Patient Survival Prediction | ICU Mortality Modeling

Predicting Patient Mortality Using Clinical and Demographic Data with R

Project Overview

This healthcare-focused ML project aims to predict the survival outcome of patients admitted to the Intensive Care Unit (ICU), using a rich dataset with over 190 clinical features from ~92,000 patients. The goal is to support critical decision-making in hospitals by modeling mortality risk using statistical inference and machine learning techniques.

Objective

To classify whether a patient will survive (hospital_death = 0) or not (hospital_death = 1) based on variables like:

• Vitals (heart rate, glucose, BMI, creatinine)
• Diagnoses (hepatic failure, diabetes)
• ICU types and treatment metadata
• Demographics (age, gender, ethnicity)

Key Techniques & Methodology

Data Preprocessing

• Handled missing values and extreme outliers using EDA
• Analyzed skewed variables using median/mean comparisons
• Removed high multicollinearity using VIF and correlation matrices

Feature Selection

• Chi-Squared Test for categorical feature impact
• LASSO Regression for variable selection
• Stepwise Regression (AIC) for optimal model simplification

Model Building

• Logistic Regression (with L1/L2 regularization)
• Decision Tree Classifier (for interpretability in clinical settings)
• Evaluation via ROC curve, accuracy, and AUC score

Key Hypothesis Testing

• Impact of hepatic failure, diabetes, and ICU types on mortality
• Age group and gender-based survival differences
• Ethnicity-specific outcome disparities

Outcomes & Insights

• Senior citizens showed the highest mortality rate across the dataset.
• Hepatic failure and diabetes mellitus showed statistically significant association with hospital deaths.
• Some ICU types had higher death probabilities, influencing decisions about triage and treatment prioritization.
• LASSO and stepwise regression were effective in highlighting key predictors like age, BMI, glucose levels, and ICU type.

Tech Stack

• Language: R
• Tools: RStudio, Tidyverse, MASS, glmnet
• Techniques: Logistic Regression, Stepwise AIC, LASSO, Chi-Square Test
• Domain: Predictive modeling in healthcare

Repository Structure

📦 Patient-Survival-Prediction/
├── PART - 1.Rmd                   # Data Cleaning, EDA, Feature Analysis
├── PART - 2.Rmd                   # Modeling, Evaluation, Hypothesis Testing
├── dataset.csv                    # ICU patient data (~92,000 records)
├── Part - 1.pptx                  # Visual Storytelling (EDA & Insights)
├── Part - 2.pptx                  # Modeling Results & Medical Implications
└── README.md                      # This file

Use Case

This model can help:

• Triage ICU patients based on real-time survival risk
• Optimize ICU staffing and treatment priorities
• Generate medical research insights using data-backed survival patterns

Author

Surya Vamsi Patiballa
M.S in Data Science – George Washington University

• Email :- svamsi2002@gmail.com
• LinkedIn :- https://www.linkedin.com/in/surya-patiballa-b724851aa/