Hospital Readmission Prediction Using Clinical Admission Notes

Clinical notes contain contextualised information beyond structured data relating to patients' past and current health conditions. Despite the richness, their unstructured, long, and high dimensional nature presents challenges to traditional text representation techniques. The advancement of deep contextual representation techniques in natural language processing (NLP) has shown remarkable performance in the biomedical and clinical domains for various information extraction and predictive tasks, including hospital readmission. However, most previous works have proposed discharge summary models where on-site medical intervention is impossible, and readmission could still occur. This paper utilises clinical notes recorded during admissions to study the risk of 30-day hospital readmissions. We employ clinical notes from MIMIC-III and consider competing baselines for clinical text representation, where a set of machine learning and deep learning algorithms are used to classify hospital readmission. The study demonstrates that notes captured during admissions play a crucial role to recognise potential readmission risk supporting healthcare practitioners for practical therapeutic intervention and discharge planning.