Abstract
Introduction
I
However, more data are not always better. With the increased data volume and complexity, the challenge becomes how to extract information relevant to the condition of a particular patient at a particular time. The classic pathway of medical logic of data → information → decision becomes difficult to follow with the traditional methods of biostatistics. New approaches are needed and are fortunately available, ranging from modeling, simulation, and optimal control methods, to rapidly developing data science tools.
Before continuing further, we should emphasize that, arguably, diabetes mellitus is the best quantified human condition. In the past 40 years, metabolic monitoring technologies have progressed from occasional assessment of average glycemia via glycated hemoglobin (HbA1c), to blood glucose monitoring a few times a day, to continuous glucose monitoring (CGM) producing data points every few minutes—time series tracking the dynamics of the metabolic system. The high temporal resolution of CGM data has enabled advanced treatments such as decision support assisting insulin injection or oral medication, or automated closed-loop control known as the “artificial pancreas.” Sophisticated metabolic models and simulators are available as well.
In this article, we review the progress of data technologies for diabetes from July 1, 2022, to June 30, 2023. We structured the results of this review in three sections: (1) closed-loop control, or automated insulin delivery (AID), which appears to be the term preferred recently; (2) decision-support systems (DSS), particularly those that use contemporary methods such as artificial intelligence, and (3) data acquisition, engineering, analytics, and visualization, which are all data science tools increasingly applied to the retrieval of electronic medical records (EMR) and real-time disease-tracking information.
Key Articles Reviewed
Phillip M, Nimri R, Bergenstal RM, Barnard-Kelly K, Danne T, Hovorka R, Kovatchev BP, Messer LH, Parkin CG, Ambler-Osborn L, Amiel SA, Bally L, Beck RW, Biester S, Biester T, Blanchette JE, Bosi E, Boughton CK, Breton MD, Brown SA, Buckingham BA, Cai A, Carlson AL, Castle JR, Choudhary P, Close KL, Cobelli C, Criego AB, Davis E, de Beaufort C, de Bock MI, DeSalvo DJ, DeVries JH, Dovc K, Doyle FJ 3rd, Ekhlaspour L, Shvalb NF, Forlenza GP, Gallen G, Garg SK, Gershenoff DC, Gonder-Frederick LA, Haidar A, Hartnell S, Heinemann L, Heller S, Hirsch IB, Hood KK, Isaacs D, Klonoff DC, Kordonouri O, Kowalski A, Laffel L, Lawton J, Lal RA, Leelarathna L, Maahs DM, Murphy HR, Nørgaard K, O'Neal D, Oser S, Oser T, Renard E, Riddell MC, Rodbard D, Russell SJ, Schatz DA, Shah VN, Sherr JL, Simonson GD, Wadwa RP, Ward C, Weinzimer SA, Wilmot EG, Battelino T
Burnside MJ, Lewis DM, Crocket HR, Meier RA, Williman JA, Sanders OJ, Jefferies CA, Faherty AM, Paul RG, Lever CS, Price SKJ, Frewen CM, Jones SD, Gunn TC, Lampey C, Wheeler BJ, de Bock MI
Wadwa RP, Reed ZW, Buckingham BA, DeBoer MD, Ekhlaspour L, Forlenza GP, Schoelwer M, Lum J, Kollman C, Beck RW, Breton MD, for the PEDAP Trial Study Group
Bionic Pancreas Research Group; Russell SJ, Beck RW, Damiano ER, El-Khatib FH, Ruedy KJ, Balliro CA, Li Z, Calhoun P, Wadwa RP, Buckingham B, Zhou K, Daniels M, Raskin P, White PC, Lynch J, Pettus J, Hirsch IB, Goland R, Buse JB, Kruger D, Mauras N, Muir A, McGill JB, Cogen F, Weissberg-Benchell J, Sherwood JS, Castellanos LE, Hillard MA, Tuffaha M, Putman MS, Sands MY, Forlenza G, Slover R, Messer LH, Cobry E, Shah VN, Polsky S, Lal R, Ekhlaspour L, Hughes MS, Basina M, Hatipoglu B, Olansky L, Bhangoo A, Forghani N, Kashmiri H, Sutton F, Choudhary A, Penn J, Jafri R, Rayas M, Escaname E, Kerr C, Favela-Prezas R, Boeder S, Trikudanathan S, Williams KM, Leibel N, Kirkman MS, Bergamo K, Klein KR, Dostou JM, Machineni S, Young LA, Diner JC, Bhan A, Jones JK, Benson M, Bird K, Englert K, Permuy J, Cossen K, Felner E, Salam M, Silverstein JM, Adamson S, Cedeno A, Meighan S, Dauber A
Boughton CK, Allen JM, Ware J, Wilinska ME, Hartnell S, Thankamony A, Randell T, Ghatak A, Besser REJ, Elleri D, Trevelyan N, Campbell FM, Sibayan J, Calhoun P, Bailey R, Dunseath G, Hovorka R; CLOuD Consortium
Pichardo-Lowden AR, Haidet P, Umpierrez GE, Lehman EB, Quigley FT, Wang L, Rafferty CM, DeFlitch CJ, Chinchilli VM
Castle JR, Wilson LM, Tyler NS, Espinoza AZ, Mosquera-Lopez CM, Kushner T, Young GM, Pinsonault J, Dodier RH, Hilts WW, Oganessian SM, Branigan DL, Gabo VB, Eom JH, Ramsey K, Youssef JE, Cafazzo JA, Winters-Stone K, Jacobs PG
Unsworth R, Armiger R, Jugnee N, Thomas M, Herrero P, Georgiou P, Oliver N, Reddy M
Shi X, He J, Lin M, Liu C, Yan B, Song H, Wang C, Xiao F, Huang P, Wang L, Li Z, Huang Y, Zhang M, Chen CS, Obst K, Shi L, Li W, Yang S, Yao G, Li X
Sng GGR, Tung JYM, Lim DYZ, Bee YM
Bailey R, Calhoun P, Bergenstal RM, Beck RW
Dunn TC, Xu Y, Bergenstal RM, Ogawa W, Ajjan RA
Schiavon M, Galderisi A, Basu A, Kudva YC, Cengiz E, Dalla Man C
Xu NY, Nguyen KT, DuBord AY, Klonoff DC, Goldman JM, Shah SN, Spanakis EK, Madlock-Brown C, Sarlati S, Rafiq A, Wirth A, Kerr D, Khanna R, Weinstein S, Espinoza J
Zaharieva DP, Senanayake R, Brown C, Watkins B, Loving G, Prahalad P, Ferstad JO, Guestrin C, Fox EB, Maahs DM, Scheinker D, the 4T Research Team
Williams DD, Ferro D, Mullaney C, Skrabonja L, Barnes MS, Patton SR, Lockee B, Tallon EM, Vandervelden CA, Schweisberger C, Mehta S, McDonough R, Lind M, D'Avolio L, Clements MA
Closed-Loop Control of Diabetes, or AID
A PubMed search on artificial pancreas, or AID, or closed loop in diabetes identified 397 results for the period of July 1, 2022, to June 30, 2023. From these, we selected 40 articles of interest (e.g., ∼10%) for inclusion in the reference list of this article. Five of these articles are reviewed in more detail in the following pages (1 –5). It is worth noting that four of them are published in the prestigious New England Journal of Medicine (impact factor, 176.082) (2 –5) and signify substantial clinical trials by different research groups. Publication of these studies in such high-ranking journals was rare until recently, and we believe the AID literature this past year has achieved a new record in terms of its impact on medical science.
A number of other notable studies were published as well (6 –15) that tested various AID systems in diverse conditions, such as the inpatient setting (7), with type 2 diabetes (T2D) (9,10), with long-standing type 1 diabetes (T1D) with hypoglycemia unawareness (12), or in very young children (14). The large randomized trials (> 100 participants) used traditional designs to reconfirm the advantages of hybrid closed loop over conventional therapy (6,15). Meta-analyses (16 –18) and comprehensive reviews (19 –23) solidified AID as an advanced, effective, and cost-effective (24) treatment of T1D and T2D. International consortiums published clinical guidelines for the use of AID in the clinical practice (1,25 –27). And several studies reported real-life data for various AID systems (28 –35), with one study reporting data for nearly 20,000 AID users (28).
The investigation of using ultra-rapid-acting insulin has continued with clinical trials comparing faster insulin aspart with standard insulin aspart (36 –38) and yielding mixed results—in some studies no effect was shown (36), in others ultra-rapid insulin contributed to better control (38). A new study using a sodium-glucose cotransporter 2 (SGLT-2) inhibitor in addition to insulin in people with T1D confirmed previous results and found that “empagliflozin at 2.5 and 5 mg increased time in range during hybrid closed-loop therapy by 11–13 percentage points compared with placebo in those who otherwise were unable to attain glycemic targets” (39). And last, but not least, a study on AID performance during swimming was published, which also presents interesting data about the interdevice (sensor–pump) communication in water (40).
Consensus Recommendations for the Use of Automated Insulin Delivery Technologies in Clinical Practice
Phillip M1,2*, Nimri R1,2*, Bergenstal RM3, Barnard-Kelly K4, Danne T5, Hovorka R6, Kovatchev BP7, Messer LH8, Parkin CG9, Ambler-Osborn L10, Amiel SA11, Bally L12, Beck RW13, Biester S5, Biester T5, Blanchette JE14,15, Bosi E16, Boughton CK17, Breton MD7, Brown SA7,18, Buckingham BA19, Cai A20, Carlson AL3, Castle JR21, Choudhary P22, Close KL20, Cobelli C23, Criego AB3, Davis E24, de Beaufort C25, de Bock MI26, DeSalvo DJ27, DeVries JH28, Dovc K29, Doyle FJ 3rd30, Ekhlaspour L31, Shvalb NF1, Forlenza GP8, Gallen G11, Garg SK8, Gershenoff DC3, Gonder-Frederick LA7, Haidar A32, Hartnell S33, Heinemann L34, Heller S35, Hirsch IB36, Hood KK37, Isaacs D38, Klonoff DC39, Kordonouri O5, Kowalski A40, Laffel L10, Lawton J41, Lal RA42, Leelarathna L43, Maahs DM19, Murphy HR44, Nørgaard K45, O'Neal D46, Oser S47, Oser T47, Renard E48, Riddell MC49, Rodbard D50, Russell SJ51, Schatz DA52, Shah VN8, Sherr JL53, Simonson GD3, Wadwa RP8, Ward C54, Weinzimer SA53, Wilmot EG55,56, Battelino T29
*Contributed equally to the manuscript and are both corresponding authors 1The Jesse Z and Sara Lea Shafer Institute for Endocrinology and Diabetes, National Center for Childhood Diabetes, Schneider Children's Medical Center of Israel, Petah Tikva, Israel; 2Sacker Faculty of Medicine, Tel-Aviv University, Tel-Aviv, Israel; 3International Diabetes Center, HealthPartners Institute, Minneapolis, MN; 4Southern Health NHS Foundation Trust, Southampton, UK; 5AUF DER BULT, Diabetes-Center for Children and Adolescents, Endocrinology and General Paediatrics, Hannover, Germany; 6Wellcome Trust-MRC Institute of Metabolic Science, University of Cambridge, Cambridge, UK; 7Center for Diabetes Technology, School of Medicine, University of Virginia, Charlottesville, VA; 8Barbara Davis Center for Diabetes, University of Colorado Denver—Anschutz Medical Campus, Aurora, CO; 9CGParkin Communications, Inc., Henderson, NV; 10Joslin Diabetes Center, Harvard Medical School, Boston, MA; 11Department of Diabetes, King's College London, London, UK; 12Department of Diabetes, Endocrinology, Nutritional Medicine and Metabolism, Bern University Hospital and University of Bern, Bern, Switzerland; 13Jaeb Center for Health Research Foundation, Inc., Tampa, FL; 14College of Nursing, University of Utah, Salt Lake City, UT; 15Center for Diabetes and Obesity, University Hospitals Cleveland Medical Center, Cleveland, OH; 16Diabetes Research Institute, IRCCS San Raffaele Hospital and San Raffaele Vita Salute University, Milan, Italy; 17Wellcome Trust-MRC Institute of Metabolic Science, Addenbrooke's Hospital, University of Cambridge Metabolic Research Laboratories, Cambridge, UK; 18Division of Endocrinology, University of Virginia, Charlottesville, VA; 19Division of Endocrinology, Department of Pediatrics, Stanford University, School of Medicine, Stanford, CA; 20The diaTribe Foundation/Close Concerns, San Diego, CA; 21Harold Schnitzer Diabetes Health Center, Oregon Health & Science University, Portland, OR; 22Diabetes Research Centre, University of Leicester, Leicester, UK; 23Department of Woman and Child's Health, University of Padova, Padova, Italy; 24Telethon Kids Institute, University of Western Australia, Perth Children's Hospital, Perth, Australia; 25Diabetes & Endocrine Care Clinique Pédiatrique DECCP/Centre Hospitalier Luxembourg, and Faculty of Sciences, Technology and Medicine, University of Luxembourg, Esch sur Alzette, GD Luxembourg/Department of Paediatrics, UZ-VUB, Brussels, Belgium; 26Department of Paediatrics, University of Otago, Christchurch, New Zealand; 27Division of Pediatric Diabetes and Endocrinology, Baylor College of Medicine, Texas Children's Hospital, Houston, TX ; 28Amsterdam UMC, University of Amsterdam, Internal Medicine, Amsterdam, The Netherlands; 29Department of Pediatric Endocrinology, Diabetes and Metabolic Diseases, UMC - University Children's Hospital, Ljubljana, Slovenia, and Faculty of Medicine, University of Ljubljana, Ljubljana, Slovenia; 30Harvard John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA; 31Lucile Packard Children's Hospital—Pediatric Endocrinology, Stanford University School of Medicine, Palo Alto, CA; 32Department of Biomedical Engineering, McGill University, Montreal, Canada; 33Wolfson Diabetes and Endocrine Clinic, Cambridge University Hospitals NHS Foundation Trust, Cambridge, UK; 34Science Consulting in Diabetes GmbH, Kaarst, Germany; 35Department of Oncology and Metabolism, University of Sheffield, Sheffield, UK; 36Department of Medicine, University of Washington Diabetes Institute, University of Washington, Seattle, WA; 37Stanford Diabetes Research Center, Stanford University School of Medicine, Stanford, CA; 38Cleveland Clinic, Endocrinology and Metabolism Institute, Cleveland, OH; 39Diabetes Research Institute, Mills-Peninsula Medical Center, San Mateo, CA; 40JDRF International, New York, NY; 41Usher Institute, University of Edinburgh, Edinburgh, UK; 42Division of Endocrinology, Department of Pediatrics, Stanford University School of Medicine, Stanford, CA; 43Manchester University Hospitals NHS Foundation Trust/University of Manchester, Manchester, UK; 44Norwich Medical School, University of East Anglia, Norwich, UK; 45Steno Diabetes Center Copenhagen and Department of Clinical Medicine, University of Copenhagen, Gentofte, Denmark; 46Department of Medicine and Department of Endocrinology, St Vincent's Hospital Melbourne, University of Melbourne, Melbourne, Australia; 47Department of Family Medicine, University of Colorado School of Medicine, Anschutz Medical Campus, Aurora, CO; 48Department of Endocrinology, Diabetes, Nutrition, Montpellier University Hospital, and Institute of Functional Genomics, University of Montpellier, CNRS, INSERM, Montpellier, France; 49School of Kinesiology & Health Science, Muscle Health Research Centre, York University, Toronto, Canada; 50Biomedical Informatics Consultants LLC, Potomac, MD; 51Massachusetts General Hospital and Harvard Medical School, Boston, MA; 52Department of Pediatrics, College of Medicine, Diabetes Institute, University of Florida, Gainesville, FL; 53Department of Pediatrics, Yale University School of Medicine, Pediatric Endocrinology, New Haven, CT; 54Institute of Metabolic Science, Cambridge University Hospitals NHS Foundation Trust, Cambridge, UK; 55Department of Diabetes & Endocrinology, University Hospitals of Derby and Burton NHS Trust, Derby, UK; 56Division of Medical Sciences and Graduate Entry Medicine, University of Nottingham, Nottingham, England, UK
The past 6 years have produced 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 individuals with diabetes achieve long-term glycemic goals while reducing hypoglycemia risk. AID systems have thus become an integral part of diabetes management, but recommendations for using AID systems in clinical settings have been lacking.
Methods
In 2021, an international panel of clinicians, researchers, and patient advocates with expertise in AID was organized by the Advanced Technologies & Treatments for Diabetes (ATTD) Congress to develop clinical guidelines for initiating AID for individuals with type 1 diabetes (T1D). The nine working groups addressed evolution of AID, clinical evidence, determining the target population for AID use, initiation of AID, education and training, utilization of AID, AID data reporting, psychological issues/user perspective, and the future of AID. The full panel voted on the working group recommendations, which became the basis of the consensus recommendations.
Results
The report provides needed guidance to clinicians who are interested in using AID and (2) serves as a comprehensive review of evidence for payers to consider when determining eligibility criteria for AID insurance coverage.
Conclusions
Guided recommendations are critical for AID success and acceptance, and all clinicians who treat diabetes need to become familiar with the available systems to eliminate disparities in diabetes quality of care. A comprehensive listing of the evidence payers should consider when determining eligibility criteria for AID insurance coverage is also provided.
Comments
These consensus recommendations published in Endocrine Reviews present the collective opinion of a select group of experts in AID (aka closed-loop control of diabetes) regarding the use of AID systems in clinical practice. This article is a highly recommended reading for anyone who wants to be introduced to the current technological and clinical state of AID systems, with a peek into their future as well.
Following a review of the closed-loop field and an account of the evidence-based advantages of AID systems over other T1D therapies, the guidelines continue with sections explaining the different types of AID algorithms and the difference between hybrid and fully automated AID systems (HCL and FCL), the latter being defined as an AID system that does not require user involvement to function.
The following are major takeaways from these consensus recommendations. (1) Evidence suggests that AID systems should be considered as a treatment option for all people with T1D to improve glycemic control, regardless of age, hypoglycemia awareness, pregnancy, or certain comorbidities. (2) To date, all clinically available systems are “hybrid,” meaning they require specific diabetes managements skills such as carbohydrate counting, so training and support for users and health-care providers are essential. (3) Interoperability between system components (e.g., sensors, algorithms, and insulin pumps) is important for the proliferation of these new technologies, but to a large extent this depends on the industry. (4) Early initiation of diabetes technologies in newly diagnosed T1D has been shown to improve and sustain long-term glycemic control, so early AID system initiation is encouraged. (5) The transition to an AID system should be individualized, and the initial settings of an AID system should be selected according to personal glycemic targets, based on recently acquired continuous glucose monitor (CGM) metrics. (6) Unified reporting of AID data is suggested, using clinically important glucose (e.g., time-in-range) and insulin (e.g., total daily insulin) metrics, plus the ambulatory glucose profile (AGP) usually aggregated over 14 days, which has become a standardized way to visualize CGM data. (7) Psychological issues related to the use of AID systems should be discussed and documented in the context of discontinuation or noncompliance with AID treatment.
In conclusion, the consensus recommendations map future directions for AID development, including FCL, use of data science methods to analyze and comprehend the vast amounts of CGM and insulin-delivery data collected by AID systems, and learning adaptive AID algorithms that continually tailor the treatment to the changing behavior and physiology of their users. The consensus strongly recommends that all payers (government and private) should reimburse for AID systems as well as for AID education and training to support the management of people with T1D.
Open-Source Automated Insulin Delivery in Type 1 Diabetes
Burnside MJ1,3, Lewis DM11, Crocket HR4, Meier RA1, Williman JA2, Sanders OJ1,3, Jefferies CA6,7, Faherty AM6, Paul RG5, Lever CS5, Price SKJ5, Frewen CM8, Jones SD8, Gunn TC10, Lampey C6, Wheeler BJ8,9, de Bock MI1,3
1Departments of Pediatrics, University of Otago, Dunedin, New Zealand; 2Department of Population Health, University of Otago, Dunedin, New Zealand; 3Department of Pediatrics, Canterbury District Health Board Christchurch, New Zealand; 4Te Huataki Waiora School of Health, Sport and Human Performance, University of Waikato, Hamilton, New Zealand; 5Waikato Regional Diabetes Service, Waikato District Health Board, Hamilton, New Zealand; 6Department of Pediatric Endocrinology, Starship Children's Health, Auckland District Health Board, Auckland, New Zealand; 7Liggins Institute, University of Auckland, Auckland, New Zealand; 8Department of Women's and Children's Health, Dunedin School of Medicine, University of Otago, Dunedin, New Zealand; 9Pediatric Department, Southern District Health Board, Dunedin, New Zealand; 10Nightscout New Zealand, Hamilton, New Zealand; 11OpenAPS, Seattle, WA
This study is also discussed in DIA-2024-2508, page S-117.
People with type 1 diabetes (T1D) use open-source automated insulin delivery (AID) systems, and more data are needed on the efficacy and safety of these system.
