In 2010, the Affordable Care Act (ACA) began financially penalizing hospitals whose 30-day readmission rates for acute myocardial infarction, heart failure, and pneumonia are higher than expected under the Hospital Readmissions Reduction Program. Since then, according to the figures released by the Centers for Medicare & Medicaid Services, nearly $2.5 billion of penalties were collected.
An efficient discharge decision-making system with machine learning
Unplanned readmissions not only incur avoidable waste of medical resources, Intensive Care Unit (ICU) readmission in particular, is exposing patients to additional morbidity and mortality risks. The mortality rate of readmitted ICU patients is believed to be around 26 to 58%. Despite so, hospitals in developed countries are still suffering from high ICU readmission rates. A study published in 2013 found that approximately 10% of the patients will be readmitted back to ICU after a hospital stay. In the US, there has also been an increase in ICU readmission rates, from 4.6% in 1989 to 6.4% in 2003.
As such, identifying patients who are likely to be readmitted render significant benefits to both healthcare providers and patients. Early on, a group of researchers at the University of Illinois at Urbana-Champaign relied on machine learning to uncover patterns in complex datasets and generate an efficient discharge decision-making system for ICU specialists and medical staff. The research team constructed a readmission dataset based on more than 40,000 ICU patients’ electronic health records (EHRs) extracted from MIMIC-III. These patients received care at the Beth Israel Deaconess Medical Center between years 2001 and 2012.
Three categories of features: patients’ clinical records and medical history, demographic details, and presence of chronic diseases were selected to design the readmission prediction model. The researchers used Recurrent Neural Networks (RNN) with Long Short-Term Memory (LSTM); so that the machine learning model will capture all the multivariate features embedded within EHRs and subtle changes (i.e., fluctuating blood pressure or glucose levels) in patients’ clinical records.
Recurrent Neural Networks with Long Short-Term Memory to enhance sensitivity
Indeed, researchers found that the LSTM-based solution can better capture the extremely volatile and unstable ICU patients, an important indicator for ICU readmission. Besides, the machine learning model also found to be more sensitive and accurate as compared to traditional prediction models. The researchers thought the model also illustrated the importance of each input feature, especially chronic diseases, which is often neglected in EHRs dataset but is highly relevant to later readmissions.
They hope that this fast and interpretable model will augment ICU specialists’ efficiency in deciding whether patients should be discharged based on their risks of readmission. Nevertheless, the researchers also noted that the model requires further training and validation, preferably with more complex real-world data obtained from other sources before it’s ready to be used clinically. All findings can be found on PLoS One.
If you are interested in learning more about the use of machine learning and deep learning, particularly, convolutional neural networks, recurrent neural networks and deep reinforcement learning in ICU for decision-support, do not miss the upcoming AIMed virtual event taking place on 22 September.
Dr. Matthieu Komorowski, Consultant in Intensive Care, Charing Cross Hospital and Clinical Senior Lecturer at the Imperial College London and Dr. Kevin Maher, Professor of Pediatrics at Emory University School of Medicine and Director of Cardiac Intensive Care at Children’s Healthcare of Atlanta will be leading a session on the topic. You may register for the virtual event and obtain a copy of its agenda here.