Use of hemodynamics to predict events and identify latent structure in the post-op population

Abstract

Coronary Artery Bypass Grafting is the most common surgery for treating Coronary Artery Disease, and one of the most common surgeries worldwide. Despite having a low mortality rate, CABG has a relatively high complication rate. Medical protocols used by physicians to drive the recovery of these patients are highly standardized, yet the mechanisms that lead to different outcomes are still not well understood. By gaining deeperinsight into the physiologic relationships that drive a patient’s recovery from surgery,healthcare providers could better adjust their treatment, potentially improving recoveryand reducing complications. The main goal of this master’s thesis is to contribute to finding these relations by looking at trends in the evolution of vital signs and medication doses. This master thesis has been carried on at the Cardiovascular Research Center of the Massachusetts General Hospital and data used for the analysis come from this hospital. The methodology developed in the project consisted of three main steps. First, a tool was created to extract, curate and merge data from two databases of the hospital. The tool, named Tensorizer, can also be used to obtain data for other analyses. Next, a web visualization application was designed and implemented to allow visual examination of signals. This tool was also used to assist in the detection of corrupted data and its use is also extensible to other studies beyond this project. Finally, different clustering methods were applied on a set of 10 vital signs (heart rate; diastolic, systolic and mean pulmonary artery pressure; diastolic, systolic and mean arterial pressure; blood temperature, oxygen saturation and central venous pressure) and 5 medication dose trajectories (epinephrine, norepinephrine, vasopressins, milrinone and phenylephrine) recorded during the first after CABG. These methods include multivariate time-series clustering on the entire stream of vital signs and univariate time-series clustering on individual vital signs and medications.Results from these methods were combined into a table (Tabular Representation) and analyzed together.Clustering results suggest that vital sign and medication trajectories contain predictive power on their trajectories and reveal repeated patterns within and between signals. These patterns can be used to predict bad recovery from surgery, identify patients at risk or assist physicians to provide more personalized treatments. Results motivate and open the doors for future research in the field

Objectius de Desenvolupament Sostenible::3 - Salut i Benestar