Event prediction by estimating continuously the completion of a single temporal pattern's instances.

Area of Science:

• Machine Learning
• Time Series Analysis
• Predictive Modeling

Background:

• Continuous prediction in temporal data is challenging.
• Existing methods struggle with heterogeneous multivariate time series.

Purpose of the Study:

• Develop a new continuous prediction method.
• Utilize single temporal patterns ending with an event of interest.
• Improve prediction accuracy and timeliness.

Main Methods:

• Employed temporal abstraction to create Symbolic Time Intervals (STIs).
• Introduced Time Intervals-Related Patterns (TIRPs) for event prediction.
• Trained models on patterns preceding events to predict occurrence and timing.

Main Results:

• Achieved an average 5% improvement in Area Under the Receiver Operating Characteristic Curve (AUROC).
• Outperformed baseline models including LSTM-FCN, RawXGB, Resnet, and ROCKET.
• Validated on real-life heterogeneous multivariate temporal datasets.

Conclusions:

• The proposed method offers a powerful tool for continuous, real-time event prediction.
• Applicable across diverse domains with complex temporal data.
• Potential applications include early prediction of panic attacks and ICU patient complications.