Sepsis is a serious cause of morbidity and mortality and yet its pathophysiology remains elusive. Recently, medical and technological advances have helped redefine the criteria for sepsis incidence, which is otherwise poorly understood. With the recording of clinical parameters and outcomes of patients, enabling technologies, such as machine learning, open avenues for early prognostic systems for sepsis. In this work, we propose a two-phase approach towards prognostic scoring by predicting two outcomes in sepsis patients - Sepsis Severity and Comorbidity Severity. We train and evaluate multiple machine learning models on a dataset of 80 parameters collected from 800 patients at Amrita Institute of Medical Sciences, Kerala, India. We present an analysis of these results and harmonize consistencies and/or contradictions between elements of human knowledge and that of the model, using local interpretable model-agnostic explanations and other methods.