[Paper Review] MIMIC-Extract: A Data Extraction, Preprocessing, and Representation Pipeline for MIMIC-III

Shirly Wang, Matthew B. A. McDermott+4 more|arXiv (Cornell University)|Jul 19, 2019

Machine Learning in Healthcare20 references42 citations

TL;DR

MIMIC-Extract provides an open-source pipeline to extract, preprocess, and represent MIMIC-III EHR data as ready-to-use time-series features, interventions, and outcomes for benchmarking ML models. It emphasizes robustness, reproducibility, and extensibility for ICU prediction tasks.

ABSTRACT

Robust machine learning relies on access to data that can be used with standardized frameworks in important tasks and the ability to develop models whose performance can be reasonably reproduced. In machine learning for healthcare, the community faces reproducibility challenges due to a lack of publicly accessible data and a lack of standardized data processing frameworks. We present MIMIC-Extract, an open-source pipeline for transforming raw electronic health record (EHR) data for critical care patients contained in the publicly-available MIMIC-III database into dataframes that are directly usable in common machine learning pipelines. MIMIC-Extract addresses three primary challenges in making complex health records data accessible to the broader machine learning community. First, it provides standardized data processing functions, including unit conversion, outlier detection, and aggregating semantically equivalent features, thus accounting for duplication and reducing missingness. Second, it preserves the time series nature of clinical data and can be easily integrated into clinically actionable prediction tasks in machine learning for health. Finally, it is highly extensible so that other researchers with related questions can easily use the same pipeline. We demonstrate the utility of this pipeline by showcasing several benchmark tasks and baseline results.

Motivation & Objective

• Provide a robust, reproducible pipeline to transform raw MIMIC-III EHR data into usable time-series formats for ML

Proposed method

• Cohort extraction focusing on first adult ICU stay (age ≥15, duration 12h–<10d)
• Unit standardization and outlier handling using clinically informed thresholds
• Hourly aggregation of vitals and labs into time-series features with clinical aggregation to reduce missingness
• Extraction of hourly intervention signals (ventilation, vasopressors, fluids) and static outcomes
• Two feature representations: raw ItemID-based features and clinically aggregated features
• Extensible design with keyword-driven configuration, resource files, and embedded SQL to support customization

Experimental results

Research questions

• RQ1 How can MIMIC-III data be transformed into standardized, robust hourly time-series suitable for prediction tasks?
• RQ2 Can a generic, reproducible data pipeline improve cross-study comparability and benchmarking for ICU ML models on MIMIC-III?
• RQ3 What are the effects of clinical aggregation, unit conversion, and outlier handling on model robustness to temporal drift?
• RQ4 What prediction tasks (mortality, LOS, and hourly interventions) are feasible with the extracted data and how do baseline models perform?

Key findings

• The pipeline outputs a 34,472-patient default cohort with both static and time-varying data suitable for multiple benchmarks
• Two output formats are provided: raw item-level features and clinically aggregated features to improve robustness
• Outlier detection and unit conversion are applied, with clinically informed thresholds guiding data cleaning
• Hourly interventions for ventilation, vasopressors, and fluids are included as time-varying signals
• Benchmark tasks include mortality and LOS predictions, and hourly intervention onset/offset prediction using multiple models (LR, RF, GRU-D)
• GRU-D and RF generally achieve strong AUROC/AUPRC across tasks, with varying F1 and accuracy patterns indicating task-specific model strengths