Creative Commons source reading · Explainable decision support
Sugar slay: a gamified decision support ecosystem for type 1 diabetes
Sundararaman Rengarajan; Nicholas Abrams; Aspen Tabar; Hariharan Sundaram; Kavya Pratap Singh; Leanne Chukoskie
Originally published in Front Digit Health · 2026-06-17
From the article
Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.
Abstract
Background/IntroductionType 1 Diabetes (T1D) management demands consistent attention to blood glucose levels, insulin dosing, physical activity, sleep, and diet-tasks that are especially burdensome for adolescents and young adults navigating new independence. While continuous glucose monitoring (CGM) systems and wearable fitness devices provide real-time physiological data, the cognitive load of interpreting this information and maintaining consistent self-care behaviors remains a significant barrier for patients and their support networks.MethodsWe developed the Sugar Slay ecosystem, comprising a gamified mobile decision support application for T1D individuals and a companion application, Sugar Slay Care, for caregivers and supporters. The main application integrates CGM and wearable device data to generate real-time, personalized…
Results
This section presents the results from our machine learning model evaluation study using the OhioT1DM dataset. We systematically compared multiple deep learning architectures to identify the best performing model for glucose prediction in the Sugar Slay application. To contextualize the error metrics reported throughout this section: Root Mean Square Error (RMSE) quantifies the average magnitude of prediction errors while giving greater weight to large deviations, making it particularly sensitive to dangerous prediction failures in glucose forecasting. Mean Absolute Error (MAE) provides a complementary measure of average error magnitude without emphasizing outliers. Prior glucose forecasting studies report RMSE values in the range of approximately 18–22 mg/dL for 30 min predictions and around 30–40 mg/dL for 60 min horizons, with…
Summary of machine learning results
Comprehensive quantitative benchmarking across LSTM, BiLSTM, TCN, ConvLSTM, and sequence-to-sequence variants was performed using 30 min and 60 min prediction horizons. RMSE and MAE values for all models are presented in the corresponding results tables, supporting the selection of the Seq2Seq BiLSTM architecture as the best-performing model. The sequence-to-sequence BiLSTM model emerged as the best-performing choice for glucose prediction in the Sugar Slay application, demonstrating: Consistent performance across all test subjectsClinically acceptable error rates for both 30 min and 60 min predictions (59)Superior ability to capture bidirectional temporal dependencies in glucose dataRobust performance without requiring complex transfer learning approachesThese results provide strong evidence for integrating the Seq2Seq BiLSTM model…
Editorial reading guide
- What data created and tested the model, and was evaluation independent?
- What is the cost of a false positive, false negative or out-of-distribution decision?
