LONGITUDINAL DATA ANALYTICS FOR CLINICAL DECISION SUPPORT IN TYPE 2 DIABETES

Type 2 Diabetes Mellitus (T2DM) is assuming epidemic proportions, which will progressively worsen as the population ages. Managing T2DM is a complex task, such complexity being embodied in long clinical histories, lasting longer than 10 years and characterized by substantial variability in the type and frequency of clinical events that are manifested across the population and within a single patient history. In addition, the pathology itself entails a number of complications and comorbidities. These issues suggest the difficulty in managing T2DM chronic patients (World Health Organization 2016). A major source of complexity in the management of T2DM patients arises from events such as hospital admissions, follow-up clinic visits, laboratory tests, and therapy changes. During these events, patients are treated by many different health professionals. These events are stored in different data repositories using different formats and occurring in temporal sequences that represent the patient careflow. Although these data are distributed in sources such as the Electronic Health Record (EHRs) and, Administrative Data Warehouses, new data management technologies are able to gather and merge them, and consequently enable researchers and other to access a huge quantity of complex data for the interpretation and exploitation of these data for a management of chronic diseases. The application of longitudinal analysis and careflow discovery to these data enable the recognition of hidden temporal patterns, population stratification and cohorts’ identification, and phenotypes definition. Temporal data analysis and careflow mining techniques can automatically detect the most frequent patterns and careflows from routinely collected data. Once identified, the enacted careflows might be used for comparison with clinical protocols to check their adherence to best practices, but they can be also exploited to identify different sub-groups of individuals in large cohorts of patients. These temporal data mining techniques can be used as a type of electronic phenotyping., which has been defined as the detection of computable phenotypes through query to EHRs and clinical data repository using specific data elements and logical expressions. Clinical guidelines and health care protocols are well-established tools used to improve and standardize health care services. Nevertheless, in the absence of effective technology-based solutions to automatically extract frequent patterns and careflows, it is often impossible to measure their implementation. Patients’ management processes can be improved through an overall system that integrates longitudinal heterogeneous data, and implements temporal data mining methods that illustrate the evolution of the disease and the individual and population variability. The detection of temporal patterns makes possible to reconstruct clinical pathways and forecast the complications that might arise during the process of care, to identify interesting clusters of patients with similar care histories and re-assess their risk profiles accordingly. The identification of healthcare pathways through methods derived from temporal and careflow mining research can be used for Decision Support. These facts suggest the need to investigate novel methods for improving the clinical decision support in T2DM and the utility of creating an analytics methodological framework. This is the overall goal of this dissertation, which was successfully retained completing these three specific aims: (i) To implement a system that integrates a large amount of unstructured and structured data from heterogeneous sources; (ii) To extend longitudinal analytic approaches to enable recognition of trending patterns and enhance temporal electronic phenotypes description; (iii) To create an expansion of existing methods for clinical decision support that is based on a more complete and easily understood description of patient health status.