Novel Detection and Progression Markers for Diabetes Based on Continuous Glucose Monitoring Data Dynamics.

PURPOSE: Develop a diabetes diagnostic tool based on two markers of continuous glucose monitoring (CGM) dynamics: CGM entropy rate (ER) and Poincaré plot (PP) ellipse area (S). METHODS: 5,754 daily CGM profiles from 843 individuals with type 1, type 2 diabetes, or healthy individuals with or without islet autoantibody status were used to compute two individual dynamic markers: ER (in bits per transition; BPT) of daily probability matrices describing CGM transitions between eight glycemic states, and the area S (mg2/dL2) of individual CGM PP ellipses using standard PP descriptors. The Youden's index was used to determine "optimal" cut-points for ER and S for health vs. diabetes (case 1); type 1 vs. type 2 (case 2); and low vs. high type 1 immunological risk (case 3). The markers' discriminative power was assessed through the area under the receiver operating characteristics curves (AUC). RESULTS: Optimal cut-off points were determined for ER and S for each of the three cases. ER and S discriminated case 1 with AUC = 0.98 (95% CI: 0.97-0.99) and AUC = 0.99 (95% CI: 0.99-1.00), respectively, (cut-offs ERcase1 = 0.76 BPT, Scase1 = 1993.91 mg2/dL2), case 2 with AUC = 0.81 (95% CI, 0.77-0.84) and AUC = 0.76 (95% CI, 0.72-0.81), respectively (ERcase2 = 1.00 BPT, Scase2 = 5112.98 mg2/dL2), and case 3 with AUC = 0.81 (95% CI, 0.77-0.84) and AUC = 0.76 (95% CI, 0.72-0.81), respectively (ERcase3 = 0.52 BPT, Scase3 = 923.65 mg2/dL2). CONCLUSIONS: CGM dynamics markers can be an alternative to fasting plasma glucose or glucose tolerance testing and identifying individuals at higher immunological risk of progressing to type 1 diabetes.

Adaptive Biobehavioral Control: A Pilot Analysis of Human-Machine Coadaptation in Type 1 Diabetes.

Background: While it is well recognized that an automated insulin delivery (AID) algorithm should adapt to changes in physiology, it is less understood that the individual would also have to adapt to the AID system. The adaptive biobehavioral control (ABC) method presented here attempts to compensate for this deficiency by including AID into an information cloud-based ecosystem. Methods: The Web Information Tool (WIT) implements the ABC concept via the following: (1) a Physiological Adaptation Module (PAM) that tracks metabolic changes and adapts AID parameters accordingly and (2) a Behavioral Adaptation Module (BAM) that provides information feedback. The safety of WIT (primary outcome) was assessed in an 8-week randomized, two-arm parallel pilot study. All participants used the Control-IQ® AID system enhanced with PAM, but only those in the Experimental group had access to BAM. Secondary glycemic outcomes were computed using the 2-week baseline period and the last 2 weeks of treatment. Results: Thirty participants with type 1 diabetes (T1D) completed all study procedures (17 female/13 male; age: 40 ± 14 years; HbA1c: 6.6% ± 0.5%). No severe hypoglycemia, DKA, or other serious adverse events were reported. Comparing the Experimental and Control groups, no significant difference was observed in time in range (70-180 mg/dL): 74.6% vs 73.8%, adjusted mean difference: 2.65%, 95% CI (-1.12%,6.41%), P = 0.161. Time in 70-140 mg/dL was significantly higher in the Experimental group: 50.7% vs 49.2%, 5.71% (0.44%,10.97%), P = 0.035, without increased time below range: 0.54% (-0.09%,1.17%), P = 0.089. Conclusion: The results demonstrate that it is safe to integrate an AID system into the WIT ecosystem. Validation in a full-scale study is ongoing.

Neural-Net Artificial Pancreas: A Randomized Crossover Trial of a First-in-Class Automated Insulin Delivery Algorithm.

