Predicting Hospital Readmissions Using XLNet-BiGRU-Attention on Patient Clinical Notes

Abstract

Hospital readmission is a key healthcare quality indicator. Predicting it from medical records enables providers to implement preventive interventions. We propose a predictive model utilizing a combination of eXtreme Language Understanding NETwork (XLNet), Bidirectional Gated Recurrent Unit (BiGRU), and Attention Mechanism, using the MIMIC-III dataset. The MIMIC-III dataset contains a rich collection of clinical data, including discharge summaries and various clinical notes, which provide a comprehensive view of patient histories and treatment outcomes. The architecture harnesses the contextual representations from XLNet, the capability of BiGRU to capture bidirectional sequential information, and the attention mechanism to focus on the most relevant parts of the medical records. The model demonstrated strong predictive performance, achieving a ROC-AUC score of 0.742, indicating its ability to distinguish between readmitted and non-readmitted patients. Additionally, the PR-AUC score of 0.723 highlights the model's effectiveness in handling imbalanced data, while the PR80 score of 0.237 reflects the model's capacity to maintain high precision at a recall rate of 80%. The proposed model predicts hospital readmissions from clinical text, helping providers identify high-risk patients and implement preventive measures.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint