The OhioT1DM Dataset for Blood Glucose Level Prediction: Update 2020 Cindy Marling and Razvan Bunescu 1 Abstract. This paper documents the OhioT1DM Dataset, which Table 1. Gender, age range, insulin pump model, and sensor band type for was developed to promote and facilitate research in blood glucose each data contributor, by cohort level prediction. It contains eight weeks’ worth of continuous glu- cose monitoring, insulin, physiological sensor, and self-reported life- ID Gender Age Pump Model Sensor Band Cohort event data for each of 12 people with type 1 diabetes. An associated graphical software tool allows researchers to visualize the integrated 540 male 20–40 630G Empatica 2020 data. The paper details the contents and format of the dataset and tells 544 male 40–60 530G Empatica 2020 interested researchers how to obtain it. 552 male 20–40 630G Empatica 2020 The OhioT1DM Dataset was first released in 2018 for the first 567 female 20–40 630G Empatica 2020 Blood Glucose Level Prediction (BGLP) Challenge. At that time, the 584 male 40–60 530G Empatica 2020 dataset was half its current size, containing data for only six people with type 1 diabetes. Data for an additional six people is being re- 596 male 60–80 530G Empatica 2020 leased in 2020 for the second BGLP Challenge. This paper subsumes 559 female 40–60 530G Basis 2018 and supersedes the paper which documented the original dataset. 563 male 40–60 530G Basis 2018 570 male 40–60 530G Basis 2018 1 INTRODUCTION 575 female 40–60 530G Basis 2018 588 female 40–60 530G Basis 2018 Accurate forecasting of blood glucose levels has the potential to im- 591 female 40–60 530G Basis 2018 prove the health and wellbeing of people with diabetes. Knowing in advance when blood glucose is approaching unsafe levels provides time to proactively avoid hypo- and hyper-glycemia and their con- comitant complications. The drive to perfect an artificial pancreas The dataset includes: a CGM blood glucose level every 5 minutes; [2] has increased the interest in using machine learning (ML) ap- blood glucose levels from periodic self-monitoring of blood glucose proaches to improve prediction accuracy. Work in this area has been (finger sticks); insulin doses, both bolus and basal; self-reported meal hindered, however, by a lack of real patient data; some researchers times with carbohydrate estimates; self-reported times of exercise, have only been able to work on simulated patient data. sleep, work, stress, and illness; and data from the Basis Peak or Em- To promote and facilitate research in blood glucose level predic- patica Embrace band. The data for individuals who wore the Basis tion, we have curated the OhioT1DM Dataset and made it publicly Peak band includes 5-minute aggregations of heart rate, galvanic skin available for research purposes. To the best of our knowledge, this response (GSR), skin temperature, air temperature, and step count. is the first publicly available dataset to include continuous glucose The data for those who wore the Empatica Embrace band includes monitoring, insulin, physiological sensor, and self-reported life-event 1-minute aggregations of GSR, skin temperature, and magnitude of data for people with type 1 diabetes. acceleration. Both bands indicated the times they detected that the The OhioT1DM Dataset contains eight weeks’ worth of data for wearer was asleep, and this information is included when available. each of 12 people with type 1 diabetes. These anonymous people However, not all data contributors wore their sensor bands overnight. are referred to by randomly selected ID numbers. All data contrib- Data for the first six individuals was released in 2018 for the first utors were on insulin pump therapy with continuous glucose moni- Blood Glucose Level Prediction (BGLP) Challenge, which was held toring (CGM). They wore Medtronic 530G or 630G insulin pumps in conjunction with the 3rd International Workshop on Knowledge and used Medtronic Enlite CGM sensors throughout the 8-week data Discovery in Healthcare Data, at IJCAI-ECAI 2018, in Stockholm, collection period. They reported life-event data via a custom smart- Sweden. Data for six additional people is being released in 2020 for phone app and provided physiological data from a fitness band. The the second BGLP Challenge, to be held at the 5th International Work- first cohort of six individuals wore Basis Peak fitness bands. Data for shop on Knowledge Discovery in Healthcare Data, at ECAI 2020, in this cohort was released in 2018. The second cohort of six individ- Santiago de Compostela, Spain. This paper subsumes and supersedes uals wore the Empatica Embrace. Data for this cohort is included in the paper which documented the original 2018 dataset [3]. In order the 2020 release. Table 1 shows the gender, age range, insulin pump to provide a unified overview of the entire dataset, this paper incor- model, and sensor band type for each data contributor, by cohort. porates most of the original paper verbatim. The following sections of this paper provide background informa- 1 Ohio University, USA, email: {marling,bunescu}@ohio.edu tion, detail the data format, describe the OhioT1DM Viewer visual- Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). ization software, and tell how to obtain the OhioT1DM Dataset and Each XML file contains the following data fields: Viewer for research purposes. 1. The patient ID number and insulin type. Weight is set 2 BACKGROUND to 99 as a placeholder, as actual patient weights are unavailable. 2. Continuous glucose monitoring (CGM) data, We have been working on intelligent systems for diabetes manage- recorded every 5 minutes. ment since 2004 [1, 4, 5, 6, 7, 8, 10, 11]. As part of our work, we 3. Blood glucose values obtained through self- have run five clinical research studies involving subjects with type 1 monitoring by the patient. diabetes on insulin pump therapy. Over 50 anonymous subjects have 4. The rate at which basal insulin is continuously infused. provided blood glucose, insulin, and life-event data so that we could The basal rate begins at the specified timestamp ts, and it continues develop software intended to help people with diabetes and their pro- until another basal rate is set. fessional health care providers. 5. A temporary basal insulin rate that supersedes Our most recent study was designed so that de-identified data the patient’s normal basal rate. When the value is 0, this indi- could be shared with the research community. All data contributors to cates that the basal insulin flow has been suspended. At the end the OhioT1DM Dataset signed informed consent documents allow- of a temp basal, the basal rate goes back to the normal basal rate, ing us to share their de-identified data with outside researchers. This agreement clearly delineated what types of data could be shared and 6. Insulin delivered to the patient, typically before a meal with whom. The data in the dataset was fully de-identified according or when the patient is hyperglycemic. The most common type of to the Safe Harbor method, a standard specified by the Health Insur- bolus, normal, delivers all insulin at once. Other bolus types can ance Portability and Accountability Act (HIPAA) Privacy Rule [9]. stretch out the insulin dose over the period between ts begin and To protect the data contributors and to ensure that the data is used ts end. only for research purposes, a Data Use Agreement (DUA) must be 7. The self-reported time and type of a meal, plus the pa- executed before a researcher can obtain the data. tient’s carbohydrate estimate for the meal. 8. The times of self-reported sleep, plus the patient’s sub- 3 OhioT1DM DATA FORMAT jective assessment of sleep quality: 1 for Poor; 2 for Fair; 3 for Good. For each data contributor, there is one XML file for training and de- 9. Self-reported times of going to and from work. Intensity velopment data and a separate XML file for testing data. This results is the patient’s subjective assessment of physical exertion, on a in a total of 24 XML files, two for each of the 12 contributors. Table 2 scale of 1 to 10, with 10 the most physically active. shows the number of training and test examples for each contributor. 10. Time of self-reported stress. Table 2 also indicates the BGLP Challenge for which the data was 11. Time of self-reported hypoglycemic episode. released. For the 2018 BGLP Challenge, the number of test examples Symptoms are not available, although there is a slot for them in was equal to the number of data points in the XML testing file. How- the XML file. ever, for the 2020 BGLP Challenge, the first hour of data in each 12. Time of self-reported illness. XML testing file is excluded from the set of points used for eval- 13. Time and duration, in minutes, of self-reported ex- uation. This is to allow unbiased comparison of prediction models ercise. Intensity is the patient’s subjective assessment of physical using all training data to predict each test point, as the first test points exertion, on a scale of 1 to 10, with 10 the most physically active. would otherwise be too close chronologically to the training data. 14. Heart rate, aggregated every 5 minutes. This Thus, for the 2020 BGLP Challenge, there are 12 more data points data is only available for people who wore the Basis Peak sensor in each XML testing file than the number of test examples shown in band. Table 2. 15. Galvanic skin response, also known as skin con- ductance or electrodermal activity. For those who wore the Ba- Table 2. Number of training and test examples per data contributor sis Peak, the data was aggregated every 5 minutes. Despite this attribute’s name, it is also available for those who wore the Em- BGLP Training Test patica Embrace. For these individuals, the data is aggregated every ID Challenge Examples Examples 1 minute. 540 2020 11947 2884 16. Skin temperature, in degrees 544 2020 10623 2704 Fahrenheit, aggregated every 5 minutes, for those who wore the 552 2020 9080 2352 Basis Peak, and every 1 minute, for those who wore the Empatica 567 2020 10858 2377 Embrace. 17. Air temperature, in degrees Fahren- 584 2020 12150 2653 heit, aggregated every 5 minutes. This data is only available for 596 2020 10877 2731 people who wore the Basis Peak sensor band. 559 2018 10796 2514 18. Step count, aggregated every 5 minutes. This data 563 2018 12124 2570 is only available for people who wore the Basis Peak sensor band. 570 2018 10982 2745 19. Times when the sensor band reported that the sub- ject was asleep. For those who wore the Basis Peak, there is also 575 2018 11866 2590 a numeric estimate of sleep quality. 588 2018 12640 2791 20. Magnitute of acceleration, aggregated every 1 591 2018 10847 2760 minute. This data is only available for people who wore the Em- patica Embrace sensor band. Note that, in de-identifying the dataset, all dates for each individ- of people with diabetes. In addition to their role in the artificial pan- ual were shifted by the same random amount of time into the future. creas project, such predictions could also enable other beneficial ap- The days of the week and the times of day were maintained in the plications, such as decision support for avoiding impending prob- new timeframes. However, the months were shifted, so that it is not lems, “what if” analyses to project the effects of different lifestyle possible to consider the effects of seasonality or of holidays. choices, and enhanced blood glucose profiles to aid in individualiz- ing diabetes care. It is our hope that sharing this dataset will help to advance the state of the art in blood glucose level prediction. 4 THE OhioT1DM VIEWER The OhioT1DM Viewer is a visualization tool that opens an XML ACKNOWLEDGEMENTS file from the OhioT1DM Dataset and graphically displays the inte- This work was supported by grant 1R21EB022356 from the National grated data. It aids in developing intuition about the data and also Institutes of Health (NIH). The OhioT1DM Viewer was originally in debugging. For example, if a system makes a poor blood glucose implemented by Hannah Quillin and Charlie Murphy, and further re- level prediction at a particular point in time, viewing the data at that fined by Robin Kelby and Jeremy Beauchamp. The authors gratefully time might illuminate a cause. For example, the subject might have acknowledge the contributions of Emeritus Professor of Endocrinol- forgotten to report a meal or might have been feeling ill or stressed. ogy Frank Schwartz, MD, a pioneer in building intelligent systems Figure 1 shows a screenshot from the OhioT1DM Viewer. The for diabetes management. We would also like to thank our physician data is displayed one day at a time, from midnight to midnight. Con- collaborators, Aili Guo, MD, and Amber Healy, DO, our research trols allow the user to move from day to day and to toggle any type nurses, Cammie Starner and Lynn Petrik, and our past and present of data off or on for targeted viewing. graduate and undergraduate research assistants. We are especially The bottom pane shows blood glucose, insulin, and self-reported grateful to the 12 anonymous individuals with type 1 diabetes who life-event data. CGM data is displayed as a mostly blue curve, with shared their data, enabling the creation of this dataset. green points indicating hypoglycemia. Finger sticks are displayed as red dots. Boluses are displayed along the horizontal axis as orange and yellow circles. The basal rate is indicated as a black line. Tem- REFERENCES porary basal rates appear as red lines. Self-reported sleep is indicated [1] R. Bunescu, N. Struble, C. Marling, J. Shubrook, and F. Schwartz, by blue regions. Life-event icons appear at the top of the pane as ‘Blood glucose level prediction using physiological models and support vector regression’, in Proceedings of the Twelfth International Confer- dots, squares, and triangles. The data in the bottom pane is clickable, ence on Machine Learning and Applications (ICMLA), pp. 135–140. so that additional information about any data point can be displayed. IEEE Press, (2013). For example, clicking on a meal (a square blue icon) displays the [2] Juvenile Diabetes Research Foundation (JDRF). Artificial Pan- timestamp, type of meal, and carbohydrate estimate. creas, 2019. Available at http://www.jdrf.org/impact/research/artificial- pancreas/, accessed January, 2020. The top pane displays sensor band data. Blue regions in the top [3] C. Marling and R. Bunescu, ‘The OhioT1DM dataset for blood glu- pane are times the sensor band detected that the wearer was asleep. cose level prediction’, in The 3rd International Workshop on Knowl- The step count is indicated by vertical blue lines. The curves show edge Discovery in Healthcare Data, Stockholm, Sweden, (July 2018). heart rate (red), galvanic skin response (green), skin temperature Available at http://ceur-ws.org/Vol-2148/paper09.pdf, accessed Jan- (gold), air temperature (cyan), and magnitude of acceleration (black). uary, 2020. [4] C. Marling, J. Shubrook, and F. Schwartz, ‘Toward case-based rea- soning for diabetes management: A preliminary clinical study and de- cision support system prototype’, Computational Intelligence, 25(3), 5 OBTAINING THE DATASET AND VIEWER 165–179, (2009). [5] C. Marling, M. Wiley, R. Bunescu, J. Shubrook, and F. Schwartz, The original 2018 OhioT1DM Dataset and the OhioT1DM Viewer ‘Emerging applications for intelligent diabetes management’, AI Mag- are now available to researchers. The full 2020 OhioT1DM Dataset azine, 33(2), 67–78, (2012). is currently being released to participants in the second BGLP Chal- [6] C. Marling, L. Xia, R. Bunescu, and F. Schwartz, ‘Machine learning lenge. The second BGLP Challenge will take place June 9, 2020, in experiments with noninvasive sensors for hypoglycemia detection’, in IJCAI 2016 Workshop on Knowledge Discovery in Healthcare Data, conjunction with the 5th International Workshop on Knowledge Dis- New York, NY, (2016). covery in Healthcare Data at ECAI 2020, in Santiago de Compostela, [7] S. Mirshekarian, R. Bunescu, C. Marling, and F. Schwartz, ‘Using Spain. LSTMs to learn physiological models of blood glucose behavior’, in After the completion of the BGLP Challenge, the entire dataset Proceedings of the 39th International Conference of the IEEE Engi- neering in Medicine and Biology Society (EMBC 2017), Jeju Island, will be made available to other researchers. To protect the data con- Korea, (2017). tributors and to ensure that the data is used only for research pur- [8] S. Mirshekarian, H. Shen, R. Bunescu, and C. Marling, ‘LSTMs and poses, a Data Use Agreement (DUA) is required. A DUA is a bind- neural attention models for blood glucose prediction: Comparative ex- ing document signed by legal signatories of Ohio University and periments on real and synthetic data’, in Proceedings of the 41st Inter- the researcher’s home institution. As of this writing, researchers national Conference of the IEEE Engineering in Medicine and Biology Society (EMBC 2019), Berlin, Germany, (2019). can request a DUA at http://smarthealth.cs.ohio.edu/OhioT1DM- [9] Office for Civil Rights. Guidance regarding methods for de- dataset.html. Once a DUA is executed, the OhioT1DM Dataset and identification of protected health information in accordance with the Viewer will be directly released to the researcher. Health Insurance Portability and Accountability Act (HIPAA) privacy rule, 2012. Available at https://www.hhs.gov/sites/default/files/ocr /privacy/hipaa/understanding/coveredentities/De-identification 6 CONCLUSION /hhs deid guidance.pdf, accessed January, 2020. [10] K. Plis, R. Bunescu, C. Marling, J. Shubrook, and F. Schwartz, ‘A ma- chine learning approach to predicting blood glucose levels for diabetes The OhioT1DM Dataset was developed to promote and facilitate re- management’, in Modern Artificial Intelligence for Health Analytics: search in blood glucose level prediction. Accurate blood gluocose Papers Presented at the Twenty-Eighth AAAI Conference on Artificial level predictions could positively impact the health and well-being Intelligence, pp. 35–39. AAAI Press, (2014). Figure 1. Screenshot from the OhioT1DM Viewer [11] F. L. Schwartz, J. H. Shubrook, and C. R. Marling, ‘Use of case-based reasoning to enhance intensive management of patients on insulin pump therapy’, Journal of Diabetes Science and Technology, 2(4), 603–611, (2008).