ICU Resource Optimization

The Problem

Intensive care units are among the most resource-constrained environments in healthcare. Beds, ventilators, and specialist staff are finite — and when they're misallocated, patients wait longer for critical care. The question was whether machine learning could model ICU demand patterns and help optimize resource allocation in ways that traditional scheduling couldn't.

What I Did

This was research work at UFMG (Universidade Federal de Minas Gerais) as part of my MSc in Computer Science, focused on machine learning.

• Developed ML models to predict ICU occupancy and resource demand patterns
• Explored optimization approaches for bed and equipment allocation under real-world constraints
• Worked with healthcare data requiring careful handling of patient privacy and regulatory compliance
• Built experimentation pipelines for iterating on model architectures and feature engineering

The Hard Parts

Healthcare data is messy, sparse, and high-stakes. Overfitting to historical patterns is dangerous because the cost of a wrong prediction isn't a bad recommendation — it's a patient who doesn't get a bed. Balancing model accuracy with interpretability was critical: clinicians need to understand and trust the system, not just accept a black-box prediction.

Why It Matters

This project shaped how I think about building systems for real people. The technical problem was interesting, but the real challenge was always remembering that every optimization decision maps to a human outcome. That mindset carries directly into my current work: data quality in AI isn't an abstract engineering problem — it's about whether the person at the other end gets a good answer or a bad one.