Methods
In this multicenter, open-label, randomized, controlled trial, patients with T1D were assigned in a 1:1 ratio to use an open-source AID system or a sensor-augmented insulin pump (control). The participants included both children (defined as 7–15 years of age) and adults (defined as 16–70 years of age). A modified version of AndroidAPS 2.8 (with a standard OpenAPS 0.7.0 algorithm) paired with a preproduction DANA-i insulin pump and Dexcom G6 CGM, which has an Android smartphone application as the user interface, was used as the AID system. The primary outcome was the percentage of time in the target glucose range of 70–180 mg/dL (3.9–10.0 mmol/L) between days 155 and 168, the final 2 weeks of the trial.
Results
A total of 97 patients (48 children and 49 adults) underwent randomization, 44 to open-source AID and 53 to the control group. At 24 weeks, the mean time in the target range increased from 61.2% ± 12.3% SD to 71.2% ± 12.1% SD in the AID group and decreased from 57.7% ± 14.3% SD to 54.5% ± 16.0% SD in the control group (adjusted difference, 14 percentage points [95% CI, 9.2–18.8], P < 0.001), with no treatment effect according to age (P = 0.56). Participants in the AID group spent 3 hours and 21 minutes more in the target range per day compared with in the control group. There were no episodes of severe hypoglycemia or diabetic ketoacidosis in either group. Two patients in the AID group withdrew from the trial owing to connectivity issues.
Conclusions
The use of an open-source AID system in children and adults with T1D resulted in a significantly higher percentage of time in the target glucose range compared with the use of a sensor-augmented insulin pump.
Comments
Open-source AID (or Do-it-yourself Loop) algorithms have been available since before commercial AID systems appeared on the market. Typically, the handling of these open-source algorithms required above-average engineering and computer skills. The lack of clinical trials of open-source AID was also a reason these systems to remain unregulated (i.e., not formally approved by the U.S. Food and Drug Administration [FDA] or other regulatory agencies). This article in the prestigious New England Journal of Medicine aims to counter these shortcomings by presenting a randomized-controlled trial of one open-source algorithm running on a smartphone.
The trial achieved results that are very similar to those in trials published to date for commercial AID systems: for instance, a 14-percentage-point increase in time in range up to 71%, a reduction of HbA1c of 0.5 percentage points, and a relatively low risk for hypoglycemia. The editorial that accompanied the article labeled the trial a “path toward expanding treatment options for diabetes” (41). As expected there were opposing opinions that emphasized the lack of regulatory approval of open-source AID: “Unapproved products and algorithms raise the undesirable prospect of clinical harm and litigation” (42). After the publication of this article, the latter argument became less critical owing to the FDA approval in January 2023 of Tidepool Loop—an FDA-regulated version of Loop, to be available in the iOS App Store, which is intended to work with commercially available insulin pumps and CGMs.
This ATTD Yearbook article takes a neutral view on the open-source AID controversy: evidence has shown that these systems are not superior to commercial devices, and their more flexible customization features do not necessarily result in improved glycemic control. However, as noted in the previously mentioned editorial (41), open-source AID offers one more path toward expanding the treatment options for insulin-requiring diabetes, which can be appealing to a perhaps select group of people who are confident in their diabetes-management skills.
Trial of Hybrid Closed-Loop Control in Young Children with Type 1 Diabetes
Wadwa RP1, Reed ZW2, Buckingham BA3, DeBoer MD5, Ekhlaspour L4, Forlenza GP1, Schoelwer M5, Lum J2, Kollman C2, Beck RW2, Breton MD5, for the PEDAP Trial Study Group
1Barbara Davis Center for Diabetes, University of Colorado, Anschutz Medical Campus, Aurora, CO; 2Jaeb Center for Health Research, Tampa, FL; 3Department of Pediatrics, Division of Pediatric Endocrinology and Diabetes, Stanford University School of Medicine, Stanford, CA; 4Division of Pediatric Endocrinology, University of California, San Francisco, CA; 5University of Virginia Center for Diabetes Technology, Charlottesville, VA
This study is also discussed in DIA-2024-2508, page S-117.
Closed-loop control systems of insulin delivery may improve glycemic outcomes in young children with type 1 diabetes (T1D), but the efficacy and safety of initiating a closed-loop system virtually have not been determined.
Methods
In this 13-week, multicenter trial, children who were at least 2 years of age but younger than 6 years of age who had T1D were randomly assigned in a 2:1 ratio to receive treatment with a closed-loop system of insulin delivery or standard care, which included either an insulin pump or multiple daily injections of insulin plus a continuous glucose monitor (CGM). The primary outcome was the percentage of time that the glucose level was in the target range of 70–180 mg/dL as measured by CGM. The secondary outcomes included the percentage of time that the glucose level was > 250 mg/dL or < 70 mg/dL, the mean glucose level, the glycated hemoglobin level (HbA1c), and safety outcomes.
Results
In randomization, a total of 102 children were assigned to two groups: 68 to the closed-loop group and 34 to the standard-care group. The HbA1c levels at baseline ranged from 5.2% to 11.5%. Initiation of the closed-loop system was virtual in 55 patients (81%). The mean percentage of time that the glucose level was within the target range increased from 56.7% ± 18.0% SD at baseline to 69.3% ± 11.1% SD during the 13-week follow-up period in the closed-loop group and from 54.9% ± 14.7% SD to 55.9% ± 12.6% SD in the standard-care group: mean adjusted difference, 12.4 percentage points (equivalent to approximately 3 hours per day [95% CI, 9.5–15.3], P < 0.001). Similar treatment effects were observed (favoring the closed-loop system) for the percentage of time that the glucose level was > 250 mg/dL, for the mean glucose level, and for the HbA1c level with no significant between-group difference in the percentage of time that the glucose level was < 70 mg/dL. The closed-loop group had two cases of severe hypoglycemia; the standard-care group had one case. One case of diabetic ketoacidosis occurred in the closed-loop group.
Conclusions
For young children with T1D, the glucose level was in the target range for a greater percentage of time with a closed-loop system than with standard care.
Comments
This study is the third in a series of publications in the New England Journal of Medicine (43,44) reporting the outcomes of a pivotal trial of the Control-IQ AID system (Tandem Diabetes Care) in children aged 2 to 6 years old. The results of this study confirmed, and were similar to, the findings in children and adults using this same system, such as the 12.4 percentage points increase in time in range, the 0.42 percentage points reduction in HbA1c after 13 weeks of system use, and the low exposure to hypoglycemia (3% or less time below 70 mg/dL). Wadwa and colleagues concluded that this particular AID system is a viable treatment option for children with T1D as young as 2 years of age.
This study's conclusion supports further the consensus recommendations for the use of AID technologies in clinical practice (1), which suggested the use of AID regardless of age. More recently, a meta-analysis of the three published randomized controlled trials of Control-IQ (3,43,44) across all studied age groups (ages 2–72 years old) concluded, “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” (16). This conclusion was also supported by analysis of real-life data for nearly 20,000 people with T1D aged 1 to 92 years old (28). Thus, this study appears in the context of expanding use of AID to progressively younger age groups, confirming AID utility early in the life with T1D.
Multicenter, Randomized Trial of a Bionic Pancreas in Type 1 Diabetes
Bionic Pancreas Research Group, Russell SJ1, Beck RW4, Damiano ER2,3, El-Khatib FH3, Ruedy KJ4, Balliro CA1, Li Z4, Calhoun P4, Wadwa RP6, Buckingham B7, Zhou K10, Daniels M8, Raskin P11, White PC11, Lynch J12, Pettus J9, Hirsch IB13, Goland R14, Buse JB15, Kruger D16, Mauras N5, Muir A17, McGill JB18, Cogen F19, Weissberg-Benchell J20, Sherwood JS1, Castellanos LE1, Hillard MA1, Tuffaha M1, Putman MS1, Sands MY1, Forlenza G6, Slover R6, Messer LH6, Cobry E6, Shah VN6, Polsky S6, Lal R7, Ekhlaspour L7, Hughes MS7, Basina M7, Hatipoglu B10, Olansky L10, Bhangoo A8, Forghani N8, Kashmiri H8, Sutton F8, Choudhary A11, Penn J11, Jafri R12, Rayas M12, Escaname E12, Kerr C12, Favela-Prezas R12, Boeder S9, Trikudanathan S13, Williams KM14, Leibel N14, Kirkman MS15, Bergamo K15, Klein KR15, Dostou JM15, Machineni S15, Young LA15, Diner JC15, Bhan A16, Jones JK16, Benson M5, Bird K5, Englert K5, Permuy J5, Cossen K17, Felner E17, Salam M18, Silverstein JM18, Adamson S18, Cedeno A18, Meighan S19, Dauber A19
1Diabetes Research Center, Massachusetts General Hospital, Boston, MA; 2Boston University, Boston, MA; 3Beta Bionics, Concord, MA; 4Jaeb Center for Health Research, Tampa, FL; 5Nemours Children's Health Jacksonville, Jacksonville, FL; 6Barbara Davis Center for Diabetes, University of Colorado, Aurora, CO; 7Stanford University School of Medicine, Palo Alto, CA; 8Children's Hospital of Orange County, Orange, CA; 9University of California, San Diego, La Jolla, CA; 10Cleveland Clinic, Cleveland, OH; 11University of Texas Southwestern Medical Center, Dallas, TX; 12University of Texas Health Science Center, San Antonio, TX; 13University of Washington, Seattle, WA; 14Naomi Berrie Diabetes Center, Columbia University, New York, NY; 15University of North Carolina, Chapel Hill, NC; 16Henry Ford Health System, Detroit, MI; 17Emory University, Atlanta, GA; 18Washington University in St. Louis, St. Louis, MO; 19Children's National Hospital, Washington, DC; 20Pritzker Department of Psychiatry and Behavioral Health, Ann and Robert Lurie Children's Hospital, Chicago, IL
This study is also discussed in DIA-2024-2508, page S-117.
For routine operation, available semiautomated insulin-delivery systems require individualized insulin regimens for the initialization of therapy and meal doses based on carbohydrate counting. By contrast, the bionic pancreas is initialized with only body weight; it makes all dose decisions, delivers insulin autonomously, and uses meal announcements without carbohydrate counting.
Methods
This 13-week, multicenter, randomized trial with persons 6 years of age and older with type 1 diabetes (T1D) randomly assigned the participants in a 2:1 ratio either to receive bionic pancreas treatment with either insulin aspart or insulin lispro, or to receive standard care (any insulin-delivery method with unmasked, real-time continuous glucose monitoring [CGM]). The primary outcome was the glycated hemoglobin (HbA1c) level at 13 weeks. The key secondary outcome was the percentage of time that the CGM-assessed glucose level was below 54 mg/dL (the prespecified noninferiority limit for this outcome was 1 percentage point), and safety was also assessed.
