AAI_2025_Capstone_Chronicles_Combined

10

With this enriched feature set, an exhaustive grid search with three-fold cross-validation

optimized parameters controlling complexity, regularization, and learning rate. The tuned model, trained

on the full resampled training set, achieved a ROC AUC of 0.8240 on the unseen test set (Figure 6). The

close match between cross-validation and test scores indicated no overfitting.

At the default 0.50 classification threshold, recall for the mortality class was 0.23, meaning most

high-risk patients would be missed. Lowering the threshold to 0.25 increased recall to 0.43, nearly

doubling correctly identified mortality cases while maintaining acceptable accuracy — an important

adjustment for clinical use.

SHAP analysis was used to interpret predictions (Figure 5). The most influential predictors

included neurological function, demographic factors, and proxies for monitoring intensity. These patterns

aligned with clinical understanding, such as elevated risk with reduced consciousness, advanced age, and

more intensive monitoring. This alignment, combined with strong predictive performance, supports the

model’s potential for trusted use in clinical decision support.

CNN – LSTM Model Design and Training

Along with the Xgboost model we also implemented CNN-LSTM using a hybrid architecture. The

CNN layers extract short-term temporal patterns from multivariate ICU time-series data, while the LSTM

layer captures long-term dependencies across the observation window (Ordóñez & Roggen, 2016;

Harutyunyan et al., 2019). This dual-path approach is well suited to ICU monitoring, where relevant

signals can appear as both transient fluctuations and sustained trends.

Architectural Design Choices

The temporal branch includes two one-dimensional convolutional layers followed by max

pooling and dropout regularization. The first convolutional layer uses 64 filters (kernel size = 3, ReLU

156