Background: Automated insulin delivery (AID) is now integral to the clinical practice of type 1 diabetes (T1D). The objective of this pilot-feasibility study was to introduce a new regulatory and clinical paradigm-a Neural-Net Artificial Pancreas (NAP)-an encoding of an AID algorithm into a neural network that approximates its action and assess NAP versus the original AID algorithm. Methods: The University of Virginia Model-Predictive Control (UMPC) algorithm was encoded into a neural network, creating its NAP approximation. Seventeen AID users with T1D were recruited and 15 participated in two consecutive 20-h hotel sessions, receiving in random order either NAP or UMPC. Their demographic characteristics were ages 22-68 years old, duration of diabetes 7-58 years, gender 10/5 female/male, White Non-Hispanic/Black 13/2, and baseline glycated hemoglobin 5.4%-8.1%. Results: The time-in-range (TIR) difference between NAP and UMPC, adjusted for entry glucose level, was 1 percentage point, with absolute TIR values of 86% (NAP) and 87% (UMPC). The two algorithms achieved similar times <70 mg/dL of 2.0% versus 1.8% and coefficients of variation of 29.3% (NAP) versus 29.1 (UMPC)%. Under identical inputs, the average absolute insulin-recommendation difference was 0.031 U/h. There were no serious adverse events on either controller. NAP had sixfold lower computational demands than UMPC. Conclusion: In a randomized crossover study, a neural-network encoding of a complex model-predictive control algorithm demonstrated similar performance, at a fraction of the computational demands. Regulatory and clinical doors are therefore open for contemporary machine-learning methods to enter the AID field. Clinical Trial Registration number: NCT05876273.

Artificial Intelligence and Machine Learning for Improving Glycemic Control in Diabetes: Best Practices, Pitfalls, and Opportunities.

OBJECTIVE: Artificial intelligence and machine learning are transforming many fields including medicine. In diabetes, robust biosensing technologies and automated insulin delivery therapies have created a substantial opportunity to improve health. While the number of manuscripts addressing the topic of applying machine learning to diabetes has grown in recent years, there has been a lack of consistency in the methods, metrics, and data used to train and evaluate these algorithms. This manuscript provides consensus guidelines for machine learning practitioners in the field of diabetes, including best practice recommended approaches and warnings about pitfalls to avoid. METHODS: Algorithmic approaches are reviewed and benefits of different algorithms are discussed including importance of clinical accuracy, explainability, interpretability, and personalization. We review the most common features used in machine learning applications in diabetes glucose control and provide an open-source library of functions for calculating features, as well as a framework for specifying data sets using data sheets. A review of current data sets available for training algorithms is provided as well as an online repository of data sources. SIGNIFICANCE: These consensus guidelines are designed to improve performance and translatability of new machine learning algorithms developed in the field of diabetes for engineers and data scientists.

Safety and Efficacy of Sustained Automated Insulin Delivery Compared With Sensor and Pump Therapy in Adults With Type 1 Diabetes at High Risk for Hypoglycemia: A Randomized Controlled Trial.

OBJECTIVE: Assess the safety and efficacy of automated insulin delivery (AID) in adults with type 1 diabetes (T1D) at high risk for hypoglycemia. RESEARCH DESIGN AND METHODS: Participants were 72 adults with T1D who used an insulin pump with Clarke Hypoglycemia Perception Awareness scale score >3 and/or had severe hypoglycemia during the previous 6 months confirmed by time below range (TBR; defined as sensor glucose [SG] reading <70 mg/dL) of at least 5% during 2 weeks of blinded continuous glucose monitoring (CGM). Parallel-arm, randomized trial (2:1) of AID (Tandem t:slim ×2 with Control-IQ technology) versus CGM and pump therapy for 12 weeks. The primary outcome was TBR change from baseline. Secondary outcomes included time in target range (TIR; 70-180 mg/dL), time above range (TAR), mean SG reading, and time with glucose level <54 mg/dL. An optional 12-week extension with AID was offered to all participants. RESULTS: Compared with the sensor and pump (S&P), AID resulted in significant reduction of TBR by -3.7% (95% CI -4.8, -2.6), P < 0.001; an 8.6% increase in TIR (95% CI 5.2, 12.1), P < 0.001; and a -5.3% decrease in TAR (95% CI -87.7, -1.8), P = 0.004. Mean SG reading remained similar in the AID and S&P groups. During the 12-week extension, the effects of AID were sustained in the AID group and reproduced in the S&P group. Two severe hypoglycemic episodes occurred using AID. CONCLUSIONS: In adults with T1D at high risk for hypoglycemia, AID reduced the risk for hypoglycemia more than twofold, as quantified by TBR, while improving TIR and reducing hyperglycemia. Hence, AID is strongly recommended for this specific population.