Results
A total of 219 participants 6–79 years of age were assigned to the bionic pancreas group and 107 to the standard care group. In the bionic pancreas group, the HbA1c level decreased from 7.9% to 7.3%; in the standard care group the level did not change (7.7% at both time points) (mean adjusted difference at 13 weeks: −0.5 percentage points [95% CI, −0.6 to −0.3], P < 0.001). The percentage of time that the CGM-assessed glucose level was below 54 mg/dL did not differ significantly between the two groups (13-week adjusted difference: 0.0 percentage points [95% CI, −0.1 to 0.04], P < 0.001 for noninferiority). The rate of severe hypoglycemia was 17.7 events per 100 participant-years in the bionic-pancreas group and 10.8 events per 100 participant-years in the standard-care group (P = 0.39). No episodes of diabetic ketoacidosis occurred in either group.
Conclusions
Use of a bionic pancreas was associated with a greater reduction in HbA1c compared with standard care in this 13-week, randomized trial involving adults and children with T1D.
Comments
The bionic pancreas was originally designed as a dual-hormone system using insulin and glucagon to control diabetes and counter the risk for hypoglycemia. While studies of the bihormonal configuration of this system continue, with some previously reviewed in the Yearbook (45), this year the Bionic Pancreas Research Group offers three publications with an insulin-only system (4,13,38). The systems pivotal trial, published in the New England Journal of Medicine, was reviewed here (4), an extension study was published as well (13), and a multicenter trial evaluated fast-acting insulin aspart in the bionic pancreas (38).
Over 300 (N = 326) participants with T1D aged 6 to 79 years old were assigned in a ratio 2:1 to the bionic pancreas group or standard care, making this the largest randomized controlled trial of any AID system to date. The primary outcome, Hb1Ac, was reduced by 0.5 percentage points from 7.9% to 7.3% in the bionic pancreas group, which is a typical change observed 3 months after initiation of AID use. The absolute final time in range in the bionic pancreas group was 66%, and the relative improvement was 11 percentage points, which is also consistent with other AID studies. The one difference between this trial and other published studies is the rate of severe hypoglycemia: this study reported 17.7 events per 100 participant-years in the bionic pancreas group and 10.8 events per 100 participant-years in the standard-care group; the previously discussed meta-analysis with a comparable sample size (N = 369 participants, ages 2–72) reported several-fold lower rates: 2.1 events per 100 person-years in the Control-IQ group versus 2.3 events per 100 person-years in those on sensor-augmented insulin pumps, accompanied by a lower final HbA1c of 7.0% in the Control-IQ group (16).
Although comparisons across studies may be generally unfair, a several-fold difference in a critical objective marker such as severe hypoglycemia, without the more intensive insulin treatment on the bionic pancreas, deserves at least a hypothesis. We can speculate that, while the control groups of these studies had different compositions (standard care vs sensor-augmented pump) and are therefore not comparable, the difference between the experimental groups may be explained by the control algorithm designs. An algorithm originally intended to work with insulin and glucagon was perhaps less aggressive in dealing with hypoglycemia when left with its insulin-only action, than a system designed from the ground up for insulin-only use.
Closed-Loop Therapy and Preservation of C-peptide Secretion in Type 1 Diabetes
Boughton CK1,3, Allen JM1,2, Ware J1,2, Wilinska ME1,2, Hartnell S3, Thankamony A2, Randell T4, Ghatak A5, Besser REJ6, Elleri D7, Trevelyan N8, Campbell FM9, Sibayan J11, Calhoun P11, Bailey R11, Dunseath G10, Hovorka R1,2, CLOuD Consortium
1Wellcome–Medical Research Council Institute of Metabolic Science, University of Cambridge, UK; 2Department of Paediatrics, University of Cambridge, UK; 3Wolfson Diabetes and Endocrine Clinic, Cambridge University Hospitals NHS Foundation Trust, Cambridge, UK; 4Paediatric Diabetes and Endocrinology, Nottingham Children's Hospital, Nottingham, UK; 5Department of Diabetes, Alder Hey Children's NHS Foundation Trust, Liverpool, UK; 6Department of Paediatrics, University of Oxford, and the National Institute for Health and Care Research Oxford Biomedical Research Centre, John Radcliffe Hospital, Oxford, UK; 7Department of Diabetes, Royal Hospital for Sick Children, Edinburgh, UK; 8Department of Paediatric Diabetes, Southampton Children's Hospital, Southampton, UK; 9Department of Paediatric Diabetes, Leeds Children's Hospital, Leeds, UK; 10Diabetes Research Group, Swansea University, Swansea, UK; 11Jaeb Center for Health Research, Tampa, FL
This study examined whether improved glucose control with hybrid closed-loop therapy can preserve C-peptide secretion as compared with standard insulin therapy in persons with new-onset type 1 diabetes (T1D).
Methods
In a multicenter, open-label, parallel-group, randomized trial, youths aged 10.0–16.9 years who were within 21 days of a T1D diagnosis were randomized to receive hybrid closed-loop therapy or standard insulin therapy (control) for 24 months. The primary end point was the area under the curve (AUC) for the plasma C-peptide level (after a mixed-meal tolerance test) at 12 months after diagnosis. The analysis was performed on an intention-to-treat basis.
Results
A total of 97 participants (mean age, 12 ± 2 SD years) were randomized to two groups: 51 received closed-loop therapy and 46 received the control therapy (standard insulin). The AUC for the C-peptide level at 12 months (primary end point) did not significantly differ between the two groups (with closed-loop therapy: geometric mean, 0.35 pmol/mL [IQR, 0.16–0.49] vs control: 0.46 pmol/mL [IQR, 0.22–0.69]; mean adjusted difference, −0.06 pmol/mL [95% CI, −0.14 to 0.03]). There was no substantial difference between the groups in the AUC for the C-peptide level at 24 months (closed-loop therapy: geometric mean, 0.18 pmol/mL [IQR, 0.06–0.22]; vs control: 0.24 pmol/mL [IQR, 0.05–0.30]; mean adjusted difference, −0.04 pmol/mL [95% CI, −0.14 to 0.06]). The arithmetic mean glycated hemoglobin level was lower in the closed-loop group than in the control group by 4 mmol/mol (0.4 percentage points [95% CI, 0–8 mmol/mol], 0.0–0.7 percentage points) at 12 months and by 11 mmol/mol (1.0 percentage points [95% CI, 7–15 mmol/mol], 0.5–1.5 percentage points) at 24 months. Five cases of severe hypoglycemia occurred in the closed-loop group in three participants, and one case occurred in the control group. One case of diabetic ketoacidosis occurred in the closed-loop group.
Conclusions
In youths with new-onset T1D, intensive glucose control for 24 months did not appear to prevent the decline in residual C-peptide secretion.
Comments
To the best of our knowledge, this is a unique large, long-term (24 months) study attempting to clarify whether intensive AID treatment in youths with new-onset T1D can preserve β-cell function longer than in a parallel control group. This question has been studied before, with inconsistent results. This new trial postulates that the answer is “no,” so a more pragmatic title for this article might have been “Closed-Loop Therapy Does Not Preserve C-Peptide Secretion in Type 1 Diabetes.” The negative result is instructive and appears definitive because the control group ended up with (nonsignificantly) higher C-peptide levels than the AID group. Thus, it is unlikely that this finding would be reversed by a larger sample size with any of the current AID systems.
However, given the glycemic outcomes of the study, some questions remain. (1) By 12 and 24 months the closed-loop group had a time in range of 64%, which was 10 percentage points lower than their baseline. (2) Simultaneously, their time below 70 mg/dL increased by 2.8 percentage points and remained high at 11.2% at the end of the trial. (3) There were five severe hypoglycemic events (incidence of 5.4 events per 100 patient-years) in the closed-loop group. It is understandable that in new-onset diabetes the glycemic control will deteriorate from baseline to 12 and 24 months, but the end result—particularly the observed exposure to hypoglycemia—is far from optimal. This prompted the authors to speculate in the discussion that “the greater mean time below the target glucose range and greater mean glycemic coefficient of variation observed in the closed-loop group may have reduced beta-cell viability.” Thus, the case is not closed: better control with better future AID systems may reverse the findings of this study.
Artificial Intelligence-Based Decision Support Systems
A PubMed search using the keywords decision support system (DSS), artificial intelligence (AI), or dose treatment recommendations in diabetes yielded 341 results for the period between July 1, 2022, and June 30, 2023. Among these, we chose five papers that represent five distinct topics to be reviewed in this article. Some others are referenced in the introduction below.
This year the introduction of AI-driven chatbots like ChatGPT has significantly propelled the domain of machines performing human tasks. The incorporation of these technologies has the potential to accelerate the application of AI tools to enhance medical care for people with diabetes. This trend is reflected by the increased number of publications related to that topic. AI-based DSS can serve as a valuable tool in diabetes management, especially in primary care settings where limited resources are available.
A review of DSS publications from the last four decades showed that 77% of the systems were knowledge based. Notably, non-knowledge-based systems have been on the rise in recent years, particularly in the context of treatment recommendations which were the systems most often used (46). Consequently, advanced DSS platforms are sought after, with improved functionalities to increase effectiveness. This year the publications follow this path: DSS based on AI technologies such as machine learning and case-based reasoning utilizing wearable devices were evaluated (47).
Another technology that pushes the DSS field is the growing use of CGM enabling personalized recommendations, not only for insulin or medication treatment but also for dietary guidance and lifestyle suggestions tailored to an individual's unique physiological responses and preferences. This year a study published in diabetes care by Lee et al. (48) showed that an integrated digital health platform with AI-based dietary management (using AI-based food recognition from a single photograph) in adults with T2D resulted in better glycemia and more weight loss. Several smart bolus calculators using CGM data and glucose trends are reviewed in this article. Another growing field of AI-based DSS in diabetes is the screening of diabetes-related complications such as foot ulcers and retinopathy (49) through advanced image recognition and diagnostic algorithms.
We can summarize this year's data with the scoping review presented at the last American Diabetes Association meeting 2023, which revealed that the majority of DSS platforms significantly improved the outcomes for individuals with diabetes (46). Nevertheless, there is a need to understand how to effectively implement these tools in clinical care. While significant advances are occurring in the field of AI, the ethical and regulatory considerations still require attention and answers.
Clinical Decision Support for Glycemic Management Reduces Hospital Length of Stay
Pichardo-Lowden AR1, Haidet P1,2,3, Umpierrez GE4, Lehman EB2, Quigley FT5, Wang L2, Rafferty CM1, DeFlitch CJ6, Chinchilli VM2
1Department of Medicine, Penn State Health, Penn State College of Medicine, Hershey Medical Center, Hershey, PA; 2Department of Public Health Sciences, Penn State College of Medicine, Hershey, PA; 3Department of Humanities and the Woodward Center for Excellence in Health Sciences Education, Penn State College of Medicine, Hershey, PA; 4Department of Medicine, Emory University, Atlanta, GA; 5Department of Medicine, Penn State Health St. Joseph Medical Center, Reading, PA; 6Department of Emergency Medicine, Office of the Chief Medical Information Officer, Penn State Health, Hershey, PA
Clinical decision support (CDS) holds promise for optimizing care by overcoming management barriers in the treatment of dysglycemia. This study assessed the impact on hospital length of stay (LOS) of an alert-based CDS tool in the electronic medical record that detected dysglycemia or inappropriate insulin use, coined as gaps in care (GIC).
