Multi-Output Learning for Rare Event Analysis

OBJECTIVES: Rare event analysis poses significant challenges due to the infrequency of occurrences, making it inherently difficult to gather sufficient data for robust analysis. Traditional statistical methods may struggle to detect patterns or establish meaningful relationships for rare events. However, in practice, we may always find correlated outcomes with large sample sizes. For instance, readmission or mortality may exhibit a high correlation with hospital admissions. In the domains of data mining and machine learning, combining multiple outcomes has the potential to alleviate the impact of suboptimal data quality and biases linked to rare events. This strategy seeks to improve the robustness and generalizability of results. This presentation delves into the viability of employing a multi-output learning approach to enhance model performance when addressing rare events in healthcare quality improvement areas.

METHODS: In our initial investigation, we employed 100 distinct sets of simulated data with different correlations to assess the potential of multi-output learning in mitigating model bias. Subsequently, we validated these findings by applying the approach to a real-world dataset comprising diabetic patient readmission data.

RESULTS: Our results showed the multi-output model always achieved better model results compared to the single output model.

CONCLUSIONS: By leveraging correlations between related events, a multi-output model can improve predictive performance for rare events. It enables the model to learn shared patterns and relationships among different outcomes, making it more robust and effective in capturing the complexity of rare events. Additionally, multi-output models can benefit from information shared across tasks, enhancing generalization capabilities.