Use of case-based reasoning to enhance intensive management of patients on insulin pump therapy

The OhioT1DM Dataset for Blood Glucose Level Prediction: Update 2020

This paper documents the OhioT1DM Dataset, which was developed to promote and facilitate research in blood glucose level prediction. It contains eight weeks' worth of continuous glucose monitoring, insulin, physiological sensor, and self-reported life-event data for each of 12 people with type 1 diabetes. An associated graphical software tool allows researchers to visualize the integrated data. The paper details the contents and format of the dataset and tells interested researchers how to obtain it. The OhioT1DM Dataset was first released in 2018 for the first Blood Glucose Level Prediction (BGLP) Challenge. At that time, the dataset was half its current size, containing data for only six people with type 1 diabetes. Data for an additional six people is being released in 2020 for the second BGLP Challenge. This paper subsumes and supersedes the paper which documented the original dataset.

Artificial Intelligence in Decision Support Systems for Type 1 Diabetes

Type 1 diabetes (T1D) is a chronic health condition resulting from pancreatic beta cell dysfunction and insulin depletion. While automated insulin delivery systems are now available, many people choose to manage insulin delivery manually through insulin pumps or through multiple daily injections. Frequent insulin titrations are needed to adequately manage glucose, however, provider adjustments are typically made every several months. Recent automated decision support systems incorporate artificial intelligence algorithms to deliver personalized recommendations regarding insulin doses and daily behaviors. This paper presents a comprehensive review of computational and artificial intelligence-based decision support systems to manage T1D. Articles were obtained from PubMed, IEEE Xplore, and ScienceDirect databases. No time period restrictions were imposed on the search. After removing off-topic articles and duplicates, 562 articles were left to review. Of those articles, we identified 61 articles for comprehensive review based on algorithm evaluation using real-world human data, in silico trials, or clinical studies. We grouped decision support systems into general categories of (1) those which recommend adjustments to insulin and (2) those which predict and help avoid hypoglycemia. We review the artificial intelligence methods used for each type of decision support system, and discuss the performance and potential applications of these systems.

A systematic review of the applications of artificial intelligence and machine learning in autoimmune diseases

Autoimmune diseases are chronic, multifactorial conditions. Through machine learning (ML), a branch of the wider field of artificial intelligence, it is possible to extract patterns within patient data, and exploit these patterns to predict patient outcomes for improved clinical management. Here, we surveyed the use of ML methods to address clinical problems in autoimmune disease. A systematic review was conducted using MEDLINE, embase and computers and applied sciences complete databases. Relevant papers included "machine learning" or "artificial intelligence" and the autoimmune diseases search term(s) in their title, abstract or key words. Exclusion criteria: studies not written in English, no real human patient data included, publication prior to 2001, studies that were not peer reviewed, non-autoimmune disease comorbidity research and review papers. 169 (of 702) studies met the criteria for inclusion. Support vector machines and random forests were the most popular ML methods used. ML models using data on multiple sclerosis, rheumatoid arthritis and inflammatory bowel disease were most common. A small proportion of studies (7.7% or 13/169) combined different data types in the modelling process. Cross-validation, combined with a separate testing set for more robust model evaluation occurred in 8.3% of papers (14/169). The field may benefit from adopting a best practice of validation, cross-validation and independent testing of ML models. Many models achieved good predictive results in simple scenarios (e.g. classification of cases and controls). Progression to more complex predictive models may be achievable in future through integration of multiple data types.

Using LSTMs to learn physiological models of blood glucose behavior

For people with type 1 diabetes, good blood glucose control is essential to keeping serious disease complications at bay. This entails carefully monitoring blood glucose levels and taking corrective steps whenever they are too high or too low. If blood glucose levels could be accurately predicted, patients could take proactive steps to prevent blood glucose excursions from occurring. However, accurate predictions require complex physiological models of blood glucose behavior. Factors such as insulin boluses, carbohydrate intake, and exercise influence blood glucose in ways that are difficult to capture through manually engineered equations. In this paper, we describe a recursive neural network (RNN) approach that uses long short-term memory (LSTM) units to learn a physiological model of blood glucose. When trained on raw data from real patients, the LSTM networks (LSTMs) obtain results that are competitive with a previous state-of-the-art model based on manually engineered physiological equations. The RNN approach can incorporate arbitrary physiological parameters without the need for sophisticated manual engineering, thus holding the promise of further improvements in prediction accuracy.

