Explainable Prediction of Adverse Outcomes Using Clinical Notes

• DOI: 10.48550/arxiv.1910.14095
• Publication Date: 2019-01-01
• Authors: Justin Lovelace; Nathan C. Hurley; Adrian D. Haimovich; Bobak J. Mortazavi
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Statistics - Machine Learning; Machine Learning (stat.ML); 01 natural sciences; 0105 earth and related environmental sciences; 3. Good health; Machine Learning (cs.LG)

Clinical notes contain a large amount of clinically valuable information that is ignored in many clinical decision support systems due to the difficulty comes with mining information. Recent work has found success leveraging deep learning models for prediction outcomes using notes. However, these fail provide relevant and interpretable clinicians can utilize informed care. In this work, we augment popular convolutional model an attention mechanism apply it unstructured ICU readmission mortality. We find addition leads competitive performance while allowing straightforward interpretation predictions. develop clear visualizations present important spans text both individual predictions high-risk cohorts. then conduct qualitative analysis demonstrate our consistently attending meaningful portions narrative all explore.