Analysis and prediction of unplanned intensive care unit readmission using recurrent neural networks with long short-term memory.

摘要: Background Unplanned readmission of a hospitalized patient is an indicator patients' exposure to risk and avoidable waste medical resources. In addition hospital readmission, intensive care unit (ICU) brings further financial risk, along with morbidity mortality risks. Identification high-risk patients who are likely be readmitted can provide significant benefits for both providers. The emergence machine learning solutions detect hidden patterns in complex, multi-dimensional datasets provides unparalleled opportunities developing efficient discharge decision-making support system physicians ICU specialists. Methods findings We used supervised approaches prediction. methods on comprehensive, longitudinal clinical data from the MIMIC-III predict within 30 days their discharge. incorporate multiple types features including chart events, demographic, ICD-9 embeddings. have utilized recent techniques such as Recurrent Neural Networks (RNN) Long Short-Term Memory (LSTM), by this we been able multivariate EHRs capture sudden fluctuations event (e.g. glucose heart rate). show that our LSTM-based solution better high volatility unstable status patients, important factor readmission. Our models identify readmissions at higher sensitivity rate 0.742 (95% CI, 0.718-0.766) improved Area Under Curve 0.791 0.782-0.800) compared traditional methods. perform in-depth deep performance analysis, well analysis each feature contribution predictive model. Conclusion manuscript highlights ability improve accuracy real-world example precision medicine hospitals. These data-driven hold potential substantial impact augmenting anticipate will counseling, administration, allocation healthcare resources ultimately individualized care.