Personalized Hybrid Models for Exercise, Meal, and Insulin Interventions in Type 1 Diabetic Children and Adolescents Naviyn Prabhu Balakrishnan, Lakshminarayanan Samavedham, and Gade Pandu Rangaiah* Department of Chemical and Biomolecular Engineering, National University of Singapore, Singapore ABSTRACT: Inter- and intrapatient variability in blood glucose (BG) metabolism imposes the need for personalized models in diabetes care. Validation of personalized models using the clinical data of type 1 diabetic (T1D) subjects and inclusion of lifestyle factors like exercise as an input in these personalized models are rarely seen in the literature. In this paper, we have developed personalized BG prediction models with a specialized structure comprising three different classes of models: a mechanistic model for meal absorption dynamics, an empirical model for insulin absorption kinetics and a transfer function model for prediction of personalized BG dynamics. Hence, the proposed model structure is termed a hybrid model (HM). Exercise intensity in these personalized HMs is quantified using a measure called rate of perceived exertion (RPE). The clinical data of 34 T1D subjects are used for model development and two different scenarios of cross validationsame day validation (SDV) and different day validation (DDV). The BG data collected during one of the clinical visits have been used for model development (85-90% of data of this visit) and SDV (the remaining 10-15% data of the same visit); while the data collected during the other day visit have been used for DDV. This is to test the ability of developed HMs in predicting the BG dynamics for a prolonged time period, thereby ensuring their potential to capture intrapatient variability. The fitness and cross validation results of personalized HMs not only show accurate prediction of BG dynamics but also reveal the potential of HMs in capturing both inter- and intrapatient variability. 1. INTRODUCTION Modern health care is gradually abandoning the conventional “one-size-fits-all” therapy and rapidly embracing a personalized approach in the treatment of many alarming diseases. The personalized medicine market in the United States alone is estimated to grow to more than $450 billion by the year 2015. 1 Such a major transformation would not have happened without advancements in measurement technology and application of various mathematical modeling techniques that catalyzed an improved understanding of many diseases. 2 Although there are many definitions for personalized medicine; the widely accepted definition 1 is “the right treatment for the right person at the right time”. Hence, a personalized model that can predict the unique patient dynamics at different periods of time can help physicians in devising safe and efficient treatment strategies in less time and cost ultimately leading to better quality of life for the patients. The two well-established products of personalized medicine that have been prevailing for years in core health care are ABO blood grouping and family history (genetic evaluation tool). 3 Incorporating the concept of personalization in diabetes care is important because of the inter- and intrapatient variations in blood glucose (BG) metabolism. Development of personalized BG prediction models for diabetic subjects is practically possible because of the two major things mentioned above the advancements in the BG sensor technology and availability of different BG prediction modeling techniques in the literature. The availability of modern sensors like continuous glucose measurement sensors (CGMS) and GlucoWatch (GW) has made frequent measurements of BG with sampling intervals of 5 min possible. Development of a personalized BG prediction model for each patient can help physicians and healthcare workers: (i) to accurately forecast the BG dynamics of a patient for various interventions, like insulin, meal, exercise, and stress levels; (ii) to devise optimal insulin, meal, and exercise intervention strategies that can help in maintaining the BG level of patients within therapeutic limits for a prolonged period of time; (iii) to educate the patients and their relatives regarding the adverse effects of protocol deviations on BG dynamics, ultimately leading to increased participation of patients and their relatives in strictly following the physician’s prescriptions, which in turn can prevent various long and short- term diabetic complications; in fact, such diabetic education is strongly recommended by the International Diabetes Feder- ation as evidence-based counselling. 4,5 The mathematical models developed for predicting BG dynamics in diabetic subjects can be broadly categorized into two classes: (i) knowledge-driven models (KDMs) and (ii) data-driven models (DDMs). The former class models (also called as mechanistic or first-principle models) are developed on the basis of the physiological knowledge behind the glucose-insulin regulatory mechanism of diabetics, 6-11 where- as the latter class models (also called as empirical or black box or correlation-based models) are developed on the basis of the data only. 12-15 Our recent review article gives an extensive summary and analysis of both KDMs and DDMs that have been developed so far to predict the BG dynamics in type 1 diabetics (T1Ds). 16 In the case of type 2 diabetics (T2Ds), a review on various KDMs available in the literature can be found in Landersdorfer et al. 17 There are also a few studies focusing Received: February 27, 2013 Revised: August 7, 2013 Accepted: August 10, 2013 Published: August 11, 2013 Article pubs.acs.org/IECR © 2013 American Chemical Society 13020 dx.doi.org/10.1021/ie402531k | Ind. Eng. Chem. Res. 2013, 52, 13020-13033