Developing the UVA/Padova Type 1 Diabetes Simulator: Modeling, Validation, Refinements, and Utility.

Arguably, diabetes mellitus is one of the best quantified human conditions. In the past 50 years, the metabolic monitoring technologies progressed from occasional assessment of average glycemia via HbA1c, through episodic blood glucose readings, to continuous glucose monitoring (CGM) producing data points every few minutes. The high-temporal resolution of CGM data enabled increasingly intensive treatments, from decision support assisting insulin injection or oral medication, to automated closed-loop control, known as the "artificial pancreas." Throughout this progress, mathematical models and computer simulation of the human metabolic system became indispensable for the technological progress of diabetes treatment, enabling every step, from assessment of insulin sensitivity via the now classic Minimal Model of Glucose Kinetics, to in silico trials replacing animal experiments, to automated insulin delivery algorithms. In this review, we follow these developments, beginning with the Minimal Model, which evolved through the years to become large and comprehensive and trigger a paradigm change in the design of diabetes optimization strategies: in 2007, we introduced a sophisticated model of glucose-insulin dynamics and a computer simulator equipped with a "population" of N = 300 in silico "subjects" with type 1 diabetes. In January 2008, in an unprecedented decision, the Food and Drug Administration (FDA) accepted this simulator as a substitute to animal trials for the pre-clinical testing of insulin treatment strategies. This opened the field for rapid and cost-effective development and pre-clinical testing of new treatment approaches, which continues today. Meanwhile, animal experiments for the purpose of designing new insulin treatment algorithms have been abandoned.

Clinically Similar Clusters of Daily Continuous Glucose Monitoring Profiles: Tracking the Progression of Glycemic Control Over Time.

Background: The adoption of continuous glucose monitoring (CGM) results in vast amounts of data, but their interpretation is still more art than exact science. The International Consensus on Time in Range (TIR) proposed the widely accepted TIR system of metrics, which we now take forward by introducing a finite and fixed set of clinically similar clusters (CSCs), such that the TIR metrics of daily CGM profiles within a cluster are homogeneous. Methods: CSC definition and validation used 204,710 daily CGM profiles in health, and types 1 and 2 diabetes (T1D and T2D) on different treatments. The CSCs were defined using 23,916 daily CGM profiles (Training set), and the final fixed set of CSCs was obtained using another 37,758 profiles (Validation set). The Testing set (143,036 profiles) was used to establish the robustness and generalizability of CSCs. Results: The final set of CSCs contains 32 clusters. Any daily CGM profile was classifiable to a single CSC, which approximated common glycemic metrics of the daily CGM profile, as evidenced by regression analyses with 0 intercept (R-squares ≥0.83, e.g., correlation ≥0.91), for all TIR and several other metrics. The CSCs distinguished CGM profiles in health, T2D, and T1D on different treatments, and allowed tracking of daily changes in a person's glycemic control over time. Conclusion: Daily CGM profiles can be classified into one of 32 prefixed CSCs, which enables a host of applications, for example, tabulated data interpretation and algorithmic approaches to treatment, database indexing, pattern recognition, and tracking disease progression. Clinical Trial Registration: N/A-not a clinical trial.

