Accidents on the shop floor remain a significant concern in the manufacturing sector, affecting both productivity and workers’ safety. This study focuses on analysing real-world accident data to identify key causes, high-risk locations, and recurring patterns associated with such incidents. The data primarily consists of qualitative variables, which introduces complexity in applying machine learning models. To address this, appropriate data encoding techniques are employed to prepare the data for classification tasks. Additionally, the dataset exhibits class imbalance, which can lead to biased model outcomes. To mitigate this, various resampling and algorithm-level solutions are explored. The study develops a classification model to predict accident scenarios and proposes a framework that supports proactive decision-making for accident prevention. The results demonstrate that careful handling of qualitative and imbalanced data significantly improves model performance and the potential for real-time accident risk assessment.