Methods
In a 12-month interrupted time series among hospitalized persons aged 18 years, our CDS tool identified GIC and, when active, provided recommendations. We compared LOS during 6-month-long active and inactive periods using linear models for repeated measures, multiple comparison adjustment, and mediation analysis.
Results
Among 4788 admissions with GIC, the average LOS was shorter during the tool's active periods. The LOS reductions occurred for all admissions with GIC (25.7 hours, P = 0.057), diabetes and hyperglycemia (26.4 hours, P = 0.054), stress hyperglycemia (231.0 hours, P = 0.054), patients admitted to medical services (28.4 hours, P = 0.039), and recurrent hypoglycemia (229.1 hours, P = 0.074). The subgroup analysis showed significantly shorter LOS in recurrent hypoglycemia with three events (282.3 hours, P = 0.006) and nonsignificant in two (25.2 hours, P = 0.655) and four or more (214.8 hours, P = 0.746). Among 22,395 admissions with GIC (4788 [21%]) and without GIC (17,607 [79%]), the LOS reduction during the active period was 1.8 hours (P = 0.053). When the active tool provided recommendations, it indirectly and significantly contributed to shortening LOS through its influence on GIC events during admissions with at least one GIC (P = 0.027), diabetes and hyperglycemia (P = 0.028), and medical services (P = 0.019).
Conclusions
Use of the alert-based CDS tool to address inpatient management of dysglycemia contributed to reducing LOS, which may reduce costs and improve patient well-being.
Comments
In theory, clinical DSS offer a promising approach to enhance workflow within health-care systems, which we all know are facing resource constraints. Although studies have demonstrated comparable and enhanced glycemic outcomes with DSS system use compared with health-care providers’ decisions, limited data exist regarding the measurable gains in terms of time and cost savings alongside the delivery of efficient health care. This study provides evidence that a clinical decision-support tool could impact the duration of hospitalization for patients experiencing dysglycemia and inappropriate insulin use in nonintensive care settings.
A third of hospitalized individuals have dysglycemia, including hyperglycemia (a quarter of hospitalized individuals have diabetes, and many experience stress hyperglycemia), and hypoglycemia increases mortality and complications (50). An alert-based tool (which includes evidence-based management recommendations) integrated into the electronic medical record (EMR) can detect gaps in hospital glycemic care in real time. This study recorded a significant 21% of all admissions had gaps in care, which are likely much higher when considering only those with predisposed glucose issues, such as hospitalized individuals with diabetes. This alone strengthens the case for the use of DSS. Indeed, previous use of the tool has shown a 10% reduction in recurrent hyperglycemia in adults with diabetes and a 43% reduction in stress hyperglycemia (51).
The present study found that periods when the tool was active were associated with shorter hospital stays, particularly for patients with specific conditions such as recurrent hypoglycemia and stress hyperglycemia. These findings suggest that such tools can indirectly contribute to reducing hospital stays by addressing gaps in care. However, it is important to note that while the results are promising, the overall reduction in hospital stay was modest and not consistently statistically significant across all categories (such as inappropriate insulin use). Further research could delve into refining and improving the tool's efficacy and exploring other variables that might influence hospital outcomes. An interesting study would be to investigate whether inpatient glycemic management had any impact on postdischarge care outcomes.
Assessment of a Decision Support System for Adults with Type 1 Diabetes on Multiple Daily Insulin Injections
Castle JR1, Wilson LM1, Tyler NS2, Espinoza AZ2, Mosquera-Lopez CM2, Kushner T2, Young GM2, Pinsonault J2, Dodier RH2, Hilts WW2, Oganessian SM2, Branigan DL1, Gabo VB1, Eom JH1, Ramsey K3, Youssef JE1,2, Cafazzo JA4,5,6,7, Winters-Stone K8, Jacobs PG3
1Department of Medicine, Division of Endocrinology, Harold Schnitzer Diabetes Health Center, Oregon Health & Science University, Portland, OR; 2Department of Biomedical Engineering, Artificial Intelligence for Medical Systems Lab, Oregon Health & Science University, Portland, OR; 3Biostatistics & Design Program, Oregon Health & Science University, Portland, OR; 4Centre for Global eHealth Innovation, Techna Institute, University Health Network, Toronto, Canada; 5Dalla Lana School of Public Health, Institute of Health Policy, Management and Evaluation, University of Toronto, Toronto, Canada; 6Institute of Biomaterials and Biomedical Engineering, University of Toronto, Toronto, Canada; 7Department of Computer Science, University of Toronto, Toronto, Canada; 8Division of Oncological Sciences, Knight Cancer Institute, Oregon Health & Science University, Portland, OR
DailyDose, a decision support system, was designed to provide real-time dosing advice and weekly insulin dose adjustments for adults living with type 1 diabetes (T1D) and using multiple daily insulin injections.
Methods
Twenty-five adults were enrolled in this single-arm study. All participants used Dexcom G6 for continuous glucose monitoring, InPen for short-acting insulin doses, and Clipsulin to track long-acting insulin doses. The participants used DailyDose on an iPhone for 8 weeks. The primary end point was percent time in range (TIR) comparing the 2-week baseline with the final 2-week period of DailyDose use.
Results
There were no significant differences between TIR or other glycemic metrics between the baseline period compared to final 2-week period of DailyDose use. TIR significantly improved by 6.3% when more than half of recommendations were accepted and followed, compared with 50% or fewer recommendations ([95% CI, 2.5%–10.1%], P = 0.001).
Conclusions
Use of DailyDose did not improve glycemic outcomes compared with the baseline period. The post hoc analysis found that accepting and following the recommendations from DailyDose were associated with improved TIR.
Comments
Population-based registries have consistently demonstrated that individuals managing diabetes through multiple daily injection (MDI) therapy experience comparatively lower glycemic control than those using pump therapy, whether with or without AID systems (52). This emphasizes the importance of implementation of comprehensive decision support systems in this population.
The feasibility of employing DSS with connected pens for adults using MDI therapy and CGM were established in this single-arm study, but the implementation did not result in improved glycemic control as measured by CGM. This adds to two other studies previously reviewed in the Yearbook (53,54) that had yielded similar findings. Nevertheless, as seen in the previous studies, a post hoc analysis showed that a subgroup of participants who accepted and followed the recommendations had significantly improved glycemic control. Individuals using MDI therapy make up a distinct population with various reasons for not choosing pump therapy. It is important to better understand the factors contributing to their low levels of recommendation acceptance. Some of these factors may relate to human aspects, such as simplicity and ease of use of the technology; others may be linked to core management of diabetes. It is worth noting that DSS can offer recommendations, but it cannot automatically compensate for missed meals boluses, correction boluses, or delayed boluses. Therefore, further studies are needed to explore strategies and educational interventions that can promote the use of DSS and enhance its impact on glycemic control.
A particularly noteworthy aspect of this study was that the study team did not interfere with the DSS decisions, instead allowing participants to make their own choices to reject or accept the DSS recommendations. However, this approach might raise concerns about the representativeness of the study population, given the potential selection bias.
Safety and Efficacy of an Adaptive Bolus Calculator for Type 1 Diabetes: A Randomized Controlled Crossover Study
Unsworth R, Armiger R, Jugnee N, Thomas M, Herrero P, Georgiou P, Oliver N, Reddy M
Department of Metabolism, Digestion and Reproduction, Imperial College London, London, UK
The Advanced Bolus Calculator for Type 1 Diabetes (ABC4D) is a decision-support system that uses the artificial intelligence technique of case-based reasoning to adapt and personalize insulin bolus doses. The integrated system comprises a smartphone application and a clinical web portal. This study assessed the safety and efficacy of the ABC4D (intervention) compared with a nonadaptive bolus calculator (control).
Methods
In this prospective, randomized, controlled crossover study, after a 2-week run-in period, the participants were randomized to ABC4D or control for 12 weeks. Then, after a 6-week washout period, the participants crossed over for 12 weeks. The primary outcome was difference in percent time in range (%TIR) (3.9–10.0 mmol/L [70–180 mg/dL]) change during the daytime (07: 00–22: 00) between the two groups.
Results
Thirty-seven adults with type 1 diabetes (T1D) on multiple daily injections of insulin were randomized: median age of 44.7 (IQR, 28.2–55.2) years, diabetes duration of 15.0 (IQR, 9.5–29.0) years, and glycated hemoglobin (HbA1c) of 61.0 (IQR, 58.0–67.0) mmol/mol (7.7% [7.5%–8.3%]). When data from 33 participants were analyzed, there was no significant difference in daytime %TIR change with ABC4D compared with control (median + 0.1% [IQR, −2.6% to + 4.0%] vs + 1.9% [IQR, −3.8% to + 10.1%]; P = 0.53). The participants accepted fewer meal dose recommendations in the intervention compared with control (78.7% [IQR, 55.8%–97.6%] vs 93.5% [73.8%–100%], P = 0.009), with a greater reduction in insulin dosage from that recommended.
Conclusions
The ABC4D is safe for adapting insulin bolus doses, and it provided the same level of glycemic control as the nonadaptive bolus calculator. The results suggest that participants did not follow the ABC4D recommendations as frequently as the control group, which impacted its effectiveness.
Comments
In this study, a DSS that employs case-based reasoning and AI to personalize insulin bolus doses was used to examine the difference in glycemic outcome between an adaptive and a nonadaptive bolus calculator using CGM data. The study design was well suited to address this question, using a randomized controlled crossover approach. All measurements were done to align between the two calculators used (both included dose adjustments for glucose rate of change, exercise, and menstrual period), except for the adaptive component. In addition, the treatment group assignments were masked from the participants, and the recommendations were given to them directly.
Interestingly, no additional improvement in glycemic control or satisfaction was observed when comparing the adaptive and nonadaptive calculators. This lack of differentiation could be attributed to a potential inadequacy in the adaptation module. The adaptation changes were not assessed in the study—for instance, the variation in recommendations for the same situation between adaptive and nonadaptive settings was not explored. Notably, the average meal submission dose taken remained comparable between the groups and was not compared with standard care.
On the other hand, it is worth considering whether higher acceptance of the DSS recommendations could have resulted in increased efficacy. The participants were more receptive to changes in insulin doses exceeding one unit in the control group than in the intervention group. This observation suggests that safety may not have been a significant concern, but perhaps there was a lack of understanding regarding the recommended adaptive calculator changes, which participants may not have followed. This might arise from the frequent changes in correction factors and insulin to carbohydrate ratio. Such an observation highlights the importance of understanding user behavior and engagement when implementing DSS. Future research could potentially concentrate on developing strategies to address this aspect.
