Let's look at two recent studies that have applied machine learning to the readmission risk identification problem.

The first study comes by Duke University (Futoma et al., 2014). In this study, 3.3 million hospital admissions from New Zealand were studied. For each admission, demographic information, background information, diagnosis-related group (DRG) codes, and International Classification of Diseases (ICD) diagnosis and procedure codes were used. The matrix was then fed into six different machine learning algorithms:

• Logistic regression
• Logistic regression with the multistep variable selection
• Penalized logistic regression
• Random forest
• Support vector machine
• Deep learning

Of the first five methods, the random forest had the best performance, achieving an AUC of 0.684. The deep learning models were used to predict readmissions on five patient cohorts: pneumonia, chronic obstructive pulmonary disease (COPD), CHF, MI, and total hip/knee arthroplasty. The highest performance was achieved for pneumonia, with an AUC of 0.734. It should be noted that no direct comparison to LACE or HOSPITAL risk scores was performed.

The second study comes from the Advocate Health Care hospital system (Tong et al., 2016). The study consisted of 162,466 index (for example, initial) hospital admissions. For each index admission, 19 data elements were collected including the four LACE score components and 15 other EMR variables, such as past medical history, previous clinical encounters, employment status, and lab results. They trained three different machine learning models on this data:

• Logistic regression with the stepwise forward-backward variable selection
• Logistic regression with lasso regularization
• Boosting

They found that when they used 80,000 index admissions for the training/validation sets, the LACE model had an AUC of ~0.65, while the three machine learning models all had AUCs of ~0.73. The results suggest the significance of additional variables outside of the four variables used for the LACE score, and they further reinforce the use of machine learning for healthcare.