AAI_2025_Capstone_Chronicles_Combined

6

enabling both strong predictive performance and the identification of key clinical factors that indicate

mortality risk.

The use of XGBoost for clinical risk prediction on ICU data is well-established in academic

research. For example, Alzubaidi et al. (2022) developed a machine learning model using XGBoost to

predict ICU mortality based on the MIMIC-IV dataset, which contains over 70,000 patient records.

Among several models tested, XGBoost outperformed others with an AUC of 0.918, underscoring its

effectiveness for ICU risk stratification. This provides a strong precedent for its application in our project

and offers a valuable benchmark.

CNN-LSTM

The next model, a hybrid convolutional neural network (CNN) – long short-term memory

network (LSTM) architecture, is a robust method for modeling multivariate time-series data in clinical

applications, particularly for intensive care unit (ICU) mortality prediction. This approach combines the

strengths of CNNs in detecting short-term, localized patterns with the ability of LSTMs to capture long

term temporal dependencies, enabling effective processing of complex, irregularly sampled physiological

data (Yen et al., 2024).

In this architecture, CNN layers apply sliding filters to sequences of physiological

measurements such as heart rate (HR), systolic and diastolic arterial blood pressure (SysABP, DiasABP),

mean arterial pressure (MAP), respiratory rate (RespRate), and oxygen saturation (SaO2) to identify

abrupt changes, anomalies, and short-term trends that may indicate acute patient deterioration. The

extracted feature maps are then passed to the LSTM layers, which use gating mechanisms to selectively

retain, update, or discard information over time. This allows the model to preserve clinically relevant

temporal context even when sampling intervals are irregular (Hochreiter & Schmidhuber, 1997).

152