Type 1 diabetes mellitus is the more severe form of diabetes mellitus. It results from a cellular-mediated autoimmune destruction of the beta-cells of the pancreas, which are responsible for the secretion of insulin that is required for a proper blood glucose regulation. Patients with type 1 diabetes need exogenous insulin injections to keep the glucose concentration in the safe range. Artificial pancreas is an autonomous system for closed-loop blood glucose control in subjects affected by type 1 diabetes. The core of the system is the control algorithm, which receives blood glucose data from the sensor, computes the required insulin amount and transmits this information to the insulin pump. Regarding the control algorithm, one of the most promising approach revealed to be the model predictive control algorithm, which exploits a glucose-insulin model of the patient to predict near-future blood glucose values and, consequently, computes the optimal insulin dose. The performance of the control algorithm is highly influenced by the quality of the model used for prediction. Moreover, the inter-patient variability characterizing subjects with T1D increases the need of patient-tailored models. Since promising results have been obtained in silico using the UVA/Padova simulator, the aim of this thesis is to investigate and test the applicability of the identification techniques on free-living data collected without ad hoc clinical protocols thanks to the availability of long term trials. The individualized models show superior prediction performance with respect to the average model that was used to synthetize the controller used during the trial. This result pushed towards a detailed data analysis to improve model identification. A multiple-model predictor with different identified models on the basis of the data analysis is proposed in this thesis. The prediction capabilities are improved if compared to the performance of a predictor built using a single model identified on a daily subset. These results represent a milestone in the development for a new generation of individualized controllers for artificial pancreas. The patient-tailored models can be exploited to predict this risk of hypoglycemia in advance and therefore to alert the patient on the risk. In order to improve the safety of the artificial pancreas system, the identified models have been evaluated in terms of hypoglycemia prevention by showing that these models are able to detect 84.53% of the hypoglycemia events occurred during a 1-month trial on average. In this thesis, model identification has been addressed by deep learning techniques, by showing that the proposed architecture obtains state-of-the-art performance on both in silico and in vivo data, considering several prediction horizons. Finally, since the post-prandial glucose regulation remains a challenging issue for diabetes treatment, machine learning methodologies have been applied in order to improve the postprandial glucose regulation. Two algorithms are proposed to provide corrections to time and/or amount of the meal bolus. They have been tested on the in silico virtual population of the UVA/Padova simulator by showing the reduction of both hypoglycemia and hyperglycemia.

Personalized Artificial Pancreas: from identification to optimal bolus computation

AIELLO, ELEONORA MARIA
2020

Abstract

Type 1 diabetes mellitus is the more severe form of diabetes mellitus. It results from a cellular-mediated autoimmune destruction of the beta-cells of the pancreas, which are responsible for the secretion of insulin that is required for a proper blood glucose regulation. Patients with type 1 diabetes need exogenous insulin injections to keep the glucose concentration in the safe range. Artificial pancreas is an autonomous system for closed-loop blood glucose control in subjects affected by type 1 diabetes. The core of the system is the control algorithm, which receives blood glucose data from the sensor, computes the required insulin amount and transmits this information to the insulin pump. Regarding the control algorithm, one of the most promising approach revealed to be the model predictive control algorithm, which exploits a glucose-insulin model of the patient to predict near-future blood glucose values and, consequently, computes the optimal insulin dose. The performance of the control algorithm is highly influenced by the quality of the model used for prediction. Moreover, the inter-patient variability characterizing subjects with T1D increases the need of patient-tailored models. Since promising results have been obtained in silico using the UVA/Padova simulator, the aim of this thesis is to investigate and test the applicability of the identification techniques on free-living data collected without ad hoc clinical protocols thanks to the availability of long term trials. The individualized models show superior prediction performance with respect to the average model that was used to synthetize the controller used during the trial. This result pushed towards a detailed data analysis to improve model identification. A multiple-model predictor with different identified models on the basis of the data analysis is proposed in this thesis. The prediction capabilities are improved if compared to the performance of a predictor built using a single model identified on a daily subset. These results represent a milestone in the development for a new generation of individualized controllers for artificial pancreas. The patient-tailored models can be exploited to predict this risk of hypoglycemia in advance and therefore to alert the patient on the risk. In order to improve the safety of the artificial pancreas system, the identified models have been evaluated in terms of hypoglycemia prevention by showing that these models are able to detect 84.53% of the hypoglycemia events occurred during a 1-month trial on average. In this thesis, model identification has been addressed by deep learning techniques, by showing that the proposed architecture obtains state-of-the-art performance on both in silico and in vivo data, considering several prediction horizons. Finally, since the post-prandial glucose regulation remains a challenging issue for diabetes treatment, machine learning methodologies have been applied in order to improve the postprandial glucose regulation. Two algorithms are proposed to provide corrections to time and/or amount of the meal bolus. They have been tested on the in silico virtual population of the UVA/Padova simulator by showing the reduction of both hypoglycemia and hyperglycemia.
26-feb-2020
Inglese
TOFFANIN, CHIARA
Università degli studi di Pavia
File in questo prodotto:
File Dimensione Formato  
ThesisAiello.pdf

Open Access dal 31/08/2020

Dimensione 5.82 MB
Formato Adobe PDF
5.82 MB Adobe PDF Visualizza/Apri

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14242/86333
Il codice NBN di questa tesi è URN:NBN:IT:UNIPV-86333