Poster (Scientific congresses and symposiums)
Improved Blood Glucose Forecasting Models using Changes in Insulin Sensitivity in Intensive Care Patients
Uyttendaele, Vincent; Dickson, Jennifer; Shaw, Geoff et al.
2017GIGA-DAY 2017
 

Files


Full Text
Poster_final.pdf
Publisher postprint (781.43 kB)
Download

All documents in ORBi are protected by a user license.

Send to



Details



Keywords :
Insulin sensitivity; Intensive care; Glycaemic control
Abstract :
[en] Introduction: Hyperglycaemia, hypoglycaemia and glycaemic variability are associated with worsened outcomes and increased mortality in intensive care units. Glycaemic control (GC) using insulin therapy has shown improved outcomes, but have been proven difficult to repeat or achieve safely. STAR (Stochastic TARgeted) is a model-based glycaemic control protocol using a stochastic model to forecast distributions of likely future changes in insulin sensitivity (SI) based on its current value. This can be used to determine likely future blood glucose (BG) levels for a given intervention, enabling the most optimal dose selection that best overlaps a clinically defined BG target band. This study presents a novel 3D model capable to predict likely future distribution of SI using both current SI and its prior variability (%ΔSI). Methods: Metabolic data from 3 clinical ICU cohorts totalling 819 episodes and 68629 hours of treatment under STAR and SPRINT protocols are used in this study. Data triplets (%ΔSIn, SIn, SIn+1) are created and binned together in a range of %ΔSI = [-100%, 200%] and SIn = [1.0e-7, 2.1e-3] in bin sizes of %ΔSI = 10% and SIn = 0.5e-4. The 5th, 50th, and 95th percentile of SIn+1 are determined for each bin where data density is high enough (>100 triplets) and compared to the previous stochastic model. The predictive power of the two models are compared by computing median [IQR] per-patient percentage prediction of SI within the 5th-95th and 25th-75th percentile ranges of model predictions. Results: Results show the previous model is over-conservative for ~77% of the data, mainly where %ΔSI is within an absolute 25% change. The percentage change in the 90% CI width in this region is reduced by ~25-40%. Conversely, non-conservative regions are also identified, with 90% CI width increased up to ~80%. Predictive power is similar for both model (60.3% [47.8%, 71.5%] vs. 51.2 [42.9%, 59.2%] within 25th-75th and 93.6% [85.7%, 97.3%] vs. 90.7% [84.4%, 94.6%] within 5th-95th range). Conclusions: The new 3D model achieved similar predictive power as the previous model by reducing the 5th-95th percentile prediction range for 77% of the data, predominantly where SI is stable. If the conservatism of the previous model reduces risk of hypoglycaemia, it also inhibits the controller’s ability to reduce BG to the normal range by safely using more aggressive dosing. The 3D new model thus better characterises patient-specific response to insulin, and allows more optimal dosing, increasing performance and safety.
Disciplines :
Anesthesia & intensive care
Author, co-author :
Uyttendaele, Vincent ;  Université de Liège - ULiège > Form. doct. sc. ingé. & techno. (aéro. & mécan. - Paysage)
Dickson, Jennifer
Shaw, Geoff
Desaive, Thomas  ;  Université de Liège > Département d'astrophys., géophysique et océanographie (AGO) > Thermodynamique des phénomènes irréversibles
Chase, Geoffrey
Language :
English
Title :
Improved Blood Glucose Forecasting Models using Changes in Insulin Sensitivity in Intensive Care Patients
Publication date :
01 February 2017
Event name :
GIGA-DAY 2017
Event organizer :
GIGA - University of Liege
Event place :
Liège, Belgium
Event date :
1/2/2017
Available on ORBi :
since 06 March 2017

Statistics


Number of views
84 (11 by ULiège)
Number of downloads
30 (0 by ULiège)

Bibliography


Similar publications



Contact ORBi