Intelligent Healthcare Process Discovery And Operational Coordination Using Discrete Event Simulation And Machine Learning

The healthcare system in the US is rapidly changing and reshaping to adopt continuously evolving demand for improved operational efficiency and treatment effectiveness from patients and providers in critical health services. Healthcare service systems and clinical treatment operations need to be more predictable to increase operational efficiency through proactive operations management. This research contributes to the literature by discovering clinical processes and calibrating discrete-event simulation models in healthcare service systems using data-driven and process-driven predictive models. Unlike the data-driven predictive approaches such as machine learning and statistical methods, the proposed methodologies in this thesis leverages and focuses on process-based methods and analysis in healthcare service systems. Our first contribution is an integrated framework for process-driven multi-variate change point detection by coupling change point detection models with machine learning and process-driven simulation modeling in healthcare service systems. Initial development and succeeding calibration of discrete-event simulation models for complex healthcare systems require precise identification of dynamically changing process characteristics. Existing data-driven change point methods assume that changes are extrinsic to the system and cannot utilize available process knowledge. Our framework leverages simulation models to generate system-level outputs that are then used to predict system characteristics and change points using neural networks. The framework’s optimization layer iterates the change points by repeating simulation and predictive model building steps until the simulated system characteristics conforms to that of the actual process data. Using an emergency department case study, we demonstrate that the developed approach significantly improves change point detection accuracy over data-driven methods’ estimates and is able to detect actual change points. Our second contribution is a time-to-event prediction approach for clinical care operations in intensive care units. By focusing on the sepsis treatment in intensive care units, we predict time-to-event for antibiotic administration at critical vital states of the sepsis-risk patients. Our approach’s most salient aspects are the feature engineering specific to sepsis care and timing and labeling of the predictions. Using a real dataset, MIMIC-III (Medical Information Mart for Intensive Care-III) dataset, we demonstrate that the approach is able to accurately predict a practical time-window for antibiotic administration. Through predicted antibiotics administration time interval, the providers can make informed decisions and the operations staff can proactively coordinate activities to ensure meeting service standards for quality of care