Artificial Intelligence Methodologies and Their Application to Diabetes

In the past decade diabetes management has been transformed by the addition of continuous glucose monitoring and insulin pump data. More recently, a wide variety of functions and physiologic variables, such as heart rate, hours of sleep, number of steps walked and movement, have been available through wristbands or watches. New data, hydration, geolocation, and barometric pressure, among others, will be incorporated in the future. All these parameters, when analyzed, can be helpful for patients and doctors' decision support. Similar new scenarios have appeared in most medical fields, in such a way that in recent years, there has been an increased interest in the development and application of the methods of artificial intelligence (AI) to decision support and knowledge acquisition. Multidisciplinary research teams integrated by computer engineers and doctors are more and more frequent, mirroring the need of cooperation in this new topic. AI, as a science, can be defined as the ability to make computers do things that would require intelligence if done by humans. Increasingly, diabetes-related journals have been incorporating publications focused on AI tools applied to diabetes. In summary, diabetes management scenarios have suffered a deep transformation that forces diabetologists to incorporate skills from new areas. This recently needed knowledge includes AI tools, which have become part of the diabetes health care. The aim of this article is to explain in an easy and plane way the most used AI methodologies to promote the implication of health care providers-doctors and nurses-in this field.

Bolus Calculator Reduces Hypoglycemia in the Short Term and Fear of Hypoglycemia in the Long Term in Subjects with Type 1 Diabetes (CBMDI Study)

BACKGROUND

In a previous study we demonstrated improvement in metabolic control and reduction in hypoglycemia in people with type 1 diabetes on multiple daily injections, after having used a bolus calculator for 4 months.

OBJECTIVE

To demonstrate whether (1) extending its use (2) or introducing it in the control group, previously subjected to treatment intensification, could further improve metabolic control and related psychological issues.

METHODS

After the previous clinical trial, in which the subjects were randomized either to treatment with the calculator or to control group for 4 months, both groups used the calculator during an additional 4-month period.

RESULTS

In the previous control group, after using the device, HbA1c did not improve (7.86% ± 0.87% vs. 8.01% ± 0.93%, P 0.215), although a significant decrease in postprandial hypoglycemia was observed (2.3 ± 2 vs. 1.1 ± 1.2/2 weeks, P 0.002). In the group in which the treatment was extended from 4 to 8 months, HbA1c did not improve either (7.61 ± 0.58 vs. 7.73 ± 0.65, P 0.209); however this group had a greater perceived treatment satisfaction (12.03 ± 4.26 vs. 13.71 ± 3.75, P 0.007) and a significant decrease in fear of hypoglycemia (28.24 ± 8.18 basal vs. 25.66 ± 8.02 at 8 months, P 0.026).

CONCLUSIONS

The extension in the use of the calculator or its introduction in a previously intensified control group did not improve metabolic control, although it did confirm a decrease in hypoglycemic episodes in the short term, while the extension of its use to 8 months was associated with a reduction in fear of hypoglycemia and greater treatment satisfaction.

Improved A1C Levels in Type 1 Diabetes with Smartphone App Use

OBJECTIVES

Smartphones are a potentially useful tool in diabetes care. We have developed an application (app) linked to a website, Intelligent Diabetes Management (IDM), which serves as both an insulin bolus calculator and an electronic diabetes diary. We have prospectively studied whether patients using this app improved control of their glucose levels.

METHODS

Patients with type 1 diabetes were recruited. There was a 4-week observation period, midway during which we offered to review the participants' records. The app was then downloaded and participants' diabetes regimens entered on the synchronized IDM website. At 2, 4, 8, 12 and 16 weeks of the active phase, their records were reviewed online, and feedback was provided electronically. The primary endpoint was change in levels of glycated hemoglobin (A1C).

