Study and analysis of insulin and glycaemic data for the development of user calibration optimisation of t:slim x2 insulin pump in children and adolescents

The present thesis reports the study and analysis focused on the insulin and glycaemic data of paediatric patients affected by Type 1 Diabetes Mellitus (T1DM) that use a hybrid closed-loop system called t:slim X2™. The Tandem™ insulin pump implements the Control-IQ™ technology that personalises the basal insulin control through the setting of 3 Personal Profile parameters: the basal rate (BR), the insulin to carbohydrate ratio and the insulin sensitivity factor. The ultimate goal of this work is to fully automate this parameters' selection. T1DM is a metabolic disorder linked to total insulin deficiency caused by the autoimmune destruction of the pancreatic β cells of Langherans. It usually attacks children and young adults and requires life-long monitoring and it can cause short- and long-term complications. The onset of a chronic disease like diabetes can be complex to face, especially at a tender age. Patients will have to develop control strategies for T1DM and their general psycho-physical well-being. The main obstacle to overcome is the newly acquired independence that occurs during the passage from childhood to adult life: the family and the clinicians’ team pass their role onto the subject, who will now have a dramatic influence in the future development of the pathology. That potentially leads to conflicts and refusal of self-management. Thus, innovative devices and algorithms that help keep glycaemia under control are needed. At the date of writing of this work, there are no published papers or articles focused on the optimisation of the choice of the parameters required by Control-IQ™ technology. The work was done on data coming from 20 patients (13 females and 7 males). The data was organised in an Excel file reporting the insulin, glycaemic and carbohydrate info and in a PDF file reporting the Personal Profile settings. The Excel data was preprocessed as, in order to be analysed, glycaemia and insulin pieces of information had to be comparable. In fact, the frequency at which blood glucose levels and insulin velocities and injections are recorded are different. Thus, the Excel files for glucose had several rows more. Moreover, insulin and carbohydrate data were not only scarcer, but they were also often recorded at completely different times compared to the glycaemic one. For this reason, a few steps were followed to homogenise data. Later, at first, machine learning models were applied on the properly preprocessed data to understand when the algorithm decides to change the BR or not (logistic regression and random forest), then the output of the latter model was applied to preprocess the data used as input for a multivariate linear regression focused on understanding how BR was changed. Then, the metrics coming from this machine learning modelling approaches were evaluated. Finally, further analysis was done to understand how the different pump settings influence glycaemia (through a parameter called "tir" inspired by guidelines' TIR). In this case, two different types of tirs were analysed: one linked to the meals (lunch and dinner) and another one linked to the interesting time intervals in which the Personal Profile settings were changed for each subject. The results were quite satisfactory for an initial overview of the data and to gather insight on how it can be used to get to the ultimate goal of a fully automated choice of parameters. Now, there is knowledge on how the data must be collected, preprocessed and studied. Despite the limitations of this work, this thesis represents a proper starting point towards deep and complete analysis that, along with further work, could make an efficient and completely automated algorithm for the setting of the Personal Profiles a reality.