Predicting Immunological Risk for Stage 1 and Stage 2 Diabetes Using a 1-Week CGM Home Test, Nocturnal Glucose Increments, and Standardized Liquid Mixed Meal Breakfasts, with Classification Enhanced by Machine Learning.

Background: Predicting the risk for type 1 diabetes (T1D) is a significant challenge. We use a 1-week continuous glucose monitoring (CGM) home test to characterize differences in glycemia in at-risk healthy individuals based on autoantibody presence and develop a machine-learning technology for CGM-based islet autoantibody classification. Methods: Sixty healthy relatives of people with T1D with mean ± standard deviation age of 23.7 ± 10.7 years, HbA1c of 5.3% ± 0.3%, and body mass index of 23.8 ± 5.6 kg/m2 with zero (n = 21), one (n = 18), and ≥2 (n = 21) autoantibodies were enrolled in an National Institutes of Health TrialNet ancillary study. Participants wore a CGM for a week and consumed three standardized liquid mixed meals (SLMM) instead of three breakfasts. Glycemic outcomes were computed from weekly, overnight (12:00-06:00), and post-SLMM CGM traces, compared across groups, and used in four supervised machine-learning autoantibody status classifiers. Classifiers were evaluated through 10-fold cross-validation using the receiver operating characteristic area under the curve (AUC-ROC) to select the best classification model. Results: Among all computed glycemia metrics, only three were different across the autoantibodies groups: percent time >180 mg/dL (T180) weekly (P = 0.04), overnight CGM incremental AUC (P = 0.005), and T180 for 75 min post-SLMM CGM traces (P = 0.004). Once overnight and post-SLMM features are incorporated in machine-learning classifiers, a linear support vector machine model achieved the best performance of classifying autoantibody positive versus autoantibody negative participants with AUC-ROC ≥0.81. Conclusion: A new technology combining machine learning with a potentially self-administered 1-week CGM home test can help improve T1D risk detection without the need to visit a hospital or use a medical laboratory. Trial registration: ClinicalTrials.gov registration no. NCT02663661.

A Meta-Analysis of Randomized Trial Outcomes for the t:slim X2 Insulin Pump with Control-IQ Technology in Youth and Adults from Age 2 to 72.

Objective: To evaluate the effect of hybrid-closed loop Control-IQ technology (Control-IQ) in randomized controlled trials (RCTs) in subgroups based on baseline characteristics such as race/ethnicity, socioeconomic status (SES), prestudy insulin delivery modality (pump or multiple daily injections), and baseline glycemic control. Methods: Data were pooled and analyzed from 3 RCTs comparing Control-IQ to a Control group using continuous glucose monitoring in 369 participants with type 1 diabetes (T1D) from age 2 to 72 years old. Results: Time in range 70-180 mg/dL (TIR) in the Control-IQ group (n = 256) increased from 57% ± 17% at baseline to 70% ± 11% during follow-up, and in the Control group (n = 113) was 56% ± 15% and 57% ± 14%, respectively (adjusted treatment group difference = 11.5%, 95% confidence interval +9.7% to +13.2%, P < 0.001), an increase of 2.8 h/day on average. Significant reductions in mean glucose, hyperglycemia metrics, hypoglycemic metrics, and HbA1c were also observed. A statistically similar beneficial treatment effect on time in range 70-180 mg/dL was observed across the full age range irrespective of race-ethnicity, household income, prestudy continuous glucose monitor use, or prestudy insulin delivery method. Participants with the highest baseline HbA1c levels showed the greatest improvements in TIR and HbA1c. Conclusion: This pooled analysis of Control-IQ RCTs demonstrates the beneficial effect of Control-IQ in T1D across a broad spectrum of participant characteristics, including racial-ethnic minority, lower SES, lack of prestudy insulin pump experience, and high HbA1c levels. The greatest benefit was observed in participants with the worst baseline glycemic control in whom the auto-bolus feature of the Control-IQ algorithm appears to have substantial impact. Since no subgroups were identified that did not benefit from Control-IQ, hybrid-closed loop technology should be strongly considered for all youth and adults with T1D. Clinical Trials Registry: clinicaltrials.gov; NCT03563313, NCT03844789, and NCT04796779.

Continuous glucose monitoring and metrics for clinical trials: an international consensus statement.

Randomised controlled trials and other prospective clinical studies for novel medical interventions in people with diabetes have traditionally reported HbA1c as the measure of average blood glucose levels for the 3 months preceding the HbA1c test date. The use of this measure highlights the long-established correlation between HbA1c and relative risk of diabetes complications; the change in the measure, before and after the therapeutic intervention, is used by regulators for the approval of medications for diabetes. However, with the increasing use of continuous glucose monitoring (CGM) in clinical practice, prospective clinical studies are also increasingly using CGM devices to collect data and evaluate glucose profiles among study participants, complementing HbA1c findings, and further assess the effects of therapeutic interventions on HbA1c. Data is collected by CGM devices at 1-5 min intervals, which obtains data on glycaemic excursions and periods of asymptomatic hypoglycaemia or hyperglycaemia (ie, details of glycaemic control that are not provided by HbA1c concentrations alone that are measured continuously and can be analysed in daily, weekly, or monthly timeframes). These CGM-derived metrics are the subject of standardised, internationally agreed reporting formats and should, therefore, be considered for use in all clinical studies in diabetes. The purpose of this consensus statement is to recommend the ways CGM data might be used in prospective clinical studies, either as a specified study endpoint or as supportive complementary glucose metrics, to provide clinical information that can be considered by investigators, regulators, companies, clinicians, and individuals with diabetes who are stakeholders in trial outcomes. In this consensus statement, we provide recommendations on how to optimise CGM-derived glucose data collection in clinical studies, including the specific glucose metrics and specific glucose metrics that should be evaluated. These recommendations have been endorsed by the American Association of Clinical Endocrinologists, the American Diabetes Association, the Association of Diabetes Care and Education Specialists, DiabetesIndia, the European Association for the Study of Diabetes, the International Society for Pediatric and Adolescent Diabetes, the Japanese Diabetes Society, and the Juvenile Diabetes Research Foundation. A standardised approach to CGM data collection and reporting in clinical trials will encourage the use of these metrics and enhance the interpretability of CGM data, which could provide useful information other than HbA1c for informing therapeutic and treatment decisions, particularly related to hypoglycaemia, postprandial hyperglycaemia, and glucose variability.

Consensus Recommendations for the Use of Automated Insulin Delivery Technologies in Clinical Practice.

The significant and growing global prevalence of diabetes continues to challenge people with diabetes (PwD), healthcare providers, and payers. While maintaining near-normal glucose levels has been shown to prevent or delay the progression of the long-term complications of diabetes, a significant proportion of PwD are not attaining their glycemic goals. During the past 6 years, we have seen tremendous advances in automated insulin delivery (AID) technologies. Numerous randomized controlled trials and real-world studies have shown that the use of AID systems is safe and effective in helping PwD achieve their long-term glycemic goals while reducing hypoglycemia risk. Thus, AID systems have recently become an integral part of diabetes management. However, recommendations for using AID systems in clinical settings have been lacking. Such guided recommendations are critical for AID success and acceptance. All clinicians working with PwD need to become familiar with the available systems in order to eliminate disparities in diabetes quality of care. This report provides much-needed guidance for clinicians who are interested in utilizing AIDs and presents a comprehensive listing of the evidence payers should consider when determining eligibility criteria for AID insurance coverage.

Biobehavioral Changes Following Transition to Automated Insulin Delivery: A Large Real-life Database Analysis.

