Abstract 2089: Improving 30 day readmission prediction for hematological cancer patients via clinical embeddings

Introduction: Unplanned 30 day readmission prediction models exist for general cancer population1. However, the models may not perform well for hematology patients as their readmission risk may be associated with a variety of population specific factors including labs and diagnoses that are not typically used in models for general population2. Including all lab and diagnostic elements in predictive models may not be practical due to large number of variables relative to number of observations. Embeddings, originated in the field of natural language processing, have the potential to represent medical concepts in low dimensional spaces3. In this study, we developed a machine learning model utilizing clinical embedding representations of ICD and LOINC codes to predict readmission risk with prospective validation. Methods: This is a single institutional study examining inpatient 30 day unplanned readmissions between Jan 2013 and Dec 2016 (n = 16361 total, n = 5685 in hematology). The overall readmission rate was 18% (24% for hematology). We used gradient boosted trees models (lightGBM) including several baseline factors: gender, age, length of stay, service line, number of admissions within 6 months, primary insurance, discharge disposition, ICU admission, year of admission, and average pain score at rest. In addition, we utilized publicly available clinical embeddings3 to generate 300 dimensional representations for ICD9 and LOINC codes in the patients9 Electronic Medical Records. We considered diagnoses (ICD9s) within individual admission or within 6 months prior to admission, and lab tests (LOINCs) ordered during the admission. We used ten-fold cross-validation to identify optimal model hyper-parameters. We used records between Jan 2017 and Dec 2017 for prospective validation. To eliminate potential data leakage, we only considered new patients (n = 3785 total, n = 1424 in hematology), resulting in 17% overall readmission rate (22% for hematology). Results: In the prospective validation, the Area Under Receiver Operating Characteristic Curve (AUC) using baseline factors were 0.73 (average precision “AP” = 0.352) and 0.667 (AP = 0.347) for the overall and hematology populations, respectively. The AUC for hematology was notably lower. With the inclusion of ICD9 and LOINC embeddings, AUCs of 0.775 (AP = 0.44) and 0.722 (AP = 0.4) were obtained for overall and hematology populations, corresponding to 6% and 8% AUC improvements, respectively. Conclusions: This study found that readmission prediction performance for hematology patients is lower than the overall cancer population, and utilizing clinical embeddings improves the prediction performance. Given the higher readmission rates for hematology patients, this model shows potential to improve patient care and reduce associated costs, by predicting and preventing readmissions. 1 J Surg Oncol. 2018; 117:1113-1118. 2 JAMA Network Open. 2019;2(7):e196476. 3 AMIA Jt Summits Transl Sci Proc. 2016;41-50. Citation Format: Chi Wah Wong, Chen Chen, Lorenzo A. Rossi, Jerry Wang, Monga Abila, Janet Munu, Zahra Eftekhari. Improving 30 day readmission prediction for hematological cancer patients via clinical embeddings [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2089.