Real-time predictive analytics for sepsis level and therapeutic plans in intensive care medicine

Abstract

This work aims to support doctor’s decision-making on predicting sepsis level and the best treatment for patients with microbiological problems. A set of Data Mining (DM) models was developed using forecasting techniques and classification models which will enable doctors’ decisions about the appropriate therapy to apply, as well as the most successful one. The data used in DM models were collected at the Intensive Care Unit (ICU) of the Centro Hospitalar do Porto, in Oporto, Portugal. Classification models where considered to predict sepsis level and therapeutic plan for patients with sepsis in a supervised learning approach. Models were induced making use of the following algorithms: Decision Trees, Support Vector Machines and Naïve Bayes classifier. Confusion Matrix, including associated metrics, and Cross-validation were used for the evaluation. Analysis of the total error rate, sensitivity, specificity and accuracy were the associated metrics used to identify the most relevant measures to predict sepsis level and treatment plan under study. In conclusion, it was possible to predict with great accuracy the sepsis level (2nd and 3rd), but not the therapeutic plan. Although the good results attained for sepsis (accuracy: 100%), therapeutic plan does not present the same level of accuracy (best: 62.8%).