RESULTS

Of the 31 patients recruited, 18 completed the study. These 18 made 572±98 entries per person on the IDM system over the course of the study (≈5.1/day). Their ages were 40.0±13.9 years, the durations of their diabetes were 27.3±14.9 years and 44% used insulin pumps. The median A1C level fell from 8.1% (7.5 to 9.0, IQ range) to 7.8% (6.9 to 8.3; p<0.001). During the observation period, glucose records were reviewed for 50% of the participants. In the active phase, review of the glucose diaries took less time on the IDM website than using personal glucose records in the observation period, median 6 minutes (5 to 7.5 IQ range) vs. 10 minutes (7.5 to 10.5 IQ range; p<0.05).

CONCLUSIONS

Our smartphone app enables online review of glucose records, requires less time for clinical staff and is associated with improved glucose control.

Clinical Safety and Feasibility of the Advanced Bolus Calculator for Type 1 Diabetes Based on Case-Based Reasoning: A 6-Week Nonrandomized Single-Arm Pilot Study

BACKGROUND

The Advanced Bolus Calculator for Diabetes (ABC4D) is an insulin bolus dose decision support system based on case-based reasoning (CBR). The system is implemented in a smartphone application to provide personalized and adaptive insulin bolus advice for people with type 1 diabetes. We aimed to assess proof of concept, safety, and feasibility of ABC4D in a free-living environment over 6 weeks.

METHODS

Prospective nonrandomized single-arm pilot study. Participants used the ABC4D smartphone application for 6 weeks in their home environment, attending the clinical research facility weekly for data upload, revision, and adaptation of the CBR case base. The primary outcome was postprandial hypoglycemia.

RESULTS

Ten adults with type 1 diabetes, on multiple daily injections of insulin, mean (standard deviation) age 47 (17), diabetes duration 25 (16), and HbA1c 68 (16) mmol/mol (8.4 (1.5) %) participated. A total of 182 and 150 meals, in week 1 and week 6, respectively, were included in the analysis of postprandial outcomes. The median (interquartile range) number of postprandial hypoglycemia episodes within 6-h after the meal was 4.5 (2.0-8.2) in week 1 versus 2.0 (0.5-6.5) in week 6 (P = 0.1). No episodes of severe hypoglycemia occurred during the study.

CONCLUSION

The ABC4D is safe for use as a decision support tool for insulin bolus dosing in self-management of type 1 diabetes. A trend suggesting a reduction in postprandial hypoglycemia was observed in the final week compared with week 1.

Use of Automated Bolus Calculators for Diabetes Management

Fewer than 30 % of patients with diabetes who are on insulin therapy achieve target glycated haemoglobin (HbA1C) levels. Automated bolus calculators (ABCs) are now almost universally used for patients on insulin pump therapy to calculate pre-meal insulin doses. Use of ABCs in glucose monitors and smart phone applications have the potential to improve glucose control in a larger population of individuals with diabetes on insulin therapy by overcoming the fear of hypoglycaemia and assisting those with low numeracy skills.

A consensus perceived glycemic variability metric

OBJECTIVE

Glycemic variability (GV) is an important component of overall glycemic control for patients with diabetes mellitus. Physicians are able to recognize excessive GV from continuous glucose monitoring (CGM) plots; however, there is currently no universally agreed upon GV metric. The objective of this study was to develop a consensus perceived glycemic variability (CPGV) metric that could be routinely applied to CGM data to assess diabetes mellitus control.

METHODS

Twelve physicians actively managing patients with type 1 diabetes mellitus rated a total of 250 24 h CGM plots as exhibiting low, borderline, high, or extremely high GV. Ratings were averaged to obtain a consensus and then input into two machine learning algorithms: multilayer perceptrons (MPs) and support vector machines for regression (SVR). In silica experiments were run using each algorithm with different combinations of 12 descriptive input features. Ten-fold cross validation was used to evaluate the performance of each model.

RESULTS

The SVR models approximated the physician consensus ratings of unseen CGM plots better than the MP models. When judged by the root mean square error, the best SVR model performed comparably to individual physicians at matching consensus ratings. When applied to 262 different CGM plots as a screen for excessive GV, this model had accuracy, sensitivity, and specificity of 90.1%, 97.0%, and 74.1%, respectively. It significantly outperformed mean amplitude of glycemic excursion, standard deviation, distance traveled, and excursion frequency.

CONCLUSIONS

This new CPGV metric could be used as a routine measure of overall glucose control to supplement glycosylated hemoglobin in clinical practice.

Automated glycemic pattern analysis: overcoming diabetes clinical inertia

The OneTouch® Verio™ IQ Meter with PatternAlert™ Technology has been approved by the U.S. Food and Drug Administration as the first self-glucose monitor that can automatically determine glycemic patterns [high and low pre-meal blood glucose (BG)] for health care providers (HCPs) and patients. In this issue of Journal of Diabetes Science and Technology, Katz and coauthors demonstrate that this device was more accurate and quicker in detecting abnormal glucose patterns than the review by HCPs of 30-day handwritten BG logs and that its interpretations were positively accepted by the HCPs. Continued development of automated pattern analysis and decision-support software to overcome the "data-overload" associated with intensive glucose monitoring and diabetes management will reduce clinical inertia and could dramatically improve diabetes outcomes.

Analysis of use of an automated bolus calculator reduces fear of hypoglycemia and improves confidence in dosage accuracy in type 1 diabetes mellitus patients treated with multiple daily insulin injections

In this issue of Journal of Diabetes Science and Technology, Barnard and colleagues evaluate the use of the ACCU-CHEK® Aviva Expert blood glucose meter/bolus advisor system in patients with type 1 diabetes mellitus. Hypoglycemia is a major limiting factor to intensive glucose control, and fear of hypoglycemia, especially in those who have experienced severe reactions, is a major barrier. The bolus advisor improved overall glucose control and increased adherence by overcoming the patients' fear of hypoglycemia, giving them more confidence to give adequate doses of insulin to control hyperglycemia. In this review, we discuss other human factors that become barriers to intensive control, which can benefit from new technologies, including numeracy literacy, information overload, time required for diabetes self-care, and device incompatibility.

Characterizing blood glucose variability using new metrics with continuous glucose monitoring data

OBJECTIVE

Glycemic variability contributes to oxidative stress, which has been linked to the pathogenesis of the long-term complications of diabetes. Currently, the best metric for assessing glycemic variability is mean amplitude of glycemic excursion (MAGE); however, MAGE is not in routine clinical use. A glycemic variability metric in routine clinical use could potentially be an important measure of overall glucose control and a predictor of diabetes complication risk not detected by glycosylated hemoglobin (A1C) levels. This study aimed to develop and evaluate new automated metrics of glycemic variability that could be routinely applied to continuous glucose monitoring (CGM) data to assess and enhance glucose control.

METHOD

Individual 24 h CGM tracings from our clinical diabetes research database were scored for MAGE and two additional metrics designed to compensate for aspects of variability not captured by MAGE: (1) number of daily glucose fluctuations >75 mg/dl that leave the normal range (70-175 mg/dl), or excursion frequency, and (2) total daily fluctuation, or distance traveled. These scores were used to train machine learning algorithms to recognize excessive variability based on physician ratings of daily CGM charts, producing a third metric of glycemic variability: perceived variability. Finger stick A1C (average) and serum 1,5-anhydroglucitol (postprandial) levels were used as clinical markers of overall glucose control for comparison.

RESULTS

Mean amplitude of glycemic excursion, excursion frequency, and distance traveled did not adequately quantify the glycemic variability visualized by physicians who evaluated the daily CGM plots. A naive Bayes classifier was developed that characterizes CGM tracings based on physician interpretations of tracings. Preliminary results suggest that the number of excessively variable days, as determined by this naive Bayes classifier, may be an effective way to automatically assess glycemic variability of CGM data. This metric more closely reflects 90-day changes in serum 1,5-anhydroglucitol levels than does MAGE.

