M.S. AAI Capstone Chronicles 2024

When combined with the novel SCT architecture, the system was able to strike a good balance between optimizing for a low rate of false negatives and maintaining a relatively low occurrence of false positives. The system also generalized well to unseen data. By managing the overfitting problem and minimizing the rate of false positives, the system was able to overcome two of the most common challenges faced by previous works in this area (Wong et al., 2021). Overall, these results were encouraging and set a solid foundation for additional work to further improve sepsis predictability in clinical settings. One particular area that warrants deeper exploration is to shorten the context window required by the model to make reliable predictions. This project experimented with both 24 and 48 hour windows, eventually landing on 48 hours due to the difficulty in obtaining reasonable performance with only 24 hours of context. This could likely be significantly improved through optimizations to the data preparation pipeline. Specifically, applying additional filtering of patients with 24 hours (or less) of context based on data quality and completeness would be a recommended area of focus, as this cleaner, more consistent data may enable the model to perform well even with the shorter context. Additionally, the team conducted some preliminary investigation into a second novel feature that would capture cross-correlation between vital signs and lab readings. This work was not completed and would be another area for further research.

281