FLIRT: A feature generation toolkit for wearable data.

Area of Science:

• Biomedical Engineering
• Data Science
• Wearable Technology

Background:

• Commercial wearable data often contains noise, missing values, and artifacts.
• Lack of standardized data processing hinders reproducibility in wearable health studies.
• Existing methods for wearable data preprocessing are often study-specific and lack uniformity.

Purpose of the Study:

• Introduce FLIRT (Feature Generation Toolkit for Wearable Data), an open-source Python package.
• Provide a standardized approach for processing physiological data from commercial wearables.
• Address challenges in data cleaning, normalization, and feature extraction for wearable data.

Main Methods:

• FLIRT employs advanced algorithms like particle filters and ML-based artifact detection for robust preprocessing.
• Utilizes a sliding-window approach to compute feature vectors with over 100 dimensions.
• Supports integrated handling of common wearable file formats, such as Empatica E4 archives.

Main Results:

• FLIRT was evaluated on the WESAD dataset for stress detection using an Empatica E4 wearable.
• The toolkit effectively filters unintended noise and artifacts from wearable sensor data.
• FLIRT-processed data achieved superior performance in classification tasks compared to previous methods.

Conclusions:

• FLIRT offers unique functionalities for physiological data processing and feature generation.
• The package ensures standardized feature generation, enhancing the reproducibility of research findings.
• FLIRT provides a highly configurable interface, accommodating diverse research needs beyond its default settings.