ClinicalICDBERT: Predicting Re-Admission Risk from Clinical Notes, Vital Signs and ICD Codes using BERT Models

Adamjee, Merzia Naeem

Etd

ClinicalICDBERT: Predicting Re-Admission Risk from Clinical Notes, Vital Signs and ICD Codes using BERT Models

Public

Readmissions are a financial burden and challenge for hospitals. Prior work has explored various structured predictors and machine learning algorithms to predict the risk of readmissions due to complications following colorectal, cardiac, and abdominal surgeries and heart failure. Models trained on clinical notes have generally resulted in a much better predictive performance for the hospital readmission task. The goal of this project is to explore and analyze the impact of ICD codes and vital signs with clinical notes for readmission risk prediction. This is achieved by concatenating clinical BERT embeddings created via pre-training on clinical notes, the vital signs data and the ICD codes embedding for each patient’s visit to predict readmission within the 30-day time period after discharge. In addition, we also compared our approach with the BERT model, ClinicalBERT, and a few other machine learning approaches that have predicted readmission within 30 days of discharge.

Creator