2022 journal article

Septic shock prediction and knowledge discovery through temporal pattern mining


author keywords: Temporal pattern mining; Sepsis; Electronic health records; Prediction; Pattern selection
MeSH headings : Critical Care; Electronic Health Records; Humans; Knowledge Discovery; Sepsis / diagnosis; Sepsis / therapy; Shock, Septic / diagnosis; Shock, Septic / therapy
Source: Web Of Science
Added: October 17, 2022

Sepsis is the body's adverse response to infection which can lead to septic shock and eventually death if not treated in a timely manner. Analyzing patterns in sepsis patients' health status over time can help predict septic shock before its onset allowing healthcare providers to be more proactive. Temporal pattern mining methods can be used to identify trends in a patient's health status over time. If these methods return too many patterns, however, this can hinder knowledge discovery and practical implementation at the bedside in acute care settings. We propose a framework to find a small number of relevant temporal patterns in electronic health records for the early prediction of septic shock. Our framework consists of a temporal pattern mining method and three pattern selection techniques based on non-contrasted group support (PST1), contrasted group support (PST2), and model predictive power (PST3, PST4). We find that model-based feature selection approaches PST3 and PST4 yield the best prediction performance among these techniques. However, PST2 identifies more multi-state patterns with abnormal health states, which can give healthcare providers indicators of patient deterioration towards septic shock. Hence, from a knowledge discovery perspective, it may be worthwhile to sacrifice a small amount of prediction power for actionable patient health information through the implementation of PST2. • A framework to identify relevant temporal patterns in data is proposed. • Recent temporal pattern mining is combined with three feature selection techniques. • Septic shock prediction models are trained using temporal patterns. • There is a tradeoff between prediction performance and knowledge discovery. • Contrasted grouping provides informative patterns of patient deterioration.