M.S. AAI Capstone Chronicles 2024

outcome and those that were not. This information was used to conduct feature selection, reducing the total variables from 40 to 21. Figure 2 Feature importance ranking used as input for feature selection

Another observation made during the interrogation of the data was that the length of stay (number of time steps) for each patient varied across a fairly large range of 8 hours minimum to 336 hours maximum with an average stay of approximately 38 hours. As mentioned previously, the ability to predict sepsis at least six hours in advance of diagnosis has been shown to improve patient outcomes (Fleischmann et al., 2016). In support of this objective, the patients would need to be filtered to those with a sufficient number of time steps to support a lookback window of at least 48 hours (a hyperparameter tuned through experimentation) and a prediction horizon of 6 hours, so 54 hours total. The dataset was filtered to remove patients that did not meet this criteria. This resulted in a smaller dataset, comprising a total of 3,882 patients.

268