A1212
Title: Predicting ICU readmission with a hybrid BERTopic-LSTM approach on electronic health records
Authors: Chih-Chou Chiu - National Taipei University of Technology (Taiwan)
Chung-Min Wu - National Taipei University of Technology (Taiwan)
Te-Nien Chien - National Taipei University of Technology (Taiwan)
Ling-Jing Kao - National Taipei University of Technology (Taiwan) [presenting]
Chengcheng Li - National Taipei University of Technology (Taiwan)
Abstract: The high incidence of ICU readmissions poses a significant challenge in healthcare, resulting in increased expenses and suboptimal patient outcomes. A novel method is presented for predicting ICU readmission from electronic health records (EHRs) using a hybrid BERTopic and Long Short-Term Memory (LSTM) network approach. The model integrates the benefits of unsupervised topic modelling with supervised deep learning to effectively capture the intricate associations between patient attributes and readmission risk. A dataset of 5,000 ICU patient records was leveraged, where BERTopic was initially employed to cluster patients based on their EHRs. Then a supervised LSTM network was utilized for training on the clustered data, incorporating both the EHRs and patient demographics as inputs to forecast readmission risk. The method outperforms existing readmission prediction models based on traditional machine learning approaches, yielding an AUC-ROC of 0.80. Additionally, it is demonstrated that the proposed model can identify key risk factors for readmission, including comorbidities, length of stay, and ICU admission type. In summary, the aim is to highlight the efficacy of a hybrid BERTopic and LSTM network approach for predicting ICU readmission from EHRs. This approach holds promise in enhancing patient outcomes and reducing healthcare costs by enabling early intervention and targeted care management.
