Dynamic 30-day readmission prediction for cancer patients via clinical embeddings.

Abstract

2040 Background: Existing models typically predict unplanned 30-day readmission for cancer patients at discharge1. Performing prediction dynamically during hospital stay may allow earlier intervention for high risk patients. In addition, readmission risk may be associated with the outcome of a variety of labs and diagnoses. Models including all those elements may not be practical due to large number of variables relative to number of samples. Embeddings have the potential to represent medical concepts in low dimensional spaces2. In this study, we developed a machine learning model utilizing embedding representations of ICD and LOINC codes to dynamically predict readmission risk. Methods: This is a single institutional study examining inpatient 30-day unplanned readmissions from Jan 2013 to Dec 2016 (n = 16361 total, n = 5685 in hematology). The readmission rate was 18% (24% for hematology). We used gradient boosted trees models with 10-fold cross validation and included baseline factors that are typically available shortly after admission: gender, age, service, admission count within 6 months, insurance, emergency admission, admission year, allogeneic or autologous stem cell transplant (hematology only). For dynamic factors, we randomly selected a timepoint (TP, median = 2.4 days) during each visit. We utilized publicly available clinical embeddings2 to generate 300 dimensional representations for ICD9s and LOINCs in the patients’ Electronic Medical Records. We considered diagnoses (ICD9) between 6 months prior to admission and TP, and lab tests (LOINC) ordered between admission time and TP. We used records from Jan 2017 to Dec 2017 for prospective validation (n = 3785 total, n = 1424 in hematology), with 17% readmission rate (22% for hematology). Results: Prospective validation Area Under Receiver Operating Characteristic Curve (AUC) using baseline factors were 0.72 (average precision “AP” = 0.33) and 0.65 (AP = 0.32) for overall and hematology populations, respectively. By including dynamic factors, we obtained AUCs of 0.74 (AP = 0.4) and 0.7 (AP = 0.39) for overall and hematology populations, corresponding to 3% and 8% AUC (21% and 22% AP) improvements, respectively. Conclusions: We found that dynamic readmission prediction utilizing clinical embeddings improves the prediction performance comparing with using baseline factors only. The model shows potential to improve patient care and reduce costs by predicting and preventing readmissions when the patient is still in the hospital. 1 J Surg Oncol 2018; 117:1113-1118. 2 AMIA Jt Summits Transl Sci Proc. 2016;41-50.