| Description |
Medicine is the art and science of diagnosis and treatment of disease - maintenance of one's health. Temporal reasoning in medicine is the art and practice of modeling one's pathophysiological processes to aid in the diagnosis and treatment of disease to improve one's health status. Type 2 diabetes mellitus (T2DM) is fast becoming a top global concern, while specialists to treat this chronic, complex disease process are in short supply, hence the need for a clinical model or set of models that can aide clinicians in their efforts to diagnose and treat this disease. Dynamic Bayesian Networks (DBN) support this temporal modeling paradigm, as they are able to capture the changes in a patient's condition due to diseases and treatments, using probabilistic relationships regarding their past and present clinical condition, to predict future clinical events. To effectively model a chronic disease such as T2DM, or any clinical malady, one requires a data process that can take the gathered the clinical elements at their source and transform them into a format where the disease progression can be modeled. As a single patient may have accumulated thousands of clinical observations over a temporal period, further consideration is required to manage these data elements; therefore, a clinical data process was developed to cleanse the data, ascertain temporal granularity, align patient data to granular intervals, and clinically aggregate and impute missing data for modeling purposes. This clinical data transformation process was also used to model an emergent disease, sepsis, with data from a different clinical repository. Employing the data transformation process, several T2DM DBN models were iv created, evaluated, and validated using data from 54,738 patients who were already diagnosed with T2DM, to predict microvascular - neuropathy, retinopathy, and nephropathy, and the macrovascular - myocardial infarction, stroke, and mortality complications that are associated with T2DM over a 15-year period. These models were tested for granularity, data preparation methods, and optimal model selection; two experiments were developed to evaluate a model's predictive perimeters using censoring and evaluate these models to the status quo. Lastly, T2DM DBN models were used to statistically compare actual to predicted macrovascular treatment outcomes for two distinct treatment cohorts compared to a control group. This dissertation provides a general data processing application, several tested and validated T2DM DBN models, unique ways DBN models may be used to reduce bias, and that it is possible to temporally reason in medicine for T2DM outcomes and treatments using DBN. |
