Time-Wheel Based Multi-Series Sequence Partitioning for Emergency Triage Modeling

Electronic health records (EHRs) are comprehensive collections of patient data that play a critical role in emergency triage for better clinical decision-making. In this work, we focus on emergency triage--a key to improve the survival rate by grading the patients' condition for orderly treatment. Based on EHRs, artificial intelligence triage can be applied by studying the data similarity among patients. Thus, the emergency triage problem turns out to be a subsequence partitioning problem. Motivated by the subsequence clustering approach: Toeplitz Inverse Covariance-based Clustering (TICC), we modeled the initial emergency triage progression. However, the data collected from patients, such as heart rate, body temperature, pulse oximetry, etc., are asynchronous multi-series with irregular time intervals, ranging from seconds to days. In order to reduce the impact of irregular time stamps, a Time-Wheel based Multi-Series TICC (TWM-TICC) is proposed. The time can be “rolled up” into a time-wheel to eliminate the irregular time gap, and “unrolled” to original state. Its shift property and scaling property are demonstrated, as a result, TWM-TICC shows translation invariance and duality relationship between the ‘'rolled'’ and ‘'unrolled'’ time. The effectiveness of TWM-TICC is validated via a case study using the EHRs over the past three years within the top three hospitals in Guangzhou. The results suggest that TWM-TICC can significantly outperform competitive baseline models, including the TICC.