Using deep learning with registry linked claims data to predict hospitalization during chemotherapy: Feasibility study.

Abstract

270 Background: Approximately half of cancer patients undergoing outpatient chemotherapy experience unplanned Emergency Department (ED) visits and Inpatient (IP) stays. Current machine learning algorithms that identify high-risk patients are based on pre-treatment variables which can not detect changes in risk over time. Deep learning recurrent neural networks can model complex longitudinal patient histories. This study tests the feasibility of using an interpretable recurrent neural network to predict a patient’s daily likelihood of ED and unplanned IP stays in the six months following chemotherapy initiation. Methods: Medicare and commercial claims data were linked with cancer registry records for patients in Washington State from 2011 to 2017. The study included patients diagnosed with any primary tumor site, excluding leukemia, and treated with chemotherapy. We used the Reverse Time Attention model (RETAIN) with a 1:10 case-control match and included registry elements; diagnoses, procedures, medication, and utilization pre-and post-chemotherapy initiation. Patients were randomly divided into internal training, validation, and test sets (75%, 10%, 15%). Model accuracy was measured by the areas under the receiver operating curve (ROC) and precision-recall curve (PRC), and the Youden sensitivity and specificity. Results: Of the 15,400 eligible patients; 4,037 (26.2%) visited the ED a median of 1 time (6,080 total visits); 5,116 (33.2%) had a median of 1 IP stay (7,839 total stays). Both models had good predictive accuracy: The top 20 predictors for ED visits included 5 chemotherapy regimes, 12 procedures, and 2 tumor characteristics; IP stays included all chemotherapy regimes. Conclusions: The promising performance of RETAIN supports the possibility of building a tool capable of estimating daily hospitalization risk. However, future research, particularly with alternative data sources, may be required to predict hospitalization in a real time clinical setting. [Table: see text]