🏥 ICU Demand Forecasting & Patient Risk Stratification

A data analytics and machine learning project built using MIMIC-IV clinical data to predict ICU admissions and analyze patient risk patterns.

📌 Project Overview

This project focuses on:

Understanding hospital admission patterns Identifying patients at risk of ICU admission Building predictive models for ICU demand Creating interactive dashboards for business insights

🎯 Problem Statement

Hospitals often struggle with:

Unexpected ICU demand Resource allocation challenges Late identification of high-risk patients

👉 This project aims to predict ICU admission risk early using patient-level data.

📊 Data Source

Dataset: MIMIC-IV v3.1 Source: PhysioNet Data accessed via Google BigQuery

Includes: Patient demographics Hospital admissions ICU stays

🛠️ Tech Stack

SQL (BigQuery) → Data extraction & joins Python (Pandas, Scikit-learn) → Data processing & ML Power BI → Dashboard & visualization

⚙️ Project Workflow

1️⃣ Data Extraction (BigQuery)

Joined: patients admissions icustays Created target variable: ICU_FLAG (1 = ICU admission, 0 = No ICU)

2️⃣ Data Cleaning & Feature Engineering

Handled missing values Removed data leakage (ICU_STAY_DAYS) Created features: HOSPITAL_STAY_DAYS AGE_GROUP Encoded categorical variables

3️⃣ Handling Imbalanced Data

Used SMOTE to balance ICU vs non-ICU classes

4️⃣ Model Building

🔹 Logistic Regression Used for interpretability Applied feature scaling
🔹 Random Forest Used for performance comparison

5️⃣ Model Evaluation

Metric Logistic Regression Random Forest ROC-AUC ~0.70 ~0.75 Recall (ICU) High Moderate Accuracy Moderate Higher

👉 Logistic Regression chosen for: Better interpretability Higher recall (important in healthcare)

6️⃣ Key Features Driving ICU Risk

Hospital Stay Duration Age / Age Group Gender

📈 Dashboard (Power BI)

The dashboard provides:
🔹 KPIs Total Patients ICU Patients ICU Rate High-Risk Patients Avg Hospital Stay

🔹 Visual Insights ICU Rate by Age Group ICU Distribution by Gender ICU Risk vs Hospital Stay Duration

📊 Key Insights

ICU admission likelihood increases with longer hospital stays Certain age groups show higher ICU risk Gender differences observed in ICU distribution Dataset is highly imbalanced, requiring careful modeling

🚀 How to Run

Clone the repository Install dependencies: pip install pandas scikit-learn imbalanced-learn Run the Python script: python main.py Open Power BI dashboard file (.pbix)

📂 Project Structure

📁 Project
│
├── 📄 main.py # ML pipeline
├── 📄 icu_prediction_data.csv
├── 📊 dashboard.pbix # Power BI dashboard
├── 📄 README.md

💡 Future Improvements

Add more clinical features (labs, vitals)

⚠️ Disclaimer

Data is de-identified (MIMIC-IV) For educational and research purposes only

⭐ Author

Pavan R V