Finally, the significance of the study outcome might have been overlooked, given the safety of providing recommendations directly to patients. The study demonstrated that the system is safe for adapting insulin bolus doses and can offer comparable glycemic control.
Comparative Effectiveness of Team-based Care with and without a Clinical Decision Support System for Diabetes Management: A Cluster Randomized Trial
Shi X1,2,3,4, He J4, Lin M1,2,3, Liu C1,2,3, Yan B1,2,3, Song H1,2,3, Wang C1,2,3, Xiao F1,2,3, Huang P1,2,3, Wang L1,2,3, Li Z5, Huang Y1,2,3, Zhang M1,2,3, Chen CS4, Obst K4, Shi L6, Li W7, Yang S1,2,3, Yao G8, Li X1,2,3
1Department of Endocrinology and Diabetes, Xiamen Diabetes Institute, the First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China; 2Xiamen Clinical Medical Center for Endocrine and Metabolic Diseases, the First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China; 3Fujian Province Key Laboratory of Diabetes Translational Medicine, Xiamen, China; 4Department of Epidemiology, Tulane University School of Public Health and Tropical Medicine, New Orleans, LA; 5Epidemiology Research Unit, the First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China; 6Department of Health Policy and Management, Tulane University School of Public Health and Tropical Medicine, New Orleans, LA; 7Department of Cardiology, the First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China; 8Xiamen Municipal Health Commission, Xiamen, China
Uncontrolled hyperglycemia, hypercholesterolemia, and hypertension are common in persons with diabetes. This study compared the effectiveness of team-based care with and without a clinical decision-support system (CDSS) in controlling glycemia, lipids, and blood pressure (BP) among patients with type 2 diabetes (T2D) in 38 community health centers in Xiamen, People's Republic of China.
Methods
The study enrolled 11,132 persons aged ≥ 50 years with uncontrolled diabetes and comorbid conditions: 5475 received team-based care with a CDSS, and 5657 received team-based care alone. The team-based care was delivered by primary care physicians, health coaches, and diabetes specialists in all centers. In addition, a computerized CDSS that generated individualized treatment recommendations based on clinical guidelines was implemented in the 19 centers delivering team-based care with a CDSS. The coprimary outcomes were mean reductions in hemoglobin A1c (HbA1c), low-density lipoprotein cholesterol (LDL-C), and systolic BP over 18 months, plus the proportion of participants with all three risk factors controlled at 18 months.
Results
During the 18-month intervention, in the group receiving team-based care with a CDSS the HbA1c levels significantly decreased by −0.9 percentage point ([95% CI, −0.9 to −0.8 percentage point]; −0.49 mmol/L [95% CI, −0.53 to −0.45 mmol/L]), LDL-C levels decreased by −19.0 mg/dL (95% CI, −20.4 to −17.5 mg/dL), and systolic BP decreased by −9.1 mm Hg (95% CI, −9.9 to −8.3 mm Hg). In team-based care alone, the HbA1c levels decreased by −0.6 percentage point ([95% CI, −0.7 to −0.5 percentage point]; −0.32 mmol/L [95% CI, −0.35 to −0.29 mmol/L]), LDL-C decreased by −12.5 mg/dL (95% CI, −13.6 to −11.3 mg/dL), and systolic BP by −7.5 mm Hg (95% CI,−8.4 to −6.6 mm Hg). The net difference for HbA1c level was 0.2 percentage point ([95% CI, −0.3 to −0.1 percentage point]; −0.17 mmol/L [95% CI, −0.21 to −0.12 mmol/L]), for LDL-C was −6.5 mg/dL (95% CI, −8.3 to −4.6 mg/dL), and for systolic BP was −1.5 mm Hg (95% CI, −2.8 to −0.3 mm Hg). The proportion of patients with controlled HbA1c, LDL-C, and systolic BP was 16.9% (95% CI, 15.7%–18.2%) in team-based care with a CDSS and 13.0% (95% CI, 11.7%–14.3%) in team-based care alone. The study used no usual care control, and the clinical outcome assessors were not masked; the analysis also did not account for multiple comparisons.
Conclusions
Team-based care with a CDSS significantly reduced cardiovascular risk factors in patients with diabetes compared with team-based care alone, but the effect was modest.
Comments
The quality of care for people with diabetes is not improving; in fact, it is deteriorating. Data published in 2021 from the U.S. National Health and Nutrition Examination Survey demonstrated that glycemic and blood pressure control has worsened over the last decade among adults with diabetes. Only one in four adults meets all three American Diabetes Association recommendations for care (glycemic, lipid, and blood pressure control), and less than half achieve the target glycemic control, with even less success seen among those who are insulin treated (55).
The study demonstrated the effectiveness of team-based care in improving diabetes care and reducing cardiovascular risk among individuals with T2D receiving treatment at community clinical centers by primary care physicians. Although the study's approach is not novel, its achievements relied on two significant strengths: first, the large-scale, long-term, randomized control design, and second, the distinctive features of the CDSS.
The CDSS generated personalized medication and dosage recommendations, test results, current treatments, and clinical guidelines. It also included insights into patients’ insurance-related medication policies and local drug availability. In addition, the CDSS offered a concise overview of risk factors and proposed a follow-up schedule. Notably, this CDSS was integrated into team-based health care.
The enhancement in diabetes control observed in the CDSS user group is linked to the system's features, which address critical issues in diabetes care. These include challenges related to a shortage in diabetes specialists, limitations in primary care, and notably therapeutic inertia along with the necessity for personalized treatment. The system provided not only simple reports or remainders but also actionable recommendations for care. The integration into the EMR streamlined its usability and increased adoption.
The study emphasized the importance of integrating actionable DSS into routine practice, particularly within primary care settings, with the aim of providing a comprehensive, all-in-one platform solution for effective diabetes management.
Potential and Pitfalls of ChatGPT and Natural-Language Artificial Intelligence Models for Diabetes Education
Sng GGR1, Tung JYM2, Lim DYZ3, Bee YM1
1Department of Endocrinology, Singapore General Hospital, Singapore; 2Department of Urology, Singapore General Hospital, Singapore; 3Health Services Research Unit, Singapore General Hospital, Singapore
Artificial intelligence–powered large language models, including chatbots like ChatGPT and more than 30 other variants such as Bard AI, can generate human-like text and engage in conversations, making it capable of answering questions, providing explanations, and simulating conversations across a wide range of topics. Some also can generate real-time data, images, voice searches, and a plethora of content creation capabilities (56). Similar to how the internet in the 1990s revolutionized access to information and communication, these chatbots are transforming humanlike interactions with machines. This short article evaluates the use of ChatGPT as an education tool for people with diabetes.
The following abstract was generated by ChatGPT:
“The article discusses the potential and challenges of using artificial intelligence (AI) models, particularly ChatGPT, for diabetes self-management and education (DSME). DSME plays a crucial role in improving diabetes care, but traditional methods face limitations such as access to educators. AI, specifically chatbots, could provide solutions. Modern AI models, like ChatGPT, can generate human-like responses based on vast textual data and engage in natural language interactions. The study aimed to assess the quality of DSME advice provided by ChatGPT.”
Methods
“Researchers evaluated ChatGPT's responses to questions on different DSME domains in both structured and unstructured interactions.”
Results
“ChatGPT performed well, offering clear and concise answers in layperson's terms. However, some inaccuracies were identified. For instance, it didn't recognize insulin storage specifics and showed inflexibility in certain scenarios like recommending diet plans. Despite these limitations, large language models like ChatGPT could potentially assist in basic diabetes education, allowing medical professionals to focus on more complex tasks.”
Conclusions
“The study also highlighted potential safety concerns. ChatGPT generates responses based on language patterns rather than direct knowledge, which can lead to inaccurate or misleading information. This could be problematic for patients seeking reliable medical advice. Healthcare providers should be aware of these AI model limitations when advising patients and consider the risks associated with using AI chatbots for clinical care augmentation. If these challenges are addressed, large language models could transform how patients access medical information, similar to how search engines revolutionized information retrieval in the past.”
Comments
ChatGPT and other AI chatbots are gaining popularity and becoming increasingly accessible worldwide, offering an option for supporting diabetes education. The advantages include delivering real-time, convenient, personalized education and support, giving help to understand complex medical information, and providing diverse information that can support informed decisions; this holds the potential to alleviate the shortage of diabetes educators. A few studies have shown improved diabetes self-management behaviors after ChatGPT education compared with control (57). Nevertheless, these tools have limitations and raise several ethical considerations. Ensuring the tool's accuracy and reliability is a challenge, requiring continuous updates and validation. ChatGPT, trained on the general dataset and updated up to 2021, may provide unreliable information. Furthermore, the lack of human empathy and ethical concerns related to privacy, bias, and discrimination must also be addressed.
This study assessed the responses provided by ChatGPT regarding diet and exercise, hypoglycemia and hyperglycemia education, and insulin management. Although the system addressed all questions, certain inaccuracies were identified. Notably, each response included the recommendation “consult your healthcare team.” This study contributes to similar studies supporting the potential of ChatGPT to provide diabetes education (58,59). Further research is needed to evaluate its effectiveness in the real world compared with traditional education methods and across diverse populations. In addition, the need for additional development in system accuracy and information updates suggests the consideration of an oversight mechanism to assess information quality, along with regulatory measures.
Diabetes Data Science
A PubMed search using the keywords “data science,” “electronic health record,” “electronic medical record” AND “diabetes” yielded 163 results for the period between July 1, 2022, and June 30, 2023. We selected five papers representing five distinct topics to be reviewed in this article. Additional articles of note are referenced as well.
Multiple investigators reported this year on how many days of data are required to provide a stable and accurate estimate of the glycemia management indicator (GMI) (60,61), the glycemia risk index (GRI), and other CGM-derived metrics (62). GMI is a population-based estimate of HbA1c; there is considerable interindividual variance between GMI and HbA1c (63,64). Dunn and colleagues have therefore proposed an alternative measure, the personalized HbA1c (pA1c), a metric that leverages knowledge about the historic relationship between HbA1c and average CGM glucose levels within an individual (65). One author has raised concerns about clinicians’ current reliance on percent time below range as a metric of glycemia, noting considerable shortfalls (66). Others have used CGM and insulin pump data to calculate a latent variable: continuous insulin sensitivity (67). Finally, Espinoza et al. (68,69) wrote about the need for data standards and implementation standards for integrating CGM and insulin pump data into EMR systems; they formed the iCoDE Working Group, which published a report of their meeting proceedings and their recommendations regarding CGM data integration (note: insulin data will be addressed by the iCoDE2 Working Group) (70 –72).
