M.S. AAI Capstone Chronicles 2024

Evaluation of sample predictions of the model also revealed interesting insights as to how the DI may support the model in building short and long term context of overall patient condition. Examples are illustrated in Figure 4 below. On the left, the DI of two sample patients are plotted together with the SCT’s predicted sepsis outcome. The SCT predicted a positive sepsis outcome for the patient with the red line in this graph and a negative outcome for the patient with the green line. Both predictions were correct. This visualization illustrates how the DI may help the SCT build long term context of the patient’s condition over time. The red line shows a larger index value trending upwards, while the green line shows a lower overall index value trending downwards. On the right is an example patient where the SCT correctly predicted a negative sepsis outcome, even though the overall trend seemed to indicate deterioration. The graph indicates that the patient seems to make a drastic recovery during the last hour of the series. This may provide some evidence to the DI capturing rapid changes and helping the model weigh both short and long term context. In this case, it helped the model potentially avoid making a false positive prediction. This supports the overall improvement in false positive performance observed during baseline testing.

279