Data-driven modeling for uncertain biological systems

Abstract

Open-loop data was used for training and then the SF detection algorithm was applied in real-time during a closed-loop clinical trial, achieving 88.0% sensitivity and only issuing 0.22 false positives per day. This is the first report of real-time SF detection in an artificial pancreas and the retrospective results are superior to those reported in previous retrospective studies using clinical data. Project (2) analyzes potential urine- and blood-based biomarkers for autism spectrum disorder, reporting a biomarker based on folate-dependent one-carbon metabolism (FOCM) and transsulfuration (TS) that achieves the highest classification accuracy to date. Project (3) uses neural networks to determine relationships between host cytokines and bacterial metabolites on the differentiation of T cells prior to adoptive transfer, uncovering novel interactions and suggesting optimal differentiation protocols. It is emphasized that the models developed in all projects are evaluated through validation data sets or cross-validation to assess the predictability of these models, rather than just their description of specific data sets.

Data-driven models couple any systems, statistical, or optimization-based model to a particular application area and model objective. These models can operate in the midst of large amounts of model uncertainty since they do not require fundamental knowledge of the system of interest. This is especially advantageous when the system is poorly understood since data-driven models are not constrained by possibly false hypotheses and relationships, allowing the unveiling of new relationships that would otherwise go unnoticed. This thesis explores data-driven modeling in three areas of interest. Project (1) develops an insulin infusion set failure (SF) detection algorithm using glucose measurements and insulin dosing schedules of patients with type 1 diabetes.