AAI_2025_Capstone_Chronicles_Combined

13

weighted sampling, and training metrics such as loss and AUROC were monitored per epoch. Optuna ran

25 trials to identify the best hyperparameter combination. The final model was evaluated on the test set

using AUROC, accuracy, precision, recall, and F1-score (see Figure 3). SHAP analysis (NHANES I

Survival Model — SHAP Latest Documentation, 2018) was applied to interpret feature importance in the

best-performing model (see Figure 5).

TimesFM with Static Feature Fusion

To complement PatchTST, we developed a fusion model named ICUStaticFusionTimesFM,

which integrates dynamic time-series data with static patient features. This architecture leverages the

pretrained TimesFM backbone ("google/timesfm-2.0-500m-pytorch") (TimesFM, 2025) for temporal

modeling, combined with a custom pipeline for static feature integration. Static features were projected

into a lower- dimensional space using a linear transformation, then concatenated with TimesFM’s

dynamic embeddings to form a unified representation. A multi-layer perceptron (MLP) classifier

processed the fused features, incorporating ReLU activation, dropout regularization, and a final linear

layer for binary classification.

Training followed the same stratified split strategy and used PyTorch DataLoaders. Class

imbalance was addressed using weighted sampling, and CrossEntropyLoss was used as the training

criterion. The AdamW optimizer was employed with tuned learning rate and weight decay. Training ran

for up to 10 epochs, with early stopping based on validation AUROC (patience = 5). Training loss and

validation AUROC were tracked throughout. Hyperparameter tuning was conducted using Optuna,

targeting static projection size, classifier hidden size, dropout rate, learning rate, weight decay, and batch

size. The objective function maximized validation AUROC across 10 trials, and the best-performing

model was saved and evaluated on the test set. Notably, the optimized model showed improved recall for

the minority class ("Died"), a critical metric in clinical risk prediction (see Figure 4). SHAP analysis

159