OBJECTIVE: To document glycemic and user-initiated bolus changes following transition from predictive low glucose suspend (PLGS) system to automated insulin delivery (AID) system during real-life use. RESEARCH DESIGN AND METHODS: We conducted analysis of 2,329,166 days (6,381 patient-years) of continuous glucose monitoring (CGM) and insulin therapy data for 19,354 individuals with type 1 Diabetes, during 1-month PLGS use (Basal-IQ technology) followed by 3-month AID use (Control-IQ technology). Baseline characteristics are as follows: 55.4% female, age (median/quartiles/range) 39/19-58/1-92 years, mean ± SD glucose management indicator (GMI) 7.5 ± 0.8. Primary outcome was time in target range (TIR) (70-180 mg/dL). Secondary outcomes included CGM-based glycemic control metrics and frequency of user-initiated boluses. RESULTS: Compared with PLGS, AID increased TIR on average from 58.4 to 70.5%. GMI and percent time above and below target range improved as well: from 7.5 to 7.1, 39.9 to 28.1%, and 1.66 to 1.46%, respectively; all P values <0.0001. Stratification of outcomes by age and baseline GMI revealed clinically significant differences. Glycemic improvements were most pronounced in those <18 years old (TIR improvement 14.0 percentage points) and those with baseline GMI >8.0 (TIR improvement 13.2 percentage points). User-initiated correction boluses decreased from 2.7 to 1.8 per day, while user-initiated meal boluses remained stable at 3.6 to 3.8 per day. CONCLUSIONS: Observed in real life of >19,000 individuals with type 1 diabetes, transitions from PLGS to AID resulted in improvement of all glycemic parameters, equivalent to improvements observed in randomized clinical trials, and reduced user-initiated boluses. However, glycemic and behavioral changes with AID use may differ greatly across different demographic and clinical groups.

Essential Continuous Glucose Monitoring Metrics: The Principal Dimensions of Glycemic Control in Diabetes.

Background: With the proliferation of continuous glucose monitoring (CGM), a number of metrics were developed to assess quality of glycemic control. Many of them are highly correlated. Thus, we aim to identify the principal dimensions of glycemic control-a minimal set of metrics, necessary and sufficient for comprehensive assessment of diabetes management. Methods: Seventy-five thousand five hundred sixty-three 2-week CGM profiles recorded in six studies by 790 individuals with type 1 or type 2 diabetes were used to compute mean glucose (MG), percentage time >180 mg/dL (TAR), >250 mg/dL (TAR2), <70 mg/dL (TBR), <54 mg/dL (TBR2), and coefficient of variation (CV). The true dimensionality of the glycemic-metric space was identified in a training set (53,380 profiles) and validated in an independent test set (22,183 profiles). Results: Correlation analysis identified two blocks of metrics-(MG, TAR, TAR2) and (TBR, TBR2, CV)-each with high internal correlation, but insignificant between-block correlation, suggesting that the true dimensionality of the glycemic-metric space is 2. Principal component analysis confirmed two essential metrics quantifying exposure to hyperglycemia (i.e., treatment efficacy) and risk for hypoglycemia (i.e., treatment safety), and explaining ∼90% of the variance in the training and test data. Conclusion: Two essential metrics, treatment efficacy and treatment safety, are necessary and sufficient to characterize glycemic control in diabetes. Thus, quantitatively, diabetes treatment optimization is reduced to a two-dimensional problem, meaning that minimizing both exposure to hyperglycemia and risk for hypoglycemia will lead to improvement in any other metric of glycemic control.

Automated Insulin Delivery with SGLT2i Combination Therapy in Type 1 Diabetes.

