AAI_2025_Capstone_Chronicles_Combined

2

Introduction

Our capstone project aims to develop a machine learning model to predict in-hospital mortality

for ICU patients using the PhysioNet 2012 Challenge dataset (msafi04, 2021). This real-world dataset

contains multivariate time-series data from over 12,000 ICU stays, including vital signs, lab results, and

demographic information recorded during the first 48 hours of admission.

We seek to answer: Can temporal patterns in physiological data accurately predict ICU mortality?

This question is both clinically important and technically complex. Early mortality prediction can guide

timely interventions, improve patient outcomes, and optimize resource allocation. It also helps hospitals

manage ICU staffing, bed capacity, and prioritization during high-demand periods like public health

emergencies (Yeh et al., 2024).

The end users of our AI model include ICU physicians, hospital administrators, and healthcare

data scientists. These stakeholders rely on accurate, real-time information to make critical decisions in

fast-paced clinical environments.

The dataset supports two modeling strategies: aggregated features for simpler models (e.g.,

logistic regression, XGBoost), and full temporal dynamics to capture key patterns like declining MAP or

heart rate spikes. While aggregation improves interpretability via SHAP analysis (NHANES I Survival

Model — SHAP Latest Documentation, 2018), it may obscure essential prognostic signals. To evaluate

both approaches, we will implement models ranging from logistic regression to XGBoost, as well as deep

learning architectures including multilayer perceptrons (MLPs), long short-term memory networks

(LSTMs), one-dimensional convolutional neural networks (1D CNNs), and Transformer-based models —

well-suited for capturing temporal dependencies in clinical time series (Harutyunyan et al., 2019; Deng et

al., 2025).

148