Democratizing EHR analyses with FIDDLE: a flexible data-driven preprocessing pipeline for structured clinical data.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Data Preprocessing

Background:

• Machine learning (ML) applied to electronic health record (EHR) data requires extensive, labor-intensive preprocessing.
• There is a growing need for systematic and reproducible EHR data preprocessing techniques.
• The development of effective ML tools for healthcare is hindered by preprocessing challenges.

Purpose of the Study:

• To introduce FIDDLE (Flexible Data-Driven Pipeline), an open-source framework designed to streamline EHR data preprocessing.
• To demonstrate the utility and flexibility of FIDDLE in preparing EHR data for ML applications.
• To reduce the manual effort and decision-making involved in EHR data preprocessing.

Main Methods:

• FIDDLE systematically transforms structured EHR data into feature vectors in a largely data-driven manner.
• The framework was applied to two public EHR datasets (MIMIC-III and eICU) from intensive care units.
• ML models were trained to predict in-hospital mortality, acute respiratory failure, and shock, with performance evaluated using AUROC.

Main Results:

• FIDDLE extracted a significant number of features (2,528–7,403) from the EHR datasets.
• Models utilizing FIDDLE achieved strong discriminative performance (AUROCs of 0.757–0.886), comparable to existing methods.
• FIDDLE demonstrated generalizability across different prediction times, ML algorithms, datasets, and robustness to user-defined arguments.

Conclusions:

• FIDDLE is an open-source pipeline that simplifies the application of ML to structured EHR data.
• By standardizing and accelerating preprocessing, FIDDLE can advance the creation of clinically relevant ML tools.
• The framework facilitates progress in leveraging EHR data for improved healthcare outcomes.