From Messy EHRs to 30-Day Readmission Predictions: Benchmarking 4 ML Models

Author(s): Marie Humbert-Droz, PhD

Originally published on Towards AI.

Patient-level splits, imputation hacks, and interpretability tips for real-world healthcare AI.Image generated with ChatGPT

In Part 1, we explored why explainability matters in healthcare AI and introduced our 30-day readmission prediction model. We discussed the critical need for transparency when AI systems influence patient care decisions. Now, we’ll dive into the practical work of building multiple predictive models while addressing the unique challenges of healthcare data.

Today, we’ll cover the essential preprocessing steps and model selection decisions that form the foundation of any trustworthy healthcare AI system. By the end of this post, you’ll understand how to handle missing healthcare data appropriately, split datasets to avoid data leakage, and why we chose four specific model types for our comparison.

Healthcare datasets are notorious for missing values, and for good reason. Unlike controlled research environments, real clinical data reflects the messy reality of patient care:

Labs not ordered: A physician might skip certain tests if they’re not clinically indicatedPatient factors: Some patients can’t complete certain procedures or refuse specific testsWorkflow variations: Different hospitals or departments may have varying documentation practicesEmergency situations: Critical care scenarios often mean incomplete initial assessments

Traditional machine learning tutorials often treat missing data as a nuisance to be quickly handled with simple imputation. In healthcare, missing data can carry clinical meaning. The… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI