Predicting 30-Day Hospital Readmissions Using Deep Learning

• Activity Number: 174 - Statistical Methods to Assess the Performance of Health Providers
• Type: Contributed
• Date/Time: Monday, July 29, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Wenshuo Liu* and Ji Zhu and Brahmajee Nallamothu and Akbar Waljee and Karandeep Singh and Andrew Ryan and Devraj Sukul and Elham Mahmoudi
• Companies: University of Michigan-Ann Arbor and University of Michigan and University of Michigan-Ann Arbor and University of Michigan-Ann Arbor and University of Michigan-Ann Arbor and University of Michigan-Ann Arbor and University of Michigan-Ann Arbor and University of Michigan-Ann Arbor
• Keywords: deep learning; word embedding; medical codes; readmission; risk-adjustment

Abstract

Risk-standardarized hospital readmission rates are used to compare readmissions across hospitals caring for patients with differing levels of acuity, but existing models based on logistic regression only account for a small number of factors contributing to risk. Through representation of the data in hidden layers, deep learning models may be able to capture a more nuanced picture of patient risk. In this work, we examined if models using deep learning improve prediction of 30-day readmission after acute myocardial infarction (AMI), congestive health failure (HF), and pneumonia (PNA) hospitalizations. We identified patients 18 years or older hospitalized with AMI, HF, and PNA in 2014 within the large Nationwide Readmissions Database (NRD). We constructed deep learning models that employ word embeddings for diagnoses and procedures from their co-occurrences within the NRD to predict 30-day readmissions and calculate risk-standardized hospital readmission rates. Different models were assessed using area under the curve (AUC) for the receiver operating characteristic and by comparing hospital rankings based on risk-standardized hospital readmission rates.