Background: Use of sodium-glucose cotransporter 2 inhibitors (SGLT2i) as adjunct therapy to insulin in type 1 diabetes (T1D) has been previously studied. In this study, we present data from the first free-living trial combining low-dose SGLT2i with commercial automated insulin delivery (AID) or predictive low glucose suspend (PLGS) systems. Methods: In an 8-week, randomized, controlled crossover trial, adults with T1D received 5 mg/day empagliflozin (EMPA) or no drug (NOEMPA) as adjunct to insulin therapy. Participants were also randomized to sequential orders of AID (Control-IQ) and PLGS (Basal-IQ) systems for 4 and 2 weeks, respectively. The primary endpoint was percent time-in-range (TIR) 70-180 mg/dL during daytime (7:00-23:00 h) while on AID (NCT04201496). Findings: A total of 39 subjects were enrolled, 35 were randomized, 34 (EMPA; n = 18 and NOEMPA n = 16) were analyzed according to the intention-to-treat principle, and 32 (EMPA; n = 16 and NOEMPA n = 16) completed the trial. On AID, EMPA versus NOEMPA had higher daytime TIR 81% versus 71% with a mean estimated difference of +9.9% (confidence interval [95% CI] 0.6-19.1); p = 0.04. On PLGS, the EMPA versus NOEMPA daytime TIR was 80% versus 63%, mean estimated difference of +16.5% (95% CI 7.3-25.7); p < 0.001. One subject on SGLT2i and AID had one episode of diabetic ketoacidosis with nonfunctioning insulin pump infusion site occlusion contributory. Interpretation: In an 8-week outpatient study, addition of 5 mg daily empagliflozin to commercially available AID or PLGS systems significantly improved daytime glucose control in individuals with T1D, without increased hypoglycemia risk. However, the risk of ketosis and ketoacidosis remains. Therefore, future studies with SGLT2i will need modifications to closed-loop control algorithms to enhance safety.

External validation of a classifier of daily continuous glucose monitoring (CGM) profiles.

As continuous glucose monitoring (CGM) sensors generate ever increasing amounts of CGM data, the need for methods to simplify the storage and analysis of this data becomes increasingly important. Lobo et al. developed a classifier of daily CGM profiles as an initial step in addressing this need. The classifier has several important applications including, but not limited to, data compression, data encryption, and indexing of databases. While the classifier has already successfully classified 99.0% of the 42,595 daily CGM profiles in a Test Set, this work presents an external validation using an external validation set (EVal Set) derived from 8 publicly available data sets. The Test Set and the EVal Set differ in terms of (but not limited to) demographics, data collection time periods, and data collection geographies. The classifier successfully classified 98.2% of the 137,030 daily CGM profiles in the EVal Set. Furthermore, each of the 483 distinct groups of classified daily CGM profiles from the EVal Set retains the same clinical characteristics as the corresponding group from the Test Set, as desired. Finally, the set of unclassified daily CGM profiles from the EVal Set retains the same statistical characteristics as the set of unclassified daily CGM profiles from the Test Set, as desired. These results establish the robustness and generalizability of the classifier: the performance of the classifier is unchanged despite the marked differences between the Test Set and the EVal Set.

A Data-Driven Approach to Classifying Daily Continuous Glucose Monitoring (CGM) Time Series.

According to the World Health Organization, about 422 million people worldwide have type 1 or type 2 diabetes (T1D, T2D), with the latter accounting for 90-95% of cases. Safe and effective treatment of patients with diabetes requires accurate and frequent monitoring of their blood sugar levels. Continuous glucose monitoring (CGM) is a monitoring technology developed to address this need, and its use among U.S. T1D patients has increased from 6% in 2011 to 38% in 2018 and continues to increase worldwide in both T1D and T2D. This paper presents a data-driven approach to determine Ω, a finite set of representative daily profiles (motifs) such that almost any daily CGM profile generated by a patient can be matched to one of the motifs in Ω. The training data set (9741 profiles) was used to identify 8 candidate sets of motifs, while the validation data set (14 175 profiles) was used to select the final set Ω. The robustness of Ω was established by using it to successfully classify (match against a representative daily profile in Ω) 99.0% of 42 595 daily CGM profiles in the testing data set. All data sets contained daily CGM profiles from six studies involving T1D and T2D patients using a variety of treatment modes, including daily insulin injections, insulin pumps, or artificial pancreas (AP). The classified profiles can be used in predictive modeling, decision support, and automated control systems (e.g., AP).