Data science

Medical data mining: abdominal aortic aneurysm records and length of stay at the hospital

Data mining on 6,000 abdominal aortic aneurysm patient records to find how medical records relate to length of stay at the hospital. After careful preprocessing, some models reached 70 to 84 percent accuracy.

Data mining
Machine learning
Healthcare

Data

The Abdominal Aortic Aneurysm (AAA) patients' dataset has 6,000 records, produced by Vascular Governance North West of England and covering 1997 to 2012.

Goal

Use data mining to discover the relationships between patients' medical records and how long they stay in hospital (length of stay, or LOS).

Approach

Medical data is sensitive, and this dataset contained noise, outliers, and missing values, so the project started with a careful preprocessing stage to produce a clean, valid dataset.

Four model learning methods were then applied, and their accuracy and validity assessed:

Decision Tree (C4.5)
Rule Induction
k-nearest neighbours (KNN)
Naive Bayes

Results

Thanks to the preprocessing, some of the models reached high accuracy, between 70 and 84 percent. Merging the results of the decision tree and rule induction models gave a set of meaningful rules that could help medical managers manage their resources better.

This was my undergraduate final project, supervised by Professor Mo Saraee, with Dr. Abdolreza Mirzaei as referee. It was graded 20 out of 20.

More research and earlier projects

See all

Research paper2020

Cutting stock with rotation: packing square items into square bins

A research paper on square packing, a two-dimensional version of bin packing where square items go into as few square bins as possible. Unlike earlier models, items may be rotated, which matches how stock cutting works in practice.

Research paper2021

On the fault-tolerant online bin packing problem (server consolidation)

A research paper on placing replicated tenants onto as few cloud servers as possible while staying safe when a server fails. Tenants arrive one at a time and can't be moved later, so this is an online bin packing problem with fault tolerance built in.

Research paper

SCoR: an online Stack Overflow code recommender

A desktop application that takes a developer's question, collects Stack Overflow results in real time, then ranks and clusters the code snippets in the answers so the useful ones are quick to find.

Have a project in mind?

Send a short note on what you're building. I'll reply with questions and a rough plan.

Start a project Email contact@pooya.website