Wavelet based time series forecast with application to acute hypotensive episodes prediction.

Area of Science:

• Signal Processing
• Computational Neuroscience
• Machine Learning

Background:

• Time series prediction is crucial in many fields.
• Existing methods may struggle with complex signal dynamics.
• Accurate forecasting of critical events like acute hypotensive episodes is vital.

Purpose of the Study:

• To present a generic methodology for time series prediction.
• To combine wavelet decomposition with feedforward neural networks.
• To validate the methodology on forecasting acute hypotensive episodes.

Main Methods:

• Signal decomposition using wavelet transform.
• Feature extraction via wavelet coefficients.
• Time series prediction using trained feedforward neural networks.
• Application to acute hypotensive episode forecasting using MIMIC-II data.

Main Results:

• Wavelet decomposition provides a compact signal representation.
• Neural networks effectively predict time series dynamics from coefficients.
• The method achieved performance comparable to leading approaches in the PhysioNet Challenge 2009.
• Successful forecasting of acute hypotensive episodes was demonstrated.

Conclusions:

• The proposed hybrid wavelet-neural network methodology is effective for time series prediction.
• This approach offers a robust way to forecast critical medical events.
• The technique shows promise for broader applications in complex signal analysis.