Predicting Road Accident Severity Across the City of London

Coursework for LSBU's Data Mining & Big Data Analytics module, but with a real question behind it: can two years of collision data tell you where and when the fatal accidents happen, so resources can go there first?

I worked with 1,983 City of London road accidents from 2022–2023, aiming to predict severity and surface the conditions that drive the worst outcomes — the kind of analysis that feeds into emergency resource allocation and council safety planning.

The hard part: imbalance

The target was badly skewed — 85.9% slight, 13.7% serious, 0.5% fatal. A model that predicts "slight" every time scores 86% accuracy and is useless. So accuracy was never the metric that mattered; recall on the rare, serious classes was.

I cleaned the data with z-score outlier removal and mode imputation, then used SMOTE to rebalance the training set before fitting a Decision Tree classifier for severity prediction. It reached 75.2% accuracy and, more importantly, 92% recall on fatal accidents — it caught almost all of the cases that matter.

Finding the patterns

Prediction tells you what; I wanted why. Mining Apriori association rules over the accident conditions surfaced concrete risk patterns — the strongest being wet dual carriageways after dark, which reached a 73% fatal-confidence rule with 58× lift over the base rate. Across the board, road type, surface condition and lighting were the strongest predictors of a fatal outcome.

What I took from it

The metric is the model. Choosing recall-on-fatal over accuracy, and rebalancing the data to make that possible, changed what the project was actually optimising for — and turned it into something you could hand to someone making safety decisions.