AAI_2025_Capstone_Chronicles_Combined

15

further revealed that each model learned distinct patterns, XGBoost relied on statistical summaries, while

CNN-LSTM and the transformer models captured temporal dynamics and feature interactions.

Looking ahead, future improvements should focus on threshold optimization, cost-sensitive

learning (Nabi, 2019), and attention-based fusion (Vaswani et al., 2017) to better handle class imbalance

and enhance model calibration. Real-time and longitudinal modeling, external validation (Harutyunyan et

al. 2019), and clinician-in-the-loop evaluation will be critical for translating these models into robust

clinical tools. By refining these architectures and expanding the dataset, we aim to build predictive

systems that are not only accurate but also clinically meaningful and trustworthy in high-stakes ICU

environments.

References

Alsentzer, E., Murphy, J. R., Boag, W., Weng, W.-H., Jin, D., Naumann, T., & McDermott, M. B. A.

(2019). Publicly available clinical BERT embeddings . https://arxiv.org/abs/1904.03323

Alzubaidi, A., Al-Shamma, O., Fadhel, M. A., Zhang, J., Duan, Y., Al-Timemy, A. H., & Santamaría, J.

(2022). Predicting ICU mortality using machine learning algorithms: A comparative study. Diagnostics,

12 (5), 1068. https://doi.org/10.3390/diagnostics12051068

Chen, T., & Guestrin, C. (2016, August). XGBoost: A scalable tree boosting system . In Proceedings of the

22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785 – 794).

ACM. https://doi.org/10.1145/2939672.2939785

Chicco, D., & Jurman, G. (2020). The advantages of the Matthews correlation coefficient (MCC) over F1

score and accuracy in binary classification evaluation. BMC Genomics, 21 (1), 6.

https://doi.org/10.1186/s12864-019-6413-7

161