] 9 [5; 16] 4.3 [3.5; 5.0] 46 [40; 51] 10.0 [8.4; 11.4] | 1795 [1645; 1934] 6 [6; 7] 12.1 [10.7; 13.4] 8.5 [7.9; 9.1] 34 [24.46] 64 [51; 75] 1 [0; 2] 4.3 [3.7; 5.0] 36 [32; 40] 9.6 [8.4; 11.6] | 0.623 0.178 <0.001 <0.001 <0.001 <0.001 <0.001 0.334 <0.001 0.815 |
SMBG results | |||
Number of blood glucose measurements (SMBG) Mean blood glucose (SMBG), mmol/L GMI (SMBG), % TIR (SMBG), % TARtotal (SMBG), % TAR-1 (SMBG), % TAR-2 (SMBG), % TBRtotal (SMBG), % TBR-1 (SMBG), % TBR-2 (SMBG), % SD (SMBG), mmol/L CV (SMBG), % MAGE (SMBG), mmol/L | 46 [42; 48] 9.0 [7.4; 10.5] 7.2 [6.5; 7.9] 53 [43; 65] 33 [20; 51] 18 [13; 25] 13 [4; 23] 11 [5; 17] 7 [4; 11] 2 [0; 8] 4.2 [3.5; 4.9] 47 [42; 52] 8.0 [6.2; 9.3] | 46 [43; 48] 11.5 [10.4; 12.8] 8.3 [7.8; 8.8] 38 [28; 50] 61 [47; 70] 27 [21; 32] 61 [47; 70] 0 [0; 2] 0 [0; 2] 0 [0; 0] 4.4 [3.8; 5.1] 38 [34; 43] 8.6 [7.1; 9.9] | 0.889 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.036 <0.001 0.024 |
BMI, body mass index; HbA1c, glycated hemoglobin; CSII, continuous subcutaneous insulin infusion (insulin pump therapy); CGM, continuous glucose monitoring; GMI, glucose management indicator; TIR, time in range (for blood glucose level); TARtotal, total time above target glucose range; TBRtotal, total time below target glucose range; SD, standard deviation (for blood glucose level); CV, coefficient of variation (in blood glucose level); MAGE, mean amplitude of glycemic excursions; SMBG, self-monitoring of blood glucose; TAR-1, time above target blood glucose range of level 1; TAR-2, time above target blood glucose range of level 2; TBR-1, time below target blood glucose range of level 1; TBR-2, time below target blood glucose range of level 2.
Most of the parameters obtained as a result of glycemia monitoring using both methods (CGM and SMBG) differed significantly between the study groups: in Group 1, we observed higher values of TIR, TBRtotal, TBR-1, TBR-2 and CV (p<0.001), and lower values of TARtotal, TAR-1 and TAR-2 (p<0.001). When comparing identical parameters calculated by different methods of BG monitoring (CGM or SMBG), we obtained the following results. In Group 1, significant differences were established in the MAGE parameter (CGM: 10.0 [8.4; 11.4] mmol/L; SMBG: 8.0 [6.2; 9.3] mmol/L; p<0.001). In Group 2, we detected statistically significant differences between the measurement methods in BGmean (p=0.032), GMI (p=0.035), TARtotal (p=0.031), TIR (p=0.027), CV (p=0.001) and MAGE (p<0.001), while such differences in other parameters were not revealed (p>0.05).
Initially, the following parameters were included in the procedure of direct stepwise selection of predictors for the binary logistic regression model: BMI, HbA1c, use of CSII, glycemic control parameters measured by SMBG (BGmean, GMI, TIR, TARtotal, TAR-1, TAR-2, TBRtotal, TBR-1, TBR-2, CV, and MAGE). The results of the regression analysis are presented in Table 2. For the final prediction model, the software-based algorithm selected four parameters as predictors of NH: HbA1c, use of CSII, TBR-1 and CV. When analyzing multicollinearity, the variance inflation factor (VIF) values for the predictors were 1.33 for HbA1c, 1.02 for use of CSII, 1.89 for TBR-1 and 1.51 for CV.
Thus, the final logistic regression equation for determining the probability of NH is as follows:
P= 1/(1-e^(-z) ) .
In this equation, z=2.294-0.924 ×Hb_A1c-1.52 ×CSII+0.357×TBR1+0.087×C_V,
where P is the probability of regular NH, e is the natural logarithm base (~2.718), HbA1c is the glycated hemoglobin level (%), СSII is the use of insulin pump therapy (1 if present, 0 if absent), TBR1 is the percentage of BG values below the target range of level 1 (%), and CV is the coefficient of variation in BG level (%).
