Improvements of LACE+ 30-day Readmission Risk Score
Published:
Improvements of LACE+ Readmission Risk Score
work by Eran Simhon, and Luoluo Liu
Background
Although the literature on 30-days readmission risk scores is rich and several scores have been developed over the years, predicting readmission during a hospital stay for general population remains a key challenge. Most models either perform poorly or requires data elements that are not easily accessible in real-time.
Main improvements of the well-known Canadian LACE+ model
- train a XGBoost model on US data of about half a million inpatients of a large multi-states healthcare network
- use a composite two-step prediction model, where in the first step we map ICD codes to clinical categories and predict risk of readmission solely based on clinical categories. The prediction is added as a feature in the next step, replacing Case-mix score suggested in the original LACE+ model
- the ability to deal with missing input data elements
the improved LACE+ algorithm:
Here is an illuatration of the improved LACE+ algorithm: 
Main results:
the weighted erformance across different hospitals with 600K inpatient encounters, LACE+ AUC=0.66, Improved LACE+ AUC=0.772, with about 17% improvement on AUC.
Using feature importance analysis, the added Clinical cateogries is the top feature.