IMPORTANCE OF IDENTIFIABILITY ANALYSIS IN RELIABLE INDIVIDUALIZED MODEL FOR USE IN ARTIFICIAL PANCREAS SYSTEMS

• Format: Web Resource
• Language: English
• Published: Morressier 01.01.2017
• DOI: 10.26226/morressier.5c3c8159e668b9000b9f8f7a

Objective:To show the importance of robust identifiability analysis on personalized glucose forecasting.Background:Standard methods for model identification usually assume population values or direct identification without analysis of parameter identifiability. Neither of these methods produces a reliable patient-specific set of parameter values since the impact of the model structure or the amount and quality of the available data is not explored.Method:The Subcutaneous Oral Glucose Minimal Model is derived from the well-known minimal model of glucose-insulin kinetics that has been extensively used in Artificial Pancreas Systems. SOGMM has 13 parameters of which 2 are a priori known. Model individualization included correlation analysis among parameters, global parameter ranking, and structural and practical identifiability analysis to define a final unique set of 5 parameters. Data from 2 clinical trials (NCT02137512 and NCT02558491), were used for analysis and parameters identification. Our method was compared to the identification of all 11 model parameters, using the same optimization procedure. Data from 6 subjects was separated in 6-hour non-overlapping intervals and 2 steps forecast was used to assess quality of fit via root mean square error (RMSE). Result:Average RMSE value was improved: 18.8u00b116.75mg/dL vs 27.5u00b127.46mg/dL. Maximum errors in glucose forecast over all cases were 32.9mg/dL vs. 102.2mg/dL.Conclusion:Methods not accounting for robust identifiability lack the precision of parameter estimation necessary for accurate forecast. On the contrary, robust identifiability analysis can be associated not only with a reduction of computational burden, but also with an improvement in the accuracy of glucose predictions.