CONCLUSION

We have developed a new automated metric to assess overall glycemic variability in people with diabetes using CGM, which could easily be incorporated into commercially available CGM software. Additional work to validate and refine this metric is underway. Future studies are planned to correlate the metric with both urinary 8-iso-prostaglandin F2 alpha excretion and serum 1,5-anhydroglucitol levels to see how well it identifies patients with high glycemic variability and increased markers of oxidative stress to assess risk for long-term complications of diabetes.

Evaluating the automated blood glucose pattern detection and case-retrieval modules of the 4 Diabetes Support System

BACKGROUND

We have developed a prototypical case-based reasoning system to enhance management of patients with type 1 diabetes mellitus (T1DM). The system is capable of automatically analyzing large volumes of life events, self-monitoring of blood glucose readings, continuous glucose monitoring system results, and insulin pump data to detect clinical problems. In a preliminary study, manual entry of large volumes of life-event and other data was too burdensome for patients. In this study, life-event and pump data collection were automated, and then the system was reevaluated.

METHODS

Twenty-three adult T1DM patients on insulin pumps completed the five-week study. A usual daily schedule was entered into the database, and patients were only required to upload their insulin pump data to Medtronic's CareLink® Web site weekly. Situation assessment routines were run weekly for each participant to detect possible problems, and once the trial was completed, the case-retrieval module was tested.

RESULTS

Using the situation assessment routines previously developed, the system found 295 possible problems. The enhanced system detected only 2.6 problems per patient per week compared to 4.9 problems per patient per week in the preliminary study (p=.017). Problems detected by the system were correctly identified in 97.9% of the cases, and 96.1% of these were clinically useful.

CONCLUSIONS

With less life-event data, the system is unable to detect certain clinical problems and detects fewer problems overall. Additional work is needed to provide device/software interfaces that allow patients to provide this data quickly and conveniently.

The role of technology and the chronic care model

Innovative technological approaches offer great promise for enhancing the quality of care and improved access. A chronic care model has been shown repeatedly to improve outcomes. The elements of the model include the health system, community, self-management support, decision support, clinical information systems, and delivery system redesign. Understanding opportunities to apply technology to the chronic care model is critically important as the rates of diabetes escalate and quality care becomes a priority for health systems.

Dose accuracy of SoloSTAR and FlexPen as assessed in a clinical setting

AIMS

This study aimed to compare the accuracy of SoloSTAR(sanofi-aventis, Paris, France) and FlexPen (Novo Nordisk, Bagsvaerd, Denmark) to deliver doses of insulin when used by insulin/device-naive people attending a local healthcare practice.

METHODS

For the determination of dose accuracy, SoloSTAR containing insulin glargine (lot 672) and FlexPen containing insulin aspart (lot SH70557) were used for all tests. The participants were given instruction by an independent monitor on how to use the pens. To be consistent, users were trained to hold the pens in situ for 10 s at the end of injection to ensure that the full dose was injected. Forty-eight subjects performed three dose deliveries of 20 units (into a sponge) with each pen type. The method of injecting into a sponge aimed to mimic real-life practice; this approach is commonly used to ensure the patient has a correct injection technique before the patient self-administers the injection.

RESULTS

The delivery (n = 144) of individual doses of 20 units was not statistically different (P = 0.187) between SoloSTAR (mean +/- SD, 19.75 +/- 0.30 units) and FlexPen (19.80 +/- 0.33 units). In total, 2% of doses from both devices were <19 units; 98% were within 19-21 units.

CONCLUSIONS

Both SoloSTAR and FlexPen were similarly accurate when used by device-naive subjects to deliver 20-unit doses of insulin. This is likely to have beneficial clinical effects in terms of glycemic control, as well as improved patient confidence that insulin pens accurately deliver doses of insulin.