Several authors have written about new cloud-based population health management solutions that integrate data and/or create visual insights that can promote quality care or enable novel care delivery interventions. Zaharieva et al. (73) described updates to the individual patient views for the Timely Interventions for Diabetes Excellence (TIDE) platform, which integrate CGM, physical activity tracker (heart rate), and insulin delivery data into graphical overlays that improve a clinicians’ ability to identify causal events for changes in glycemia. Others have described the creation of glycemic management dashboards within the EMR to aid clinicians in active surveillance of patient glycemia during inpatient hospitalizations (74,75). There were no articles addressing optimal approaches to data visualization during our coverage period; however, because this is the first time this article has addressed diabetes data science, we will point out that in early 2022 Bergenstal et al. (76) did propose color standardization of CGM tracings and ambulatory glucose profile reports to aid in glycemic pattern identification.
Many mature clinical decision-support tools incorporate artificial intelligence/machine learning (AI-ML) algorithms that predict or classify. Multiple groups have conducted the initial development and early validation of new AI-ML algorithms to predict diabetes-related outcomes or classify patient subgroups. One group reviewed the opportunities for implementation of AI-ML tools across the spectrum of diabetes care as well as clinical considerations for their use (77). Still other authors have offered insights on how to implement bioinformatics tools and clinical protocols to support risk identification and population health management (78,79). Other authors developed methods to identify T1D patients who have misdiagnoses in EMR data (80) as well as to predict hypoglycemia during and after physical activity (81), postoperative complications (82), and hospitalization for diabetic ketoacidosis among youth with T1D (83). As the need for near-real-time data insights to drive improvements in care increases, diabetes-related data science is taking center stage. The arrival, maturation, and proliferation of AID systems have finally freed up diabetes researchers’ attention so that they can apply themselves to improving systems of care via innovations in bioinformatics and data science. We provide a focused review of five of the data science articles we evaluated.
Assessment of the Glucose Management Indicator Using Different Sampling Durations
Bailey R1, Calhoun P1, Bergenstal RM2, Beck RW1
1Jaeb Center for Health Research, Tampa, FL; 2International Diabetes Center, HealthPartners Institute, Minneapolis, MN
This study compared the glucose management indicator (GMI) calculated using 14 days of continuous glucose monitor (CGM) data with GMI calculated using < 14 days.
Methods
The analysis included 581 individuals with type 1 diabetes or type 2 diabetes from five clinical trials.
Results
The correlation between the 14- and 7-day GMI was 0.95, and the correlation between 14 days versus 10, 5, and 3 days’ GMI was 0.98, 0.91, and 0.86, respectively. The percentages of GMI values within 0.3% of the 14-day GMI were 98% with 10-day GMI, 87% with 7-day GMI, 77% with 5-day GMI, and 60% with 3-day GMI. Minimal differences were observed between GMI computed using 14 days of data compared with GMI computed with 7 days.
Conclusions
Although 10–14 days of CGM data are preferred for computing GMI, for most patients a satisfactory estimate of HbA1c can be obtained with 7 days of data.
Comments
The GMI has become an extremely useful data feature across the field. First, it allows clinical trials to analyze an outcome measure that correlates well with HbA1c, without incurring the additional cost of biochemically measuring HbA1c. Second, it has the potential to reduce missing data in registry-based and EMR-based studies where CGM data are present but HbA1c is partially or completely missing. Third, it continues to allow investigators to relate current and future research to historical research from the pre-CGM era.
A fundamental question that has arisen is how many days of CGM data are required to create the most accurate and stable estimate of GMI? The present article answers this question, at least for the population included from the five clinical trials evaluated by Bailey and colleagues. Another article cited in the introduction to this section noted that in young children > 14 days of CGM data may be required to create a stable estimate of GMI with the highest correlation to HbA1c.
Personalized Glycated Hemoglobin in Diabetes Management: Closing the Gap with Glucose Management Indicator
Dunn TC1, Xu Y1, Bergenstal RM2, Ogawa W3, Ajjan RA4
1Clinical Affairs, Abbott Diabetes Care, Alameda, CA; 2International Diabetes Center, HealthPartners Institute, Minneapolis, MN; 3Division of Diabetes and Endocrinology, Kobe University Graduate School of Medicine, Kobe, Japan; 4The LIGHT Laboratories, Leeds Institute of Cardiovascular and Metabolic Medicine, University of Leeds, Leeds, UK
Glycated hemoglobin (HbA1c) has played a central role in the management of diabetes since the end of the landmark Diabetes Control and Complications Trial 30 years ago. However, it is known to be subject to distortions related to altered red blood cell (RBC) properties, including changes in cellular life span. On occasion, the distortion of HbA1c is associated with a clinical pathological condition affecting RBCs; however, the more frequent scenario is related to interindividual RBC variations that alter the HbA1c–average glucose relationship. Clinically, these variations can potentially lead to over- or underestimating glucose exposure of the individual to an extent that may put the person at excess risk of over- or undertreatment. Furthermore, the variable association between HbA1c and glucose levels across different groups of people may become an unintentional driver of inequitable health care delivery, outcomes, and incentives. The subclinical effects within the normal expected physiological range of RBCs can be large enough to alter clinical interpretation of HbA1c, and addressing this will help with individualized care and decision-making.
Methods
This review describes a new glycemic measure, personalized HbA1c (pA1c), which may address the clinical inaccuracies of HbA1c.
Results
The authors demonstrate how pA1c takes into account interindividual variability in RBC glucose uptake and life span.
Conclusions
pA1c represents a more sophisticated understanding of glucose–HbA1c relationship at an individual level. Future use of pA1c, after adequate clinical validation, has the potential to refine glycemic management and the diagnostic criteria in diabetes.
Comments
Although clinicians and investigators have an increased interest in CGM-specific metrics like percent time in range, the field still has a considerable interest in HbA1c-related metrics because they allow investigators to relate new research to historic research. Additionally, a considerable portion of the global population with diabetes mellitus still has no access to CGM, creating continued reliance on biochemical HbA1c measurement. GMI offers a formula for a reasonable population-based estimate of HbA1c derived from CGM data, but it has limitations. The personalized A1c (pA1c) offers the potential promise of a more precise estimate of HbA1c as long as one has information on the individual's own historic HbA1c and CGM data. The idea of calibrating glucose readings with laboratory HbA1c to account for an individual's rate of glycation and clearance is not new—it was proposed 3 years ago by Fabris et al., who concluded, “Using a model individualized with one HbA1c measurement, TIR provides an accurate approximation of HbA1c for at least 6 months, reflecting blood glucose fluctuations and nonglycemic biological factors” (84). These methods have the potential to allow more robust health outcomes analyses on large registry-based and EMR-based datasets, to improve the handling of missing data in clinical trials, and to reduce missing data for AI-ML model development in diabetes care.
A New Index of Insulin Sensitivity from Glucose Sensor and Insulin Pump Data: In silico and In vivo Validation in Youths with Type 1 Diabetes
Schiavon M1, Galderisi A2,3, Basu A4, Kudva YC5, Cengiz E6, Dalla Man C1
1Department of Information Engineering, University of Padova, Padova, Italy; 2Department of Woman and Child's Health, University of Padova, Padova, Italy; 3Department of Pediatrics, Yale University, New Haven, CT; 4Division of Endocrinology, University of Virginia, Charlottesville, VA; 5Division of Endocrinology, Diabetes, Metabolism, and Nutrition, Department of Internal Medicine, Mayo Clinic and Foundation, Rochester, MN; 6Pediatric Diabetes Program, University of California San Francisco (UCSF) School of Medicine, San Francisco, CA
Estimation of insulin sensitivity (SI) and its daily variation are key for optimizing insulin therapy in patients with type 1 diabetes (T1D). We recently developed a method for SI estimation from continuous glucose monitoring (CGM) and continuous subcutaneous insulin infusion (CSII) data in adults with T1D (SISP) and validated it under restrained experimental conditions. Herein, we validate in vivo a new version of SISP performing well in daily-life unrestrained conditions.
Methods
The new SISP was tested in both simulated and real data. The simulated dataset consists of 100 virtual adults of the UVa/Padova T1D Simulator monitored during an open-loop experiment, whereas the real dataset consists of 10 youths with T1D monitored during a hybrid closed-loop meal study. In both datasets, the participants underwent two consecutive meals (breakfast and lunch, at 7:00 and 11:00
Results
SISP was estimated with good precision (median coefficient of variation < 20%) in 100% of the real and 91% of the simulated meals. SISP and SIMM were highly correlated, both in the simulated and real datasets (R = 0.82 and R = 0.83, P < 0.001), and exhibited a similar intraday pattern.
Conclusions
SISP is suitable for estimating SI in both closed- and open-loop settings, provided that the patient wears a CGM sensor and a subcutaneous insulin pump.
Comments
In many real-world data-derived studies and in most AI-ML models, insulin sensitivity is an unmeasured (latent) variable. It remains to be determined whether a point estimate of insulin sensitivity could improve automated insulin delivery algorithms, health outcomes analyses, and AI-ML models to predict near-term diabetes outcomes. Yet this newly derived feature calculated from widely available CGM and insulin delivery device data promises to open up multiple new lines of investigation that could advance important objectives in the field.
The Launch of the iCoDE Standard Project
Xu NY1, Nguyen KT1, DuBord AY2, Klonoff DC2,3, Goldman JM4, Shah SN5, Spanakis EK6,7, Madlock-Brown C8, Sarlati S2,9, Rafiq A10, Wirth A11, Kerr D12, Khanna R2, Weinstein S13, Espinoza J14
1Diabetes Technology Society, Burlingame, CA; 2University of California, San Francisco, San Francisco, CA; 3Mills-Peninsula Medical Center, San Mateo, CA; 4Massachusetts General Hospital, Boston, MA; 5Netspective Communications LLC, Silver Spring, MD; 6Baltimore VA Medical Center, Baltimore, MD; 7University of Maryland, Baltimore, MD; 8The University of Tennessee Health Science Center, Memphis, TN; 9Anthem, Inc., Indianapolis, IN; 10National Aeronautics and Space Administration, Washington, DC; 11MedCrypt, San Diego, CA; 12Sansum Diabetes Research Institute, Santa Barbara, CA; 13McDermott Will & Emery, Washington, DC; 14Division of General Pediatrics, Department of Pediatrics, Children's Hospital Los Angeles, University of Southern California, Los Angeles, CA
On January 27, 2022, the first meeting of the Integration of Continuous Glucose Monitor Data into the Electronic Health Record (iCoDE) project, organized by Diabetes Technology Society, was held virtually.
Methods
The panels and breakout groups of clinicians, government officials, data aggregators, attorneys, and standards experts addressed three themes: (1) why digital health data integration into the electronic medical record (EMR) is needed, (2) what integrated continuously monitored glucose data will look like, and (3) how this process can be achieved in a way that will satisfy clinicians, health-care organizations, and regulatory experts.
Results
These eight sessions were held to address the meeting's overall themes. (1) What Do Inpatient Clinicians Want to See with Integration of CGM Data into the EMR? (2) What Do Outpatient Clinicians Want to See with Integration of CGM Data into the EMR? (3) Why Are Data Standards and Guidances Useful? (4) What Value Can Data Integration Services Add? (5) What Are Examples of Successful Integration? (6) Which Privacy, Security, and Regulatory Issues Must Be Addressed to Integrate CGM Data into the EMR? (7) Breakout Group Discussions. (8) Presentation of Breakout Group Ideas.
Conclusions
Data standards and workflow guidance must be created as necessary components of the Integration of Continuous Glucose Monitor Data into the Electronic Health Record (iCoDE) standard project. This initial meeting to launch iCoDE will be followed by more working group meetings to create the needed standards.
Comments
The continued barriers to open and protected data mobility must come to an end. The iCoDE initiative offers a framework for health systems to think about integrating CGM data into EMR systems. This review nicely addresses the work that was performed, and the framework in which the workgroup participants tackled the problems of data integration. There are two notable methods for integrating CGM data into EMRs: (1) direct integration between data source (third-party software) and EMR via an application programming interface (API), and (2) indirect integration between data source and a middleware software package that can serve as a staging ground for subsequent transfer of data from diverse sources into the EMR. The latter approach holds considerable promise with regard to scalability. It is expensive and time consuming to build new API connections into the highly regulated EMR environment. A middleware approach would allow health systems to create a “funnel” with all external data funneling into a single data “pipe” (HL7/FHIR connection) into the EMR; it seems logical that this is more scalable, controllable, and cost-efficient than the “many pipes” approach.
Adding Glycemic and Physical Activity Metrics to a Multimodal Algorithm-enabled Decision Support Tool for Type 1 Diabetes Care: Keys to Implementation and Opportunities
Zaharieva DP1, Senanayake R2, Brown C3, Watkins B3, Loving G3, Prahalad P1,4, Ferstad JO5, Guestrin C2, Fox EB2,6,7, Maahs DM1,4,8, Scheinker D1,3,4,5,9, the 4T Research Team
1Division of Endocrinology, Department of Pediatrics, Stanford University, School of Medicine, Stanford, CA; 2Department of Computer Science, Stanford University, Stanford, CA; 3Stanford Children's Health, Lucile Packard Children's Hospital, Stanford, CA; 4Stanford Diabetes Research Center, Stanford University, Stanford, CA; 5Department of Management Science and Engineering, Stanford University School of Engineering, Stanford, CA; 6Chan Zuckerberg Biohub, San Francisco, CA; 7Department of Statistics, Stanford University, Stanford, CA; 8Department of Health Research and Policy, Stanford University School of Medicine, Stanford, CA; 9Clinical Excellence Research Center, Stanford University School of Medicine, Stanford, CA
Although regular exercise and physical activity are essential to increasing cardiovascular fitness, increasing insulin sensitivity, and improving overall well-being of youths and adults with type 1 diabetes (T1D), exercise can lead to fluctuations in glycemia during and after the activity. Algorithm-enabled patient prioritization and remote patient monitoring (RPM), which have been used to improve clinical workflows, have been associated with improved glucose time in range in newly diagnosed youths with T1D. This novel algorithm-enabled care model currently integrates continuous glucose monitoring (CGM) data to prioritize patients for weekly reviews by the clinical diabetes team, but additional data may help clinical teams make more informed decisions around T1D management, particularly with youths undergoing regular physical exercise.
Methods
This work provides an overview of the essential steps of integrating exercise data into an RPM program and examines the most promising opportunities for the use of these data.
Results
The authors describe the technical path and methodologies for integrating physical activity metrics from wrist-worn fitness wearables into a CGM-based population health management software. Visualizations of physical activity metrics in the context of CGM data are proposed. Future iterations of the CGM-based and RPM-driven care model will integrate physical activity metrics (e.g., heart rate and step count) and physical activity flags to help identify the patients whose needs are not fully captured by CGM data. This technical methodology may help health-care professionals improve patient care with a better integration of CGM and physical activity data.
Conclusions
Incorporating exercise data into the current CGM-based care model will produce specific, clinically relevant information such as identifying whether patients are meeting exercise guidelines.
Comments
While this manuscript is not a scientific study, it provides a technical blueprint for extracting physical activity data from Garmin fitness wearable devices via the Garmin Health cloud data system, and for presenting those data side by side with CGM data in a CGM-based population health management platform. This represents an important milestone, because data from physical activity wearables are not widely integrated into electronic health records, and these data have largely been absent from population health management platforms. The authors propose an approach to aligning data related to glucose, heart rate, and type and intensity of activity, which is no small task given the differences in frequency and regularity with which the measures are each captured. The authors then propose an approach to visualizing these data to inform clinical decision-making related to diabetes management among CGM users.
An “All-Data-on-Hand” Deep Learning Model to Predict Hospitalization for Diabetic Ketoacidosis in Youth with Type 1 Diabetes: Development and Validation Study
Williams DD1, Ferro D2,3, Mullaney C4, Skrabonja L4, Barnes MS2, Patton SR5, Lockee B2, Tallon EM2, Vandervelden CA2, Schweisberger C2, Mehta S6, McDonough R2, Lind M7,8,9, D'Avolio L4, Clements MA2
1Health Services and Outcomes Research, Children's Mercy - Kansas City, Kansas City, MO; 2Department of Endocrinology, Children's Mercy - Kansas City, Kansas City, MO; 3IRCCS, Bambino Gesù Children's Hospital, Rome, Italy; 4Cyft, Inc., Cambridge, MA; 5Center for Healthcare Delivery Science, Nemours Children's Health, Jacksonville, FL; 6Joslin Diabetes Center, Boston, MA; 7Department of Medicine, NU-Hospital Group, Uddevalla, Sweden; 8Department of Molecular and Clinical Medicine, University of Gothenburg, Gothenburg, Sweden; 9Department of Internal Medicine, Sahlgrenska University Hospital, Gothenburg, Sweden
Although prior research has identified multiple risk factors for diabetic ketoacidosis (DKA), clinicians continue to lack clinic-ready models to predict these dangerous and costly episodes. Deep learning, specifically the use of a long short-term memory (LSTM) model, may be the answer to accurately predicting the 180-day risk of DKA-related hospitalization for youths with type 1 diabetes (T1D). This LSTM model was developed to predict the 180-day risk of DKA-related hospitalization for youths with T1D.
Methods
The clinical data from 17 consecutive calendar quarters (January 10, 2016, to March 18, 2020) for 1745 youths aged 8 to 18 years with T1D from a pediatric diabetes clinic network in the midwestern United States included demographics, discrete clinical observations (laboratory results, vital signs, anthropometric measures, diagnosis, and procedure codes), medications, visit counts by type of encounter, number of historic DKA episodes, number of days since last DKA admission, patient-reported outcomes (answers to clinic intake questions), and data features derived from diabetes- and nondiabetes-related clinical notes via natural language processing. The model was trained using this input data from quarters 1 to 7 (n = 1377). It was validated using the input data from quarters 3 to 9 in a partial out-of-sample cohort (OOS-P; n = 1505). It was further validated in a full out-of-sample cohort (OOS-F; n = 354) with input from quarters 10 to 15.
Results
In both OOS cohorts, admissions for DKA occurred at a rate of 5% per 180-days. In the OOS-P and OOS-F cohorts, the median age was 13.7 (IQR, 11.3–15.8) years and 13.1 (IQR, 10.7–15.5) years, respectively. The median glycated hemoglobin (HbA1c) levels at enrollment were 8.6% (IQR, 7.6%–9.8%) and 8.1% (IQR, 6.9%–9.5%), respectively. Recall was 33% (26 of 80) and 50% (9 of 18), respectively, for the top-ranked 5% of OOS-P and OOS-F youths; 14.15% (213 of 1505) and 12.7% (45 of 354), respectively, had had prior DKA admissions (after the T1D diagnosis). For lists rank ordered by the probability of hospitalization, precision increased from 33% to 56% to 100% for positions 1 to 80, 1 to 25, and 1 to 10 in the OOS-P cohort and from 50% to 60% to 80% for positions 1 to 18, 1 to 10, and 1 to 5 in the OOS-F cohort, respectively.
Conclusions
The proposed LSTM model for predicting 180-day DKA-related hospitalization was valid in this sample. Future research should evaluate the model's validity in multiple populations and settings to account for health inequities that may be present in different segments of the population, such as with racially or socioeconomically diverse cohorts. Rank ordering youths by their probability of DKA-related hospitalization will allow clinics to identify the youths who are most at risk. Clinics may then be able to create and evaluate novel preventive interventions based on their available resources.
Comments
The creation of clinical DSS that incorporate AI-ML-based prediction or classification must be preceded by considerable foundational research to develop and validate the AI-ML models that drive them. This work is among the first to attempt to use all available EMR data to predict a clinical outcome in a several-month time frame. Although the work has several limitations, it (1) sets the current bar for the level of positive predictive value and sensitivity achievable for predicting DKA in a 6-month horizon; (2) paves the way for the creation of more parsimonious models with fewer data features that can support the goals of explainable AI; and (3) proposes a practical roadmap for how to turn the model's predictions into an eventual DSS.
Footnotes
Author Disclosure Statement
Revital Nimri has received grants from Helmsley Charitable Trust, Dexcom, Medtronic, Abbott Diabetes Care, and Insulet; personal fees and others from DreaMed Diabetes Ltd; personal fees from Novo Nordisk and Eli Lilly; and owns stock in DreaMed Diabetes Ltd.
Moshe Phillip is an advisory board member of Medtronic Diabetes, Pfizer, Sanofi, and DOMPE; he has received consulting fees and honoraria from Eli Lilly, Medtronic Diabetes, Novo Nordisk, Pfizer, Sanofi, Qulab Medical, Ascensia, ProventionBio, and Bayer. The institute he is heading has received research grants from Dexcom, Eli Lilly, Insulet, Medtronic Diabetes, Novo Nordisk, Pfizer, Roche Diagnostics, Sanofi, DreaMed-Diabetes, NG Solutions, Dompe, Lumos, GWAVE, OPKO, and ProventionBio. He owns stock in DreaMed-Diabetes and NG Solutions.
Mark A. Clements has received personal consultant fees as chief medical officer from Glooko and research support handled by Children's Mercy Kansas City from Dexcom and Abbott Diabetes Care.
Boris Kovatchev has received patent royalties handled by UVA from DexCom, J&J, Novo Nordisk, and Sanofi and has received research support handled by UVA from Dexcom and Novo Nordisk.