Table 2. Results of regression analysis used for developing a model for predicting nocturnal hypoglycemia
Predictors | β-coefficient | Standard error of β | Wald | p | Odds ratio (95% CI) |
Step 1 | |||||
TBR-1 (SMBG) Constant | 0.514 -2.781 | 0.07 0.311 | 54.53 80.02 | <0.001 | 1.67 (1.46-1.92) – |
Step 2 | |||||
CSII TBR-1(SMBG) Constant | -1.554 0.535 -2.232 | 0.454 0.072 0.331 | 11.73 54.66 45.43 | 0.001 <0.001 | 0.21 (0.87-0.52) 1.71 (1.48-1.97) – |
Step 3 | |||||
HbA1c CSII TBR-1(SMBG) Constant | -0.791 -1.611 0.468 4.559 | 0.258 0.463 0.075 2.186 | 9.39 12.12 39.26 4.35 | 0.002 <0.001
0.037 | 0.45 (0.27-0.752) 0.2 (0.08-0.49) 1.6 (1.34-1.85) – |
Step 4 | |||||
HbA1c CSII TBR-1(SMBG) CV Constant | -0.924 -1.52 0.357 0.087 2.294 | 0.28 0.467 0.084 0.037 2.413 | 10.91 10.61 17.98 5.43 0.9 | 0.001
<0.001 0.02 0.342 | 0.4 (0.23-0.69) 0.22 (0.09-0.55) 1.43 (1.21-1.69) 1.09 (1.01-1.17) – |
TBR-1, time below target blood glucose range of level 1; SMBG, self-monitoring of blood glucose; CSII, continuous subcutaneous insulin infusion (insulin pump therapy); HbA1c, glycated hemoglobin; CV, coefficient of variation (in blood glucose level).
The ROC curve for the developed prognostic model is presented in Figure 2. The area under the curve was 0.917; 95% confidence interval (CI): 0.866-0.967. The optimal cut-point value for the predicted probability of NH was 0.317, at which the sensitivity of the model was 86%, and its specificity was 90%.
Discussion
To date, there are several available models for predicting NH, among which the most effective are those based on a large volume of data obtained via CGM [7, 15, 16]. Prognostic models of NH based on SMBG have been developed to a lesser extent. K. Sakurai et al. proposed a formula for assessing the risk of NH that includes age, fasting BGmean, and daily basal insulin requirement as predictors. However, according to reviewed publications, data on the validation of this approach to predicting NH are scarce [8]. In the study by S. Wang et al., the authors proposed to use the largest amplitude of glycemic excursions (LAGE) estimated on the basis of a 4-point SMBG. It was suggested that the LAGE value implying a high risk of NH is greater than 3.48 mmol/L, area under the curve is 0.587 (95% CI: 0.509-0.665), sensitivity is 66.7%, and specificity is 50% [9]. It should be emphasized that a promising method for predicting hypoglycemia is the use of machine learning and artificial intelligence technologies, but this approach is employed mainly in conjunction with CGM [17].
A feature of our study is the use of a specialized mobile application, with the help of which, in the course of conventional SMBG, the principles of data analysis inherent to CGM are used [12,13]. The clinical predictors of NH in the developed prediction model were HbA1c and the use of CSII, which is consistent with the results of previously conducted studies in this area [17]. CV and TBR-1 were the predictors of NH directly calculated using the mobile application. Currently, it is known that CV calculated based on CGM data is the most meaningful parameter of glycemic variability associated with the risk of hypoglycemia (including NH) [18]. The results of our study confirmed the possibility of using SMBG for calculating CV to predict NH. According to the results of regression analysis, a high level of TBR-1 was the most significant (and qualitatively new in comparison with previously published data) predictor of NH. Hence, a high percentage of glycemic values in the range from 3.0 to 3.9 mmol/L during the daytime based on SMBG data may be a noteworthy risk factor for NH.
The developed prognostic model has preliminarily demonstrated a fairly high efficacy in identifying NH. The limitations of our study were the sample of patients whose follow-up care was carried out mainly in hospital settings and the use of hypothetical SMBG models to calculate glycemic monitoring parameters. Further research on the proposed prognostic model is required for internal and external validation.
The developed formula for predicting regular NH in type 1 DM can be used in clinical practice if there is information on the HbA1c level and 7-point glycemic control data over the period of 6-7 days. The original mobile application presented by us substantially simplifies the process of data collection, calculation of glycemic control parameters, and transfer of results to the attending physician. Yet standardized methods for calculating CV and TBR-1 allow using the developed prognostic model without the mandatory use of the proposed application. With a predicted probability of NH≥0.317% in patients with type 1 DM, we recommend improving self-monitoring measures by using CGM or additional BG measurements at night in case of SMBG. As part of the further development of the DiaLog GM mobile application, we plan to incorporate the developed algorithm for predicting NH into this software, along with prospective studies in the field of machine learning for resolving similar issues.
Conclusion
Despite the active use of modern insulin preparations and insulin pump therapy in clinical practice, NH remains a fairly common problem among people with DM. Specialized mobile applications for BG monitoring supplemented with prediction algorithms can significantly simplify the detection of NH, which is especially important for patients with impaired awareness of hypoglycemia. Future use of such technologies can reduce the risk of severe hypoglycemia and chronic complications of DM, along with improving the patient’s quality of life.
Author contributions: All authors contributed equally to the preparation of the manuscript.
Conflict of interest. None declared by the authors.
Yeh JS, Sung SH, Huang HM, et al. Hypoglycemia and risk of vascular events and mortality: A systematic review and meta-analysis. Acta Diabetol. 2016; 53 (3): 377-92. https://www.doi.org/10.1007/s00592-015-0803-3
Klimontov VV. Impaired hypoglycemia awareness in diabetes mellitus: Epidemiology, mechanisms, and therapeutic approaches. Diabetes Mellitus 2018; 21 (6): 513-23. (In Russ.). https://www.doi.org/10.14341/DM9597
Ólafsdóttir AF, Polonsky W, Bolinder J, et al. A randomized clinical trial of the effect of continuous glucose monitoring on nocturnal hypoglycemia, daytime hypoglycemia, glycemic variability, and hypoglycemia confidence in persons with type 1 diabetes treated with multiple daily insulin injections (GOLD-3). Diabetes Technol Ther. 2018; 20 (4): 274-84. https://www.doi.org/10.1089/dia.2017.0363
Divan V, Greenfield M, Morley CP, Weinstock RS. Perceived burdens and benefits associated with continuous glucose monitor use in type 1 diabetes across the lifespan. J Diabetes Sci Technol. 2022; 16 (1): 88-96. https://www.doi.org/10.1177/1932296820978769
Kompala T, Wong J, Neinstein A. Diabetes specialists value continuous glucose monitoring despite challenges in prescribing and data review process. J Diabetes Sci Technol. 2023; 17 (5): 1265-73. https://www.doi.org/10.1177/19322968221088267
Schrangl P, Reiterer F, Heinemann L, et al. Limits to the evaluation of the accuracy of continuous glucose monitoring systems by clinical trials. Biosensors 2018; 8 (2): 50. https://www.doi.org/10.3390/bios8020050
Klimontov VV, Myakina NE. Glucose variability indices predict the episodes of nocturnal hypoglycemia in elderly type 2 diabetic patients treated with insulin. Diabetes Metab Syndr. 2017; 11 (2): 119-24. https://www.doi.org/10.1016/j.dsx.2016.08.023
Sakurai K, Kawai Y, Yamazaki M, Komatsu M. Prediction of lowest nocturnal blood glucose level based on self-monitoring of blood glucose in Japanese patients with type 2 diabetes. J Diabetes Complications 2018; 32 (12): 1118-23. https://www.doi.org/10.1016/j.jdiacomp.2018.09.007
Wang S, Tan Z, Wu T, et al. Largest amplitude of glycemic excursion calculating from self-monitoring blood glucose predicted the episodes of nocturnal asymptomatic hypoglycemia detecting by continuous glucose monitoring in outpatients with type 2 diabetes. Front Endocrinol. 2022; 13: 858912. https://www.doi.org/10.3389/fendo.2022.858912
Fleming GA, Petrie JR, Bergenstal RM, et al. Diabetes digital app technology: Benefits, challenges, and recommendations. A consensus report by the European Association for the Study of Diabetes (EASD) and the American Diabetes Association (ADA) Diabetes Technology Working Group. Diabetologia 2020; 63 (2): 229-41. https://www.doi.org/10.1007/s00125-019-05034-1
Doupis J, Festas G, Tsilivigos C, et al. Smartphone-based technology in diabetes management. Diabetes Ther. 2020; 11 (3): 607-19. https://www.doi.org/10.1007/s13300-020-00768-3
Dedov II, Shestakova MV, Mayorov AYu, et al. Algorithms of Specialized Medical Care for Patients with Diabetes Mellitus. I.I. Dedov, Shestakova M.V., Mayorov A.Yu. (Eds.). 11th edition. Diabetes Mellitus 2023; 26 (2S): 1-157. (In Russ.). https://www.doi.org/10.14341/DM13042
American Diabetes Association Professional Practice Committee. 6. Glycemic Goals and Hypoglycemia: Standards of Care in Diabetes-2024. Diabetes Care 2024; 47: 111-25. https://www.doi.org/10.2337/dc24-S006
Rusanov AN, Rodionova TI. A novel computer algorithm for calculating the mean amplitude of glycemic excursions using self-monitoring blood glucose measurements. Contemporary Issues of Science and Education 2020; (3): 132. (In Russ.) https://www.doi.org/10.17513/spno.29909
Vu L, Kefayati S, Idé T, et al. Predicting nocturnal hypoglycemia from continuous glucose monitoring data with extended prediction horizon. AMIA Annu Symp Proc. 2020: 874-82. https://pubmed.ncbi.nlm.nih.gov/32308884
Mosquera-Lopez C, Dodier R, Tyler NS, et al. Predicting and preventing nocturnal hypoglycemia in type 1 diabetes using big data analytics and decision theoretic analysis. Diabetes Technol Ther. 2020; 22 (11): 801-11. https://www.doi.org/10.1089/dia.2019.0458
Zhang L, Yang L, Zhou Z. Data-based modeling for hypoglycemia prediction: Importance, trends, and implications for clinical practice. Front Public Health 2023; (11): 1044059. https://www.doi.org/10.3389/fpubh.2023.1044059
Rodbard D. Glucose variability: A review of clinical applications and research developments. Diabetes Technol Ther. 2018; 20 (S2): S25-215. https://www.doi.org/10.1089/dia.2018.0092