Time-intervals-related pattern selection for continuous event prediction

Symbolic time intervals (STIs) are used in different domains to represent real-life events with varying durations, like traffic light timing or medical treatments. Multivariate temporal data may include STIs and event-driven or manual measurements, such as traffic accidents or blood tests. This study proposes temporal abstraction to uniformly represent such heterogeneous multivariate temporal data (time point values, instantaneous events, or time intervals) using STIs and to develop a model to continuously predict events of interest. We introduce the use of multiple TIRPs that end with an event of interest for continuous prediction while using multiple TIRP instance completion predictors simultaneously. Since often there are dozens of discovered patterns, in this paper, we introduce novel discriminative pattern-selection metrics, such as the differences in the frequencies or duration between the entities (e.g., patients) having or not having the event of interest. The proposed methods achieved an average improvement of 5% AUROC over LSTM-FCN, the best-performing baseline, out of the evaluated baseline models (RawXGB, Resnet, LSTM-FCN, and ROCKET) that were applied to the raw medical and non-medical datasets. The proposed continuous event prediction method has the potential for broad real-world and real-time applicability in diverse domains with heterogeneous multivariate temporal data, such as early panic attack prediction using wearable devices or early complication prediction in intensive care unit patients.