M.S. AAI Capstone Chronicles 2024

layers, with the first stages performing convolution along the spatial axis and the latter stages performing convolution along the time axis. Since overfitting was an observed issue in research, several regularization techniques are employed, including batch normalization, dropout and skip-layer connections. The feature maps output from the Convolutional Module feed into the Transformer Module, where it is passed through multiple encoder blocks, each with a multi-head attention mechanism and feed-forward network (Vaswani et al., 2017). Additional regularization techniques are employed, including layer normalization and skip connections. Finally, the output from the transformer is fed to the Classifier Module, where it runs through multiple fully-connected layers before reaching the output layer, which is a single-neuron fully connected layer with a sigmoid activation function to perform binary classification. In order to prepare for experimentation, the team first sought to identify appropriate metrics that would provide an effective, comparable measurement of model performance. Recall and precision were identified as important metrics for the task of binary classification. While binary accuracy was also measured, it was not used as a primary performance metric due to it being an unreliable indicator of overall performance when dealing with a highly imbalanced dataset (Ghanem et al., 2023). The business understanding of this task also led the team to slightly optimize for recall over precision. In this case, strong recall performance translates into a lower rate of positive sepsis cases being missed by the model. This is critical in a clinical setting, as a missed positive sepsis case could lead to a negative outcome for a patient. On the other hand, strong precision performance indicates a low occurrence of false positives - that is, patients that do not develop sepsis being flagged as a sepsis risk. The team determined that some level of false positives were acceptable, though it was still important to minimize as much as possible. Overall, model performance was measured using both metrics with slightly more

274