Predicting hyperlactatemia in the MIMIC II database

Sepsis, which occurs when an infection leads to a systemic inflammatory response, is believed to contribute to one in two to three hospital deaths in the United States. Using the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC II) database of electronic medical records from Boston's Beth Israel Deaconess Medical Center (BIDMC), we worked to characterize sepsis at BIDMC's intensive care units. Additionally, we developed a real-time algorithm to stratify patients with infectious complaints into different risk categories for progressing to septic shock. From time series of heart rate and arterial blood pressure, as well as estimates of cardiac output and total peripheral resistance, we developed a variety of classifiers to predict high serum lactate levels, a proxy for hypoperfusion and imminent circulatory shock. The records from 146 patients met our selection criteria. In discriminating patients whose measured serum lactate stays below 2.5 mmol/L from those whose value drifts above, the best of our classifiers perform with area under the receiver operating characteristic exceeding 0.8 on test data.