Exploring the impact of safety culture on incident reporting: Lessons learned from machine learning analysis of NHS England staff survey and incident data

Abstract

Safety culture is one of the key factors contributing to safety, even though limited evidence supports its impact on safety outcomes. This study uses supervised machine learning algorithms to explore the association between safety culture and incident reporting. The study used National Health Service (NHS) England annual staff survey data as a proxy of safety culture to predict eighteen incident reporting variables. The study did not achieve high accuracy rates in the prediction models. The highest association was found between safety culture and the number of incidents reported in class low, medium and high. LightGBM was the best-performed algorithm. SHAP plots were used to explain the model. Findings suggest that compassionate culture, violence and harassment and work pressure are critical in predicting the number of incidents reported. More specifically, the violence and harassment had a more significant impact on predicting the number of incidents reported in class high than in class medium and low. The involvement had more effect on predicting class low. The results demonstrated different behaviours in predicting different incident reporting classes. The findings facilitate lessons learned from staff surveys and incident reporting data in NHS England. Consequently, the findings can contribute to improving the safety culture in hospitals.