Extrema-weighted feature extraction for functional data.

Area of Science:

• Statistics
• Functional Data Analysis
• Biostatistics

Background:

• Existing methods for functional predictors often focus on dimension reduction, which may not capture dynamics at extreme values.
• Applications exist where the dynamics of a predictor during extreme values are highly informative for the response.
• Physicians need to understand how blood pressure dynamics during surgery, especially at extreme levels, relate to post-surgery outcomes.

Purpose of the Study:

• To propose a novel class of extrema-weighted feature (XWF) extraction models for functional predictors.
• To develop algorithms for fitting XWF-based regression and classification models.
• To compare XWF models with existing methods using simulations and a real-world application.

Main Methods:

• Extrema-weighted feature (XWF) extraction models are defined using the predictor's marginal density and a function up-weighting extreme quantiles.
• Functionals characterizing local dynamics are incorporated into the XWF definition.
• Algorithms for fitting XWF regression and classification models are developed and tested.

Main Results:

• XWF models successfully identify features in intraoperative blood pressure trajectories predictive of postoperative mortality.
• Simulations and a blood pressure application demonstrate the efficacy of XWF models.
• Many features identified by XWFs are novel and cannot be detected by previous functional data analysis methods.

Conclusions:

• Extrema-weighted feature (XWF) extraction models offer a powerful new approach for analyzing functional predictors, particularly when extreme values are informative.
• XWF models provide unique insights into the relationship between intraoperative blood pressure dynamics and patient outcomes.
• The proposed methods offer significant advantages over existing techniques for specific applications in functional data analysis.