M.S. AAI Capstone Chronicles 2024

The final issue observed in the dataset was a significant class imbalance in the target variable (sepsis outcome). This is to be expected, as most patients in the ICU do not actually develop sepsis. However, this imbalance would need to be handled to ensure that our model does not become biased towards predicting negative sepsis outcomes. The team decided that rather than perform this balancing on the dataset itself (ie: via over- or under-sampling), the model would be weighted to handle the imbalance instead. With the dataset cleaned, feature-selected and filtered to patients with suitable data density and time series lengths, two additional preparation tasks were performed. First, the continuous features were scaled using a standard scaler and our single categorical feature (sex) was one-hot encoded. Second, derived features were created for each of the continuous lab variables to capture the “lag” in the last time step where the value was updated. The goal of these features is to help the model place the appropriate weight on the value of these readings as they progress through time without regular updates. Finally, a novel feature, the Deterioration Index (DI), was engineered to encode changes in overall patient health over time. The DI was a calculated score designed to monitor changes in a patient’s vital signs and lab values, enabling the detection of rapid trends or deviations that may signal clinical deterioration, such as the onset of sepsis. Effectively, the DI encodes medical knowledge that a human doctor or nurse may apply in a clinical setting in order to detect degradation in a patient’s overall condition, potentially leading to sepsis. The feature also incorporates a rate of change from the previous time step, providing time-based trend information as to whether a patient is deteriorating or improving. In the case of the improving condition, where a patient moves from a moderate or severe condition back to normal, the score for the feature at that time step is -1, which reduces the overall index. A higher DI value

269