Diabetes mellitus is a chronic disease of disordered glucose metabolism due to defects in either insulin secretion by the pancreatic cells or insulin action. The basic effect of insulin lack or insulin resistance is to prevent the efficient uptake and utilization of glucose by most cells of the body, mainly the skeletal muscles, resulting in abnormally high blood sugar levels (hyperglycemia). Sustained hyperglycemia is associated with various health complications which are irreversibile once they develop and can mean serious disability for the person who experiences them. Already one of the most widespread diseases, the incidence of diabetes is increasing at an alarming rate on a global scale. During my graduate studies at Lund University, I participated in the European FP7-IST research project DIAdvisor. My objective was the development of a personalized blood glucose predicting system and an advisory control system, the DIAdvisor tool, to be used on the spot by the user in different daily situations, predicting glycemic excursions following meals, insulin intakes and exercise and giving them advices about how to adjust their treatments. Under the aegis of the project, acquisition of bio-clinical data linked or potentially involved in blood glucose control from insulin-treated diabetic subjects was accomplished in three separate clinical trials taking place in Italy, France and the Czech Republic. The experiments provided me with a wealth of subjects data, including specific patient parameters (e.g., gender, age, BMI, weight), characteristics related to diabetes (e.g., disease duration, insulin delivery), associated health conditions and therapies, food intakes and administered insulin doses registered in a logbook, capillary glucose strips, interstitial glucose levels, plasma glucose and plasma insulin concentration from drawn blood samples as well as vital signs from wearable sensors. I mined the data, extracting knowledge of the glucoregulatory system in type 1 diabetes. First of all, I performed exploratory data analysis to obtain some of the most interesting features of the datasets such as the distribution of blood glucose measurements, their time variability, the blood-to-interstitial lag and correlations between data. Next, I investigated the use of data-driven techniques to the purpose of personalized glucose metabolism modelling and short-term blood-glucose predictions. The approaches I took to describe the glucose metabolism were discrete-time and continuous time models on input-output form and state-space form, while the blood glucose short-term predictors, i.e., up to 120 minutes ahead, used ARX-, ARMAX- and subspace-based prediction. I also considered the problem of estimating multi-step-ahead linear blood glucose predictors directly from data, without any prior knowledge of the underlying mechanism generating the time series. Some of the algorithms I produced during my PhD were implemented in the actual DIAdvisor tool and underwent clinical trials. In-vivo tests on a population of 43 subjects following at least 70% of the DIAdvisor insulin therapy recommendations showed an increase by 7.6% in the time spent in normo-glycemia and a reduction by 42% in the time spent in hypoglycemia. In addition to impact within the academic and technology in diabetes communities, our project was featured on the Futuris column of Euronews and had popular press coverage in Ehealth News, EE Times  as well as national swedish magazines.

My work on diabetes care continued at the University of California, Santa Barbara, in the biomedical research group led by Prof. Frank Doyle. At that time, a clinical trial performed jointly at Stanford School of Medicine and Barbara Davis Center for Diabetes had just been concluded. The study had been specifically designed to allow investigation of actuator faults in insulin pump therapy. I took advantage of my previous training in diabetes blood glucose dynamics modelling and the accrued knowledge about the underlying physiology to develop a novel method for early detection of insulin infusion set failures. I set the problem in a performance monitoring framework, moving away from previous works on the topic. I designed an algorithm able to capture the degradation of control performances and hence the actuator faults, using solely measurements from the subjects glucose sensor and insulin pump. The method was evaluated retrospectively and demonstrated favourable results detecting the onset of faulty events before